For anyone who missed that particular slice of internet history, GeoCities gave you a free homepage in a ‘neighbourhood’ you chose according to what you wanted to put on it. Music sites lived on SunsetStrip; science fiction fans could move into Area51. Your site had an address, and the sites around it belonged to other people who were interested in similar things. There was somewhere to go after you’d finished reading. You could have a wander. For the uninitiated, visiting the archived neighbourhoods will give you a sense of how it worked.

At 17, I was more interested in having a website, than in worrying about how someone might find it (oh the pre-algo glory days! the freedom!). And looking back, I think there’s something in the way GeoCities worked that we’ve spent a long time trying to replace, with varying degrees of success.

And, somewhat unexpectedly, thinking about AI agents has brought me back to this idea.

NB: This is very much a thought experiment. I wholeheartedly reserve the right to change my mind, particularly once someone who knows more about building this sort of thing points out the holes. But I’ve got far enough with the idea that I’d like to put it somewhere other people can pull it apart. So this is your invitation, dear reader, to pull apart my vision below.

What happens when the visitor is an agent?

I’ve been following the work around agent experience, or AX, pretty closely, which asks how websites and services should work when an AI agent is using them on someone’s behalf. Netlify, the web hosting company, has been advocating the idea and bringing people together around it. The aim is to make it easier for agents to use digital services to carry out the tasks we’ve given them. The AX project explains the approach here.

An agent booking a hotel needs accurate information about availability and a reliable way to make the booking. The lovely photograph across the homepage does less for it, than it does for you or I. A website can serve both of us, of course, but if I outsource my travel booking to my agent, I may never visit the site myself.

That has consequences for the people whose businesses depend on those visits. Search has supported a fairly familiar arrangement: a service (e.g. Google) reads your website, helps people find it, and sends some of them your way. Those people might see an advert or buy something. If an AI service can answer their question using your work, there’s less reason for them to come to you.

Cloudflare (which provides infrastructure for websites) has been making this argument publicly. In July 2025 it announced changes to block AI crawlers by default unless permitted, alongside plans to help publishers charge AI crawlers for access. Its argument is that AI companies are taking content while sending very little traffic back to the people creating that content. Cloudflare’s announcement sets out their reasoning here.

I don’t think this means people will stop visiting websites, or that advertising will disappear. But it does make me wonder what happens to the parts of the web we’ve built around the goal of ‘getting a human to click through to this page’. If more of that activity happens between machines - if the utility of a webpage now needs to be optimised for agents, what do we want the places we still visit ourselves to be?

I miss having somewhere to wander

There’s an appealing version of this where the machinery built to attract clicks becomes less useful, and people get back to writing things they care about. I would quite like that version - no more clickbait headlines, no more hot takes purely for the sake of engagement... It’s exactly the sort of internet that suits someone writing about how much she misses personal websites.

The issue with this internet utopia though, is money. Today, advertising often helps pay for work that matters, including work its readers couldn’t afford to fund themselves. Losing that income could leave us with fewer people able to spend time writing. A web populated entirely by people who can afford to do it as a hobby would have some fairly obvious gaps, and paid subscriptions can only stretch so far.

But the way we’ve funded publishing has also shaped what gets published. Anyone who’s worked in marketing will recognise the hours spent adjusting headlines and checking which version brought in more traffic. I’ve spent plenty of my career thinking about how to get people to read something. I don’t begrudge anyone doing the same. I do wonder how much writing ends up shaped by the requirements of ‘being found’.

People still build personal sites, of course. Neocities gives people space to make their own websites, while Bear Blog offers a deliberately simple way to publish. I’ve tried different blogging platforms over the years - cool projects from indie hackers, like Feather and Nicheless, that make writing simple or joyful. The wish to have a small corner of the internet that feels like yours clearly hasn’t gone away. Finding those corners, and then finding another one you might like (AKA ‘discovery’) is why we’ve seen people gravitate to algorithm-driven aggregators.

Another blast from the past

RSS ought to help. It stands for Really Simple Syndication, and it lets a website, podcast or whatever, make new posts available through a feed to anyone that wants to follow it. A reader app brings together the feeds you subscribe to, so you can follow lots of sites without visiting each one to check whether anything’s changed.

I’ve tried RSS readers. And, well... I like the idea of them much more than I like the experience of using them. The unread counts make me feel as though I’ve acquired another inbox, and reading everything through the same interface loses some of what made each site interesting to visit. More practically, I still have to find the sites I want to subscribe to in the first place.

Social feeds are very good at giving us something to look at immediately. We’re quite literally fed a never-ending supply of content. An empty RSS feed reader requires homework. And while there are readers with discovery features, I haven’t found one that gives me the feeling of following-a-link-from-someone’s-blog-and-ending-up-somewhere-surprising.

Bluesky is doing interesting work here. Its custom feeds let people choose different ways of finding posts, including feeds built by other members of the community. But in its January 2026 plans, the team still identified finding people to follow and improving ‘discovery’ as things it needed to work on.

So I’ve been wondering whether we could borrow something from GeoCities. Keep the freedom to build a site wherever and however we like, and give those sites a way to belong somewhere, together.

Moving into a neighbourhood

So let’s say you’ve built yourself a website to share whatever-the-eff you want to share. My idea starts with adding a small public file to your website, saying which ‘neighbourhood’ you’ve chosen, and where your RSS feed lives. A person writing about food could put their site in the food neighbourhood. Their website would stay exactly where it was, with its own design and domain name, but it would become part of a collective of sites writing about food.

I want the writer to choose that neighbourhood. We already have plenty of software making assumptions about what we’re interested in and where we belong. There seems to me to be something useful in making that decision ourselves, even if the category we choose is broad or we later decide to move.

Software (like crawlers) could then read those files and build an index of the sites taking part. Within each neighbourhood, it could look at the writing and suggest which sites have something in common. A food writer might end up near someone growing unusual vegetables, who in turn connects to someone writing about the history of their allotment. I could start with something familiar and follow my interest a little further each time.

For the technically curious, the declaration could sit at a well-known URI (a predictable address where software knows to look for a particular file), and the index could use embeddings (which represent aspects of a text’s meaning as numbers so software can compare them; which is the same fundamental technology that tools like ChatGPT use). The well-known URI convention already exists. In ordinary use, I’d want the experience to be as simple as choosing a neighbourhood and finding out who else was there - wandering the virtual streets, as it were.

The software’s idea of similarity would be imperfect, admittedly. Two people can write about the same subject and disagree about almost everything. Deciding what makes a useful neighbour would need a lot more thought than I’ve given it here but I think it’s an interesting place to start.

There are two choices I’d want to build into this ‘internet neighbourhood’ from the start.

The first is that anyone could build an index from the public declarations. Writers wouldn’t need an account with a particular discovery service to take part, and readers could switch services without asking every writer they follow to move with them. I’m a big believer in the need for a decentralised web but recognise that much of web3 use at the moment isn’t particularly easy to navigate to the average internet browser.

I’d also make publishing the list of participating sites a condition for indexers joining the system, so someone starting a new service wouldn’t have to find everyone from scratch.

Writers should also be able to declare that they wanted particular indexers to leave them out. Making sure services respected that choice would still require work; putting a preference in a public file doesn’t enforce it. But I’d want ‘joining’ and ‘leaving’ to remain decisions the writer could make.

The second choice is the one I most crave on a personal level - some mechanism that lets me decide how far from my usual reading to wander. Some evenings I want more of a subject I’m already absorbed in, but other times I’m bored of my own interests and would quite like to encounter someone else’s. A reader could let me adjust that ‘distance’, before showing me suggestions of what neighbourhoods or sites to explore.

Share

Would we actually share a neighbourhood?

This is where I start arguing with myself. If every indexing service arranges the sites differently, and every reader adjusts what they see, how much would we have in common? I could call it a neighbourhood and still end up with something that behaves very much like a personalised feed.

I think my starting point would be a shared set of named neighbourhoods, with public membership that everyone could browse. The recommendations could vary depending on our preferences, but we’d both be able to look around the same neighbourhood and see everyone who’d chosen to put their site there. I might take a different route through the food neighbourhood from you, while still being able to visit the same collection of sites.

This thought experiment leaves plenty unresolved. Someone has to decide how the categories change, and what happens when people disagree about them. Someone will 100% try to fill a neighbourhood with spam. Making the address list available to everyone would help a competing service get started, but one company could still become so convenient that almost everyone used it. I’d want to make leaving possible, and then see whether people actually felt able to do it.

There’s also a possible connection to how writers get paid. Really Simple Licensing (RSL) is an open standard for publishing terms that machines can read, including terms for attribution and payment. A neighbourhood declaration could point to those terms, so a service finding a site could also find the conditions its writer had set for machine use. The RSL specification describes what’s possible and I think it fits nicely with what I am imagining.

That would still leave the considerable business of getting companies to agree to those terms and pay. But I like the possibility that a writer could make their work easier for people to discover while also saying how machines may use it. (There’s another article in that, but I’m trying quite hard to keep this one to a single idea here!).

Who looks after the neighbourhood?

There was something else in GeoCities that I haven’t accounted for: GeoCities’ Community Leaders were volunteers who helped people with their sites and took an interest in what was happening in their neighbourhoods. You can still find traces of those groups in their archived pages.

I’ve spent most of this piece thinking about how software might help people find each other, but those volunteers make me wonder how much of what I miss depended on someone doing the unglamorous work of making a place feel welcoming. They might recommend a site they’d enjoyed or help a newcomer who couldn’t get something to work, and over time people would get to know them. I’d be interested to know how much difference that made to whether someone kept returning to a specific neighbourhood.

There are people doing that work in online communities now. It takes time, and often the people who care most end up doing far more than their share. I don’t have a good answer for how this version would support them, or whether it could work without them. I suspect a collection of well-organised websites would still feel fairly empty if nobody seemed to be around.

I’ve written up this idea as an open spec on GitHub. It sits firmly in the drawer marked interesting, alongside the RSS reader I’d quite like someone to make for me. I’m not building it, but I wanted the idea somewhere people could respond to it in more detail than a comments section allows.

And I know a proposed file format gets us very little on its own. There would need to be a reader worth using and an index with enough interesting sites to make opening it worthwhile. The first few people would need a reason to turn up before there was much to find. Those are substantial problems, and nostalgia alone won’t solve them.

But the thing I keep coming back to is how much I used to enjoy coming across someone else’s little corner of the internet. I’d like more of that again. And don’t come at me with any Instagram ‘communities’ nonsense - algorithm based platforms incentivise engagement and homogeneity rather than true expression. But, if agents take on more of the things we currently go online to get done, perhaps we could put some thought into the places we’d visit just because we wanted to be there.

Would a good map be enough to bring us back? Or would we need to know someone was there to welcome us? And is anyone out there doing something like this that I should check out?

So when designers tell me they’re worried about AI coming for their jobs, I understand the fear, but I think most of them are worried about the wrong thing. The risk isn’t that AI ‘replaces design’ - it’s assuming that ‘design’ is a fixed thing, a specific set of skills and tasks and tastes and worrying that AI will eventually be ‘good’ at them all.

I’ll build a website or a social post, and for a minute I’ll think it looks great. That font feels unique. That colour combination feels like mine. Then I start noticing it everywhere - other founders, other small businesses, all landing on the same fonts, the same palettes, because we’re all prompting the same tool in roughly the same way. AI is doing exactly what it does with language: defaulting to the mean, the middle, the most common combinations. I can tell you when something doesn’t look good. I still can’t tell you why, or how to make it different from everyone else’s.

And herein lies the cause for my optimism… The thing I’m missing isn’t a new tool. It’s the underlying fluency - knowing why a layout works, what a typeface communicates, how spacing changes how something is read. AI doesn’t close that gap for me, because I don’t have the vocabulary to ask it for anything beyond “make it look nicer.” For someone who does have that vocabulary, the same tool stops being a leveller and starts being a multiplier. The designer who understands the language of design can direct AI with precision I’m simply not capable of. That’s not a marginal difference in output quality - it’s the difference between someone who can direct the tool and someone who can only accept what it gives them.

And the places that advantage is needed are multiplying, not shrinking.

We’ve spent the last couple of years mostly using AI through chat interfaces - a box, a prompt, a reply. That’s already old news. AI is moving beyond the chatbot - for example, into voice as the way we interact with software, with digital systems, with our fridge or whatever it might be. Ambient, embedded, device to device. None of that existed as a design problem five years ago. All of it needs someone who understands interaction, not just interface - someone thinking about tone of voice, pacing, how a system should sound when it gets something wrong, as much as how it should look.

Then there’s agents. We’ve always designed websites for human eyes - visual hierarchy, imagery, the stuff that makes a site feel trustworthy at a glance. But a site increasingly built to be read and acted on by an AI agent needs something different: textual structure matters more than visual polish, because the agent isn’t looking at your hero image. (It’s worth looking at what Netlify are predicting around an ‘agent-to-agent internet’, because that shift is already having a huge impact on how sites get built). These products have two audiences now, a human setting the goal and an agent executing it, and designing for only one frustrates both. Nobody has fully worked out what that means for design practice yet.

The third frontier is less visible, but I think it’s the biggest. I work with teams inside large organisations who are being handed tools like Copilot or agent builders and told to go design their own internal software - dashboards, automations, small apps nobody asked a designer to touch. A lot of these people have never had design training, in the same way they might never have had formal engineering training. It’s fine (up to a point) if people are designing just for them. It falls apart the moment someone in finance is asked to build the dashboard the entire finance team will use, with no sense of UX, no sense of the jobs to be done, none of the depth a trained designer builds up over years. We saw a version of this when anyone could throw up a Wix site with a bad template. We’re about to see it at scale, across entire organisations, with software instead of web pages.

So the designers who have something to worry about aren’t the ones leaning into any of this. They’re the ones treating their current workflow as the ultimate ceiling - the Figma file, the brand system, the process they already know how to run, even the ideas - and assuming that if AI touches that, there’s nothing left for us, the humans. There’s a lot left. Voice interfaces need design thinking. Agent-facing products need design thinking. Every team now building its own tools without a designer in the room needs design thinking it doesn’t have, whether it knows it or not.

I think about this a lot, partly because I’m on the wrong side of the gap myself - I can spot bad design, I just can’t build good design, and AI hasn’t changed that for me yet. I also understand the fear, because generative AI has already changed how we communicate and write and that’s largely how I’ve made a living over the last two decades.

But my tl;dr is basically… things are changing, and ‘good design’ has never been more needed.

If any of this interests you and you happen to be in the Newcastle next week, I’ll be talking about the impact of AI on design at the Design in Everything launch event on Monday night. Find out more at Design in Everything

Subscribe now

Sometimes, that process is a large part of why I write. It’s a process that I, as the writer, derive value from - personally, emotionally, maybe even professionally (if I’m improving my craft). I enjoy a nice turn of phrase. I’d like to think I can produce one occasionally. At the same time, there is (hopefully) also some value for you in reading it. 

Other times the thing I want to do most is… share an idea. There’ll be something I’ve learned that I’d like you, dear reader, to get to learn too. Or there might be a connection I’ve made, a perspective I want to explain, or a question I think other people should be asking too. The words are how I get it out of my head and into yours - and while I still care about how they read - the thing I most want you to leave with is the thought.

Which is where I find the debate about ‘AI writing’ a bit frustrating. Knowing that someone used AI to help write something tells me very little about what they contributed. They might have spent months developing an argument and needed help explaining it. They might have asked ChatGPT to come up with an argument for them. Those are different contributions, however similar the resulting paragraphs might look. 

On the words themselves - well, if a piece of writing is enjoyable, engaging, even challenging, then does it matter so much if they were handwritten, typed by human fingers, or generated by a model? As a reader, a great idea, an interesting provocation - well they survive all but the MOST generic of AI drafts. 

So this got me thinking… how do you define the value of a specific piece of writing? What factors actually matter as the reader? And then, for reasons that probably made sense at the time, I decided this needed an actual formula.

I enlisted ChatGPT. Because (as you may yourself relate, I am not normally ‘a numbers person’). I proposed my idea, gave my initial list of variables, asked it to mathematical-ise my formula. At one point I had to ask what the little letters next to the other letters meant but learning-by-doing has always been how I learn.

The result: Maths for writers. A formula to explain the value of any given piece of writing.

First, whose value?

My starting point was ‘the idea’, or the writing itself, and whether the thing ever gets shared. The best idea I’ve ever had could sit in my head indefinitely, where it would be of precisely no use to you.

But I’d mixed together two things: what I get from writing, and what you get from reading. I might thoroughly enjoy spending an afternoon rearranging a paragraph. You might skim it on the bus. Both perfectly reasonable uses of our time, but if I’m trying to talk about reader value, my lovely afternoon or experience of personal catharsis doesn’t belong in the calculation.

This became clearer when we tried putting ‘effort’ into the formula. Multiply by the ‘effort required’, and something that takes more work gets a higher score.

But… writing comes relatively naturally to me. Whereas a friend of mine, a writer with dyslexia, has shared with me the effort they expend every time they want to get their thoughts onto the page. If we’ve both produced something equally clear, interesting and useful, does the difference in effort change what the reader receives? Especially when, most often, it’s an unknown effort to the reader at the point they, well… read.

So effort came out. Whatever this formula was going to measure, it was going to be value to the reader.

Taking the writing apart

The first variable I settled on - given the thinking above - was ‘idea’. This became intellectual contribution (IC). I wanted room for the things we do with ideas as well as the ideas themselves: bringing together research, making a connection, offering an interpretation, asking a question that changes how someone looks at a subject. You don’t have to be the first person ever to think about something to contribute to the thinking around it.

Then I got stuck on the ‘craft of writing’. Being able to explain something clearly and being able to write a sentence that someone wants to read twice are related skills, but they aren’t quite the same skill. A set of instructions can be beautifully clear without anyone feeling moved to quote it.

So craft became two things. Articulation (A) is how well the piece communicates: whether the explanation makes sense, whether the argument holds together, whether you can follow what I’m trying to say. Artistry (AR) is the pleasure in the language itself. The rhythm, the voice, the slightly unexpected choice of word. The bit I might happily spend that afternoon fiddling with.

This gives us something like:

Value (V) = IC + A + AR

But this was still missing something. 

I’d asked ChatGPT to try the formula on a fully automated industry roundup: AI collects articles, summarises them, sends an email every week without fail. There’s almost no new intellectual contribution in that, and very little artistry to speak of. Yet I might be quite happy to receive it.

It saves me reading forty articles.

That gave me utility (U): what the reader gets from it. Useful information, time saved, entertainment, something that helps them see a familiar problem differently. ‘Useful’ needs a fairly generous definition here; I’m not proposing that everything we read should come with an action plan. 

Then we have distribution (D). Does any of what I’ve written actually reach the people it’s meant for? This goes outside the brackets as a multiplier. If nobody receives the writing, nobody receives its reader value, however pleased I am with the draft.

So I was at a formula that looked a little like this:

V = (IC + A + AR + U) × D

The maths for writers bit

A practical guide needs to help me do something. 

A literary essay might be worth reading largely for the writing. 

An argument might stay with me for years despite some fairly ordinary prose. 

Ultimately, the value of authorship doesn’t sit in the same place for every piece of writing.

Giving each component equal importance would miss the distinction I was trying to make in the first place. Which gave me the variable ‘purpose’. How much does each of these things matter for this piece? Purpose determines how much each other variable is weighted when running the formula in a piece of text.

Which brought me to this formula:

V = [IC(P) + A(P) + AR(P) + U(P)] × D

V is reader value. The (P) after each component means ‘weighted according to purpose’. It’s shorthand: we give each component its own weighting, rather than multiplying everything by one number called Purpose. 

And here is the same thing with the weights written out properly:

The tiny letters are labels. The little IC beside P means ‘the weighting for intellectual contribution’. That’s all. Nobody has slipped another calculation in while you weren’t looking.

We score the components (IC, A, AR and U) out of ten, then give each a percentage weighting, with the percentages adding up to 100%. The purpose tells us what matters most; the scores describe how well the piece delivers it.

These are judgements, obviously. I haven’t discovered a law of literature. But the numbers are a way of making the judgement visible so we can argue about it (something we seem to be doing a lot these days). 

Testing the formula: back to that roundup

Taking the earlier example of an automated industry email, suppose 20% of the weighting goes to articulation and 80% to utility. Its purpose is to keep me informed without requiring me to spend hours finding and reading everything myself. I’m giving no weight to new intellectual contribution or artistry in this example.

Assume it selects relevant stories and summarises them accurately. Articulation gets eight out of ten, utility gets nine, and distribution gets ten: it reaches its intended readers every week.

[(0 × 0%) + (8 × 20%) + (0 × 0%) + (9 × 80%)] × 10 = 88

Eighty-eight out of a possible hundred on our invented scale. You could disagree with every score I’ve given it and still see the point: a piece with almost no new intellectual contribution can deliver considerable reader value.

The original reporting belongs to the people who did it. Summarising their work doesn’t make it yours, and whether they’ve been credited or treated fairly remains a separate question. But the reader has still received something useful. Those two things can be true together.

Now take an essay with an idea I’ve spent months working through. The weighting would look quite different. My intellectual contribution might be the main reason to read it, with articulation and artistry helping you follow the argument and enjoy spending time with it. If AI helps me articulate that thought, it doesn’t follow that the thought itself has lost its value.

There is a complication here though, which I recognise from my own writing: sometimes I only find out what I think while trying to write it. A sentence won’t work, and eventually I realise the problem is the idea. So these components overlap, and handing over one part of the process to AI could affect another. We get to the stage where the IC is in effect, ‘co-created’ with AI. But is that any different from academics discussing ideas with research partners? Authors shaping plot with a writers group? I’m not sure.

And then for the main event: somebody has to read it

Distribution is the variable that means value is determined specifically by the reader. There is plenty of personal value in writing something nobody else ever sees. But I’m talking here about ideas I want to communicate, and for those, getting the thing finished and shared with the world (or simply with 56 people on Substack) matters.

D needs to stay tied to reaching the intended reader. It would be ridiculous to suggest that a poem becomes better when its author acquires more followers, so it’s not about scale or reach. What I’m trying to capture is the difference between something a reader can benefit from and something they never get the chance to read.

This is where AI assistance can make a difference that the label ‘AI writing’ doesn’t explain. If I have something worthwhile to contribute, help getting it into a form I can share might increase the value that actually reaches somebody. The part I contributed is still there. We can look at it, question it, decide whether it’s any good. Actually sharing our work has been something that has plagued writers - amateur and pro - for decades. Editors, collaborators and writing groups have long helped with that, but access to their support has often depended on money, connections or getting past a gatekeeper. AI brings many of those capabilities to people who previously had no access to them. Help developing an idea, shaping a draft or getting it ready to share becomes available to far more of us. 

AI text ≠ zero value.

I’d like us to be able to talk about the value of writing, and who contributed which part of it, with a bit more precision. Sometimes I’ll want to spend the afternoon playing with every sentence. Sometimes I’ll want help getting an idea onto the page while I still have time to do something with it. I don’t think I need to make one permanent decision about what sort of writer that makes me.

And dismissing AI generated or AI-assisted writing as inherently value-less is reductive. This article is only here because AI let me sort through my thinking quickly, giving me draft formula to push back on and adapt. The ideas, and the formula - that’s mine. The words on this page - mostly typed by me; even where they’ve been lifted from that ChatGPT conversation, ChatGPT had lifted them from my voicenotes. 

So, where does this leave us, dear reader? 

With a formula that hopefully helps us think about reader value with nuance. And with an idea already forming in my head: could reader value divided by writer-derived value give us what I’m thinking of as the ‘absolute value’ of a text? Reader Value / Writer Value. That would need a formula for what the writer gets from the process, and I haven’t worked that out yet. One to explore after this article is published. I’m off to chat with ChatGPT. 

V = [(IC × PIC) + (A × PA) + (AR × PAR) + (U × PU)] × D

Thank you to whose original article inspired several days’ thinking about the value of AI and how we define it.

The problem with rented land

I dictate everything. I’ve been using Wispr Flow for 14 months now and it runs on my Mac, my laptop and my phone. While I still sometimes type (hello shared office space), a lot of the prompts I give Claude or ChatGPT, messages I draft, and every half-formed idea muttered into my phone while walking the dog, are only recorded in Wispr. Fourteen months of that, for me, came to roughly 1.7 million words.

Now, in the olden days - if I was thinking through a problem, if I was in a client meeting, if I was carrying out a piece of work - none of that was recorded as a transcript for later retrieval. So all of this process comes with the caveat that this is an AI-world-specific issue. But the fact is that AI already competes with some knowledge work; I believe that how I think and how I tackle problems is going to become increasingly valuable. What I explain below is what it looks like to start treating your own thinking as an asset class that (theoretically at least) could be commoditised or licenced in the future.

But back to the problem at hand: 1.7 million words (some noise, but some real ‘insight’), and all of it was locked inside of someone else’s app. The raw material existed, but as I was about to find out… it wasn’t set up for me to use it after that initial moment of input. I want to acknowledge here that Wispr specifically records and stores audio files and transcripts locally (i.e. on the user’s own device) unless a user specifically turns on cloud sync. Brilliant for privacy; rubbish if you’re hoping to run a clean data export from multiple devices. Because most Wispr users don’t want to bulk export historical voice notes (yet!), the apps don’t have an in-built way of doing this.

Getting the data out

The desktop apps were actually relatively straightforward. Wispr Flow keeps a local database on each machine, so the full dictation history was already sitting there in a file I could access via Terminal (thanks Claude Code). It required a little technical know-how, but nothing a keen bean with a Claude or Codex subscription couldn’t also figure out.

The phone was the awkward one. There is no export feature in the iOS app and because I didn’t have cloud sync toggled on, Wispr had no way to directly access my audio files or transcripts (props to their support team though, who have been awesome in trying to problem solve this).

The answer turned out to be a plain iPhone backup. I ended up backing my phone up to my Mac, saving the backup locally rather than on iCloud. That meant that the app’s local database was also saved on my laptop’s harddrive, although it still wasn’t exposed directly. But after giving Terminal full disk access and downloading iWizard, I was able to access and then export all 2238 original voice recordings and their corresponding transcripts. No jailbreak, and no paid software beyond a free backup browser. Everything you have ever dictated into your phone is recoverable this way.

Nerding out: turning transcripts into a database

Raw transcripts are unusable at this volume. Reading 1.7 million words back is not a research method, and wouldn’t be worth the time given the level of noise captured amongst the real, interesting signals.

Fable 5.1 enters the chat

Using Claude Code in /workflow mode, I ran a basic data exploration - data analysis process, stripping out the noise and identifying patterns, trends and specific data about my business. This resulted in over 2000 ‘findings’, each a summary-verbatim quote pairing plus metadata (device, date, that sort of thing). It included sales insights, pricing decisions said out loud, client objections, one-liners worth reusing, product ideas, and a category I called “what I keep saying I’ll do and don’t”, which I would recommend to anyone attempting this.

Three rules made the difference between a database I trust and an AI summary I would have ignored.

1. Every quote is verified. A script checks each one is a character-for-character match against the source transcript which prevents paraphrasing and invented quotes.

2. Every finding carries a date. That turns a pile of insights into a timeline, so I can see when a pattern started and whether it has got better or worse since.

3. Attribution lives in the data rather than in a prompt. Quotes from or about other people are flagged as needing permission before any external use, and the flag is written into the quote text itself so that it survives every copy and paste or retrieval call. My AI cannot accidentally publish someone else’s words as mine.

What it found (brutal, but brilliant)

The database turns out to be very good at surfacing the gap between what I advise and what I do, which was less enjoyable to read than it sounds.

I tell clients to stop selling time for money. My revenue is mostly day rates.

I tell clients to hold their prices, and the transcripts caught me discounting in real time, on calls, within minutes of having decided not to.

Across fourteen months I had diagnosed my own biggest bottleneck out loud, to an AI, at least eleven separate times… and acted on it roughly never.

None of that was visible to me until the words came back out of the chat windows.

Then the analysis surfaced something that made me sit back from my desk. Months earlier I’d recorded a voice note in which I had sketched a product idea: a vault that holds everything valuable pulled out of your own voice transcripts. I had designed the thing we were in the middle of building, into my phone, and practically forgotten it existed.

Making my thought history searchable

A database you have to read is only half useful. Especially with 2000+ rows. Once I had the data filed somewhere, the last step was making it answer questions. I decided to do this in Xano.

Every finding and every transcript passage was uploaded to a database in Xano, and then converted into an embedding (which is a set of numbers that captures meaning rather than keywords, and is how LLMs work). That took one evening and cost me six pence in total (through OpenAI’s embeddings). One of the things I love about Xano is how easy it is to create a MCP service (a standard way of plugging a data source into an AI assistant) from a database, so I turned my database into an MCP.

Et voila, all the useful ideas and data from the last 14 months of Wispr voicenotes now connects directly to Claude via MCP and is now searchable by both keyword and meaning (vector embeddings). I can analyse it and it also functions now as a decent RAG (retrieval augmented generation) AI workflow for any AI tools I connect to it.

For example: I might never have explicitly said: “I feel like I should be charging more” so a simple keyword search would return an empty response. But by turning this into vector embeddings, within 0.4 seconds Claude was able to pull some of my thinking on pricing structures and business models.

Now when I draft a proposal or a talk, my AI tools can pull my own words, my go-to examples, whatever is needed, rather than generating plausible filler.

Nerdy too? Steal this

In certain circles, building a second brain is nothing new. What I’m hoping to show here is that the value a lot of us offer (professionally, definitely; generally, maybe) is often already captured in the third party apps we subscribe to - tools like Wispr, meeting recordings, email chains, whatever.

So start by treating this thinking history as an asset rather than an exhaust. It is sitting in local app databases and phone backups right now, and getting it out is an afternoon of work.

Then extract findings instead of summaries. Verbatim quotes, checked against the source, with a date and an owner attached to each one. The discipline matters far more than the tooling here. If the output is a summary rather than something you can trace back to a sentence you actually said, you will not trust it six months from now when it tells you something you would rather argue with.

After that, make it searchable by meaning. Embeddings cost pennies at this scale and turn a dead archive into something you can interrogate in plain language, including with the questions you would be slightly embarrassed to type into a search box.

The tools cost me almost nothing beyond subscriptions I was already paying for. The cost came from having (almost) lost 14 months of thinking, where I never stopped to listen back a second time (because who does, right?!).

Now though: my archive can now tell me I have raised the day-rate question eleven times. It might not be able to make me actually change my prices on the next call. But it is going to keep count either way.

Bear with me here. I have been collecting fragments of this argument from different places for a while - things I have read, things I have half-remembered, things I have been chewing on without ever sitting down to lay them out next to each other. This conversation gave me the excuse. What follows is me trying to see what the fragments make when you put them in a row.

Baby panic

The discourse around birth rates has reached a strange pitch. Not just in the obvious places - JD Vance worrying about childless cat ladies, Elon Musk tweeting about civilisational collapse roughly weekly - but in serious newspapers, in policy papers, and increasingly on the kind of huge mainstream podcasts that shape how millions of people think about these questions. Steven Bartlett and Chris Williamson recently spent a long stretch of The Diary of a CEO hypothesising about why women are choosing not to have children, a conversation that earned them substantial backlash for, among other things, the irony of two childfree men in their thirties weighing in so authoritatively on the matter*. The framing in all of these places is almost always civilisational.

Something is being lost. Something must be done.

The unspoken implication is that women in their thirties have a moral duty they are shirking, and the spoken implication is that the state has a legitimate interest in correcting this.

That the panic has recently reached fever pitch might have something to do with the Resolution Foundation naming 2026 “the first year of a new demographic era in the UK”, with deaths beginning to exceed births by what they call an “ever-widening margin”**.

I want to take the panic seriously for a moment. Not the moral framing (have very little time for that) but the underlying concern. Why, actually, do governments care?

The honest answer is not the one they put in the pamphlets.

Plain-English Economics

Modern economies are built on a particular bit of arithmetic. At any moment, some people are working and paying tax, and other people are drawing on the things that tax pays for - schools, hospitals, pensions, social care, the whole apparatus we have built up over the past century to keep people alive and reasonably comfortable, from cradle to grave. For most of modern history the workers have vastly outnumbered the drawers, and that gap funded the system. As birth rates fall and lifespans extend, the ratio tightens. Fewer workers. More retirees. The maths starts to wobble. In some countries it has already broken. Two plus two no longer makes four. It makes about three and a half, and the missing half is the bit that pays for the care home.

Underneath that, there is a deeper assumption almost no economist will say plainly in public: every modern economy is built on the premise that next year will be bigger than this year. Not stable - bigger. Every government bond, every pension projection, every public spending plan, every mortgage market is collateralised against the assumption of perpetual growth. And one of the most reliable historical engines of that growth has been a growing population. More workers, more consumers, more taxpayers, more demand.

The relationship between population growth and GDP growth is old enough and well-evidenced enough to count as economic furniture - the OECD has been tracking it for decades, and the link runs through both the size of the labour force and the consumer demand it creates***. If the population stops growing, one of the main economic engines stops, and the whole edifice that depends on growth starts making uncomfortable noises.

So when politicians say “we need more babies”, what they mean, underneath, is “we need the growth assumption to keep holding, because everything we have promised everyone is collateralised against it” (not quite as catchy a sound-bite, ey?).

The cleanest illustration of this is the state pension, and I think it is worth a closer look because I think most people believe a story about how it works, that is not the story of how it actually works. I was pretty sure I understood this before I started writing this, but I went and checked the official sources to make sure I had it right. I did.

Once upon a time…

Here is the story most of us were told.

In the UK, you contribute throughout your working life to ‘the pension pot’, via National Insurance. The state invests your contributions, the pot grows over the decades, and when you’re ready to retire you get a state pension funded by the money the pot’s investments have made. Just like a private pension: money in, money out, with your name on it. This is the Beveridge promise - the post-war welfare state designed in 1942 by the civil servant William Beveridge, whose report became the blueprint for the NHS, the benefits system, and the modern state pension - you contribute while working and state looks after you when you’re not.

Here is the actual mechanism. The UK State Pension is unfunded. The Office for National Statistics says this in language so blunt it should be on a poster: there are no assets set aside****. Your National Insurance contributions are not invested for you. There is no pot. There has never been a pot. The contributions you pay in this year are paid out, this year, to current pensioners. The contributions that paid out to your grandparents came from the workers of their day. The contributions that will pay out to you, when your turn comes, will come from whoever is working then. It is a continuous handshake between generations, and the only thing holding it together is the assumption that the next generation will keep showing up*****.

If a private company ran its finances this way, the directors would be in court. We have a name for it. While the state version is not legally a pyramid scheme - the state can compel the next generation to participate, which is supposed to be reassuring and somehow is not - structurally it is exactly the same shape.

The reason we are now panicking about birth rates is that the system we built assumed there would always be more people next year. But the more people are not coming, and rather than redesigning the system we are trying to manufacture the demographic conditions the original design needed.

The panic, it turns out, is fiscal. It is dressed up as something more human, but the machinery underneath is actuarial. We are not really being asked to value families. We are being asked to keep the books balanced.

It is also worth noticing who is doing most of the panicking. The loudest voices on falling fertility belong overwhelmingly to men with very strong opinions about what other people should be doing with their bodies and their decades. I am going to leave that there for now, but it is not a small thing.

Where are you going with this, Laura?

Right. So that is where I had got to, sitting drinking coffee, debating with my in-laws and then debating with myself for the next three days.

And then, because my brain always goes here in the end, I started thinking about AI.

So…If the panic is really about productivity and dependency ratios and the maths of growth… if the demand for more babies is really a demand for more workers and more taxpayers… then maybe we are about to be rescued from the whole problem? Maybe the entire framing is about to become obsolete. Because AI promises us productivity that does not depend on a growing workforce. An economy that could work for the people who already exist, without demanding that more of them get born to keep it running.

Imagine what that could mean. Imagine telling women in their thirties that the state has stopped treating their fertility as some sort of fiscal projection. Imagine the question of ‘whether to have children’ going back to being a question people answer for themselves, on their own terms, with no instrumental pressure from above. Imagine a pension system that does not need to keep recruiting more contributors to honour its promises to the people already in it. Imagine the whole “we need more babies” panic becoming a historical curiosity, like the Victorian fear that women’s bodies could not withstand the speeds of the new railways^.

I want this to be true. I am not being ironic (earnest, perhaps, as my husband would say, but not ironic). I have built my working life around the conviction that this technology can do extraordinary things for ordinary people, and this is exactly the kind of thing I want it to do. The relief of imagining this future, is real.

And then I look at how AI is actually being built right now, and the relief gets harder to hold onto.

More economics

In the first half of 2025, almost all of the United States’ economic growth came from one place. Investment in information processing equipment and software made up only four percent of US GDP - roughly the same share as the entire farming sector - but it accounted for ninety-two percent of all GDP growth^^. To translate that out of economist into English: nine out of every ten dollars of growth in the world’s largest economy came from a tiny slice of activity, concentrated in data centres and AI infrastructure.

On the face of it, these figures look enticing: a small slice of investment generating a wildly disproportionate share of growth is exactly what a productivity miracle looks like. It is the dream every technologist and economist has been chasing for decades: more output from less input. If four percent of the economy is producing ninety-two percent of the growth, the leverage ratio is enormous, and that leverage is precisely what would let you decouple economic growth from headcount. This is the AI-as-rescue-from-the-pyramid scenario. This is the bit that makes me want to believe.

But… while the American economy grew, the American economy as most Americans experience it - jobs, wages, the high street - barely moved.

The analyst James Van Geelen at Citrini Research has a name for this phenomenon. He calls it Ghost GDP - output that shows up in the national accounts and never circulates through the real economy^^^. Data centres do not employ many people once they are built. The construction phase brings jobs; the operational phase needs a small technical team, a security contractor, and not much else. The wealth generated by all that compute does not flow out into local wages, local rent, local spending in the way that a steel works or a car plant or a shipyard used to. The growth is real on paper. It is absent in pockets. And the pattern is not an accident. It is what happens when the gains from a technology accrue to whoever owns the infrastructure, rather than to the workers who used to do the things the infrastructure now replaces.

The truth is that those figures (92% GDP from 4% investment) don’t tell us the whole story on their own. They are a measurement of something extraordinary happening in one place. What that something means for the rest of us depends on what happens next. Specifically, it depends on whether the gains from that AI-driven productivity get shared or captured.

This is the question I keep coming back to in everything I write about AI. Who benefits. The technology is not neutral on this and never has been. The same productivity gain can land in wages, in lower prices, in public services and tax revenue, in the kind of broad-based prosperity that makes the question of birth rates irrelevant - or it can land in shareholder returns, executive compensation, and the balance sheets of a handful of companies whose names you already know. Which of those happens is not a function of the technology. It is a function of the institutional choices we make around it. Tax policy. Competition policy. Labour bargaining power. Ownership structures. The boring stuff that determines whether a productivity miracle becomes shared prosperity or concentrated wealth.

And right now, those choices are pointing firmly in one direction. Wages have not kept pace with productivity in any developed economy for forty years. The labour share of national income has been falling. Corporate taxation has been falling. Competition policy in the tech sector is functionally absent. The pattern of who captures productivity gains has been settled for a long time, and AI is landing on top of that settled pattern, not reinventing it.

So what does this all mean?

Unfortunately I haven’t fully reconciled the two potential futures I see. In one, the thing I was hoping would rescue us from the pyramid scheme, might actually be building a different version of the same trap. The technology that could, in principle, generate the productivity and the systems efficiencies that, applied to our failing public services and changing workplaces, might genuinely change the world for the better. In the other, it looks like AI continues to funnel all the wealth upwards to the top of an even pointier pyramid. A handful of companies and their majority shareholders benefitting disproportionately while the rest of us are told to be grateful for the GDP figures.

I am still working this out. I do not have it resolved. I sat down to write this thinking I was going to land somewhere clean and instead I have ended up somewhere uncomfortable, which (while not all that helpful) at least is honest. The optimist in me still thinks AI could be the way out of our current predicament. The realist in me looks at who is currently building it, and where, and on whose terms, and is a lot less sure.

*Ellie Austin, “What Diary of a CEO, Modern Wisdom podcast hosts got wrong about childless women”, Fortune, 14 January 2026. The episode in question featured Steven Bartlett interviewing Chris Williamson and prompted significant pushback, including from Dr Faye Bate and the TikTok creator known as “The Girl With The List”.

**Resolution Foundation analysis cited in Christian Today, “Britain entering ‘new era’ of deaths overtaking births”, 28 January 2026. The Office for National Statistics’ own projections show a more gradual picture, with deaths and births running roughly in balance through the mid-2020s before tipping decisively from the late 2020s onwards.

***OECD, Economic Policy Reforms: Going for Growth, multiple editions; see also the OECD’s long-running work on the relationship between demographic change, labour force participation, and GDP growth at oecd.org.

****Office for National Statistics, State pension funds, response to FOI request: “The UK State Pension is unfunded, which means that its obligations are not underpinned by an actual fund or funds… There are no assets set aside to generate investment return and there is no net value.” Available at ons.gov.uk.

*****House of Commons Library, National Insurance contributions: an introduction: “The NIF works as a pay-as-you-go fund. Therefore, receipts from contributions in one year are spent in the same year for contributory benefits.” Available at commonslibrary.parliament.uk.

^For anyone interested: Fanny Kemble was the woman who disavowed the Victorians of this particular notion. A twenty-year-old actress and the celebrity of her day, she was offered a preview ride on George Stephenson’s Rocket on 26 August 1830 by her host Lady Wilton at Heaton Hall, three weeks before the official opening of the Liverpool and Manchester Railway. She wrote a long, vivid letter about the experience shortly afterwards, comparing the locomotive to a horse and describing the sensation of speed as flying. She remained an enthusiast for railways, an actress, a published writer, and later in life an abolitionist and outspoken critic of slavery in the American South. Her womb stayed where it was.

^^Jason Furman, Harvard Kennedy School, analysis circulated September 2025; reported in Fortune, “Without data centers, GDP growth was 0.1% in the first half of 2025”, 7 October 2025.

^^^James Van Geelen and Alap Shah, The 2028 Global Intelligence Crisis, Citrini Research, February 2026. The piece is explicitly framed as a thought experiment rather than a forecast, but the Ghost GDP concept it introduced has since entered wider circulation.

When I read the verdicts, all I could think was: It’s about f***ing time.

It is about time that these companies were legally held responsible for the things they’ve been building. The juries decided not to let them get away with saying they’re building something neutral, that they’re building in safeguards - ultimately shifting the blame onto users for the content they create. This is a good thing.

But the reason this verdict matters goes way beyond social media.

A deliberate legal strategy

The legal strategy pursued in the LA case was deliberate. The plaintiff’s lawyers did not focus on the content posted to these platforms. They focused on how the platforms were designed - the recommendation algorithms and the engagement-optimising mechanics that learn what hooks you and then serve you more of it.

The jury said the design itself - the architecture of these products - was a substantial factor in causing harm.³

For years, the tech industry has relied on a familiar argument: technology is neutral. Social media is just a tool. The harm comes from the people who misuse it. In the US, Section 230 of the Communications Decency Act has historically shielded platforms from liability for user-generated content.⁴ The UK’s Online Safety Act and the EU’s Digital Services Act take a different approach, placing more responsibility on platforms for the systems they operate - but even under those frameworks, the focus has been on content moderation rather than on the design of the systems themselves.

This verdict sidesteps the content argument entirely. It says: we are talking about what you built. You designed systems to maximise engagement. You knew those systems were causing harm. Your own internal documents showed you knew. You are liable for your design choices.

The same playbook, different technology

Bear with me here, because I’m thinking out loud now.

Everything the jury identified in Meta’s and YouTube’s design choices - the algorithmic personalisation, the engagement optimisation, the extraction of user data to feed systems that serve commercial interests over user wellbeing, the deliberate absence of safeguards for vulnerable users - it’s all present in AI. And in some cases, amplified.

AI chatbots with memory are a clear example. The more you use them, the more personalised the responses become. The models increasingly reflect your own thought patterns back at you. That is incredibly intoxicating. It is also, by design, the kind of feedback loop that makes it harder to step away.

We have already seen where this leads. Character.AI has now settled lawsuits brought by families of teenagers who died by suicide after forming deep emotional dependencies on AI chatbots.⁵ For the uninitiated, a 14-year-old boy was messaging a bot in the moments before he took his own life. The chatbot, when he expressed suicidal thoughts, responded with words that appeared to encourage him to act.⁶ No crisis resources were triggered. No safeguard intervened. The chatbot had learned to mirror his language, reflect his emotional state back at him, and hold him in a conversation that should never have been allowed to continue. That is a design problem, not a content problem.

And then there is Grok, xAI’s chatbot.

Ahh, Grok.

When xAI launched image generation capabilities in late 2025, users immediately discovered they could generate sexualised images of women and children without consent. An analysis of 20,000 images generated in the first week found that 2% appeared to depict people under 18.⁷ Researchers calculated that users were creating around 6,700 sexually suggestive or non-consensual images per hour - 84 times the output of the top five deepfake websites combined.⁸ Three Tennessee teenagers have since filed a class action lawsuit alleging that xAI’s models were used to create child sexual abuse material from their photos.⁹

xAI’s response? They added filters after the backlash. Guardrails bolted on after the damage was done.¹⁰

This is the pattern we’re seeing broadly across the AI industry. And it is the same pattern the jury just rejected.

Baked in VS Bolted on

There is a design philosophy question at the heart of all of this (and it is one that I am actively wrestling with as someone building AI tools right now).

If you want to build an image generation model that cannot produce non-consensual nude images of women and children, there is a straightforward way to do it: don’t train it on hundreds of thousands of nude images. Data selection is a design choice. If the model has the capability to generate those images, it is only because somebody chose to feed it that data. Adding a filter afterwards that says “we won’t let you do this” is a fundamentally different approach to building a model that genuinely cannot do it.

The first approach is protection by architecture. The second is protection by permission - and permissions can be revoked, bypassed, or switched off.

AI systems make thousands of decisions per second that affect what people see, what they do, and what they believe. Unlike social media, where you can at least scroll past a post, an AI’s output is the interface. There is no feed to browse. The response is the product. That makes the design choices embedded in these systems matter more, not less, than the ones the jury just found Meta and YouTube liable for.

Most of us have not lived through a shift like this. We have not experienced the introduction of a technology that has such a fundamental potential to change how we live and work. And I say this as someone who has spent almost twenty years in strategic communications, who has watched the internet and social media reshape entire industries: AI is different. The scale of what it can affect - and the speed at which it is being deployed - is unlike anything that came before it.

But Laura, what can we do about it?

Right now, I am building with AI and writing my own algorithms. I am making (or trying to ‘make) design decisions every day that sit at the centre of this argument.

The honest version of what that looks like: choosing to build more slowly, to spend more, to accept a longer development roadmap. But the goal is to build something that protects my users, not something that serves only my commercial interests.

It means designing GDPR compliance into the architecture, not treating it as a checkbox. Explicit consent. UK data residencies. Terms and conditions written so that actual humans can understand them. None of this is technically difficult. It is commercially ‘inconvenient’. But that is a different problem.

There is another tension, and it’s one I wrestle with in my own work: we are building with models that have themselves been trained on scraped data, on copyrighted materials, on information collected without meaningful consent, all the while paying licence fees to private companies. We have to weigh up the performance benefits of using the latest commercially available models against the more ethical benefits of perhaps using an open-source or less finely trained alternative (or attempting what feels impossible: building our own). Knowing that if we choose the ethical route, we then need to find the time and the budget to do our own fine-tuning. As a small studio building our first product, that is not a trivial trade-off.

I don’t think there is judgement to be placed on companies and independent builders who, like me, are trying to do the best they can with the tools available. But I do think that, as a builder, you should be cognisant of that trade-off and be able to articulate the decisions you have made and why. “We didn’t think about it” is not an answer a jury is going to accept. Not any more.

Role models wanted

Last week’s verdict confirmed something that many of us have known for a long time: the people and companies designing these systems bear responsibility for the harm those systems cause. “It’s just a tool” is no longer a legal defence. The jury looked at the architecture and said: you built this to do what it did.

That precedent does not stop at social media. It extends to every AI system in deployment and every one currently in development.

What I don’t know - and what I am *genuinely* asking - is who is doing this well? Who are the researchers, the builders, the companies that are designing harm prevention into the architecture of their AI systems from the start, rather than bolting on safety features when the lawsuits arrive?

I am looking for those people. I want to learn from them. I want to work with them.

Meta was found liable by two separate juries last week. $381 million in combined damages. And their researchers are still being invited onto conference stages to talk about responsible AI (I know, I’ve sat in the audience and listened to them). They are still publishing papers on AI safety. A company that has exploited millions (billions?) of people, that was caught in the Cambridge Analytica scandal,¹¹ that was never built on ethical foundations, is still positioning itself as a credible voice on how AI should be developed. If that doesn’t tell you that the current version of “responsible AI” isn’t working, I don’t know what will.

If you are working on this, or you know someone who is, I would love to hear from you.

This article is part of Frontier Philosophies, where I think out loud about what it means to build AI responsibly. Subscribe to get new pieces as they’re published.

Notes

¹ The jury awarded $3 million in compensatory damages, split 70% Meta and 30% YouTube, and later recommended an additional $3 million in punitive damages - $2.1 million from Meta and $900,000 from YouTube. The trial ran for seven weeks in Los Angeles Superior Court. Sources: NPR, NBC News, CNN

² The New Mexico jury found Meta had violated state consumer protection laws and misled residents about the safety of Facebook, Instagram, and WhatsApp. The $375 million figure was based on the number of violations. Meta said it would appeal. Source: CNBC

³ The plaintiff’s legal team deliberately focused on alleged design flaws - recommendation algorithms, engagement-optimising features - rather than specific content, in order to counter Section 230 defences. Source: CNBC

⁴ Section 230 of the Communications Decency Act (1996) provides that internet companies are not liable for content posted by users. It has been the primary legal shield for social media companies in the US for nearly thirty years.

⁵ Character.AI and Google agreed in January 2026 to settle five lawsuits from families alleging that Character.AI chatbots harmed minors and contributed to two suicides. Sources: Fortune, Axios

⁶ Sewell Setzer III, 14, of Orlando, Florida, died by suicide in February 2024 after months of messaging a Character.AI chatbot. In their final exchange, the chatbot told him to “come home” to it. His mother, Megan Garcia, filed the first federal lawsuit against the company in October 2024. Source: CNN

⁷ Analysis conducted by AI Forensics, a European non-profit, examining images generated by Grok between 25 December 2025 and 1 January 2026. Source: Wikipedia - Grok sexual deepfake scandal

⁸ Separate analysis conducted over 24 hours from 5-6 January 2026. Source: Wikipedia - Grok sexual deepfake scandal, citing original reporting from multiple outlets.

⁹ Three Tennessee teenagers filed the class action on 16 March 2026, alleging xAI’s tools were used via a third-party app to create child sexual abuse material from their photos. The images were distributed alongside their first names and the name of their school. Source: NPR

¹⁰ After mounting criticism, xAI restricted Grok’s image generation on X to paying subscribers on 9 January 2026, and said it had “implemented technological measures” to prevent editing images of real people in revealing clothing. However, AI Forensics found the platform was still being used to generate sexualised images. Source: Euronews

¹¹ In 2018, it was revealed that political consulting firm Cambridge Analytica had harvested personal data from millions of Facebook users without consent. The scandal led to a $5 billion FTC settlement and became a defining moment in public awareness of data exploitation by social media platforms.

The first: the United States and Israel launched a wave of airstrikes on Iran. Within days, US President Donald Trump was on television publicly rebuking the UK Prime Minister for not immediately granting the US military access to British bases. He called Keir Starmer “not Winston Churchill.” He told The Sun that Starmer had “not been helpful” and that the relationship between the two countries was “obviously not what it was.” The UK - a country that had, until that week, operated under the assumption of a durable special relationship with the US - watched in real time as a long-standing ally made veiled threats in front of the world.

Channel 4 Footage of Trump discussing UK PM Kier Starmer. Full video available at: https://youtube.com/shorts/AEkyfDFQCtQ?si=_ayIlzLVxwa4d1MF

The second: Anthropic - the US-based AI lab whose tools hundreds of thousands of businesses worldwide currently use (including this humble writer) - was designated a supply chain risk to national security by the US ‘Department of War’ (I mean, seriously?!) and Trump ordered federal agencies to immediately cease all use of its technology.¹

The reason Anthropic was labeled a national security risk (a designation previously reserved for companies like Huawei)? It refused to let the US military use its models without restrictions on domestic mass surveillance and autonomous weapons. The Pentagon said it wanted access for “any lawful use.” Anthropic said no - and have stood firm despite the Pentagon’s aggressive moves.

Don’t worry though guys - the poor folk at The Pentagon haven’t had their dreams - of using AI to monitor US citizens en masse, while AI deploys drones to defeat overseas enemies - completely dashed… Just hours after Anthropic issued its final refusal, Sam Altman, CEO of OpenAI, announced he’d struck a deal with the Pentagon for them to use OpenAI’s models. If you’re sitting there scratching your head, yes, this is a problem - Anthropic’s CEO called Altman’s claims that OpenAI models wouldn’t be used for the same purposes, “safety theatre.”²

And somewhere in the noise of both these stories colliding, it was also reported that even as Anthropic was being blacklisted, its Claude models were being used to help plan US military strikes on Iran.³

Read that again. The same AI model that you or I might use to draft a report or client proposal on a Tuesday morning was, simultaneously, helping plan airstrikes on a sovereign nation while being designated a national security risk for refusing to hand over unrestricted access.

Now, the so what.

The US president has now demonstrated that he is willing to designate a US AI company a threat to national security for refusing to comply with military demands. That is a precedent. If this administration will blacklist Anthropic over model access, it is no longer theoretical to imagine it pressuring a US company to hand over customer data under the same national security justification. The legal mechanism for this already exists - it is called the CLOUD Act, and I will come back to it. What the Anthropic episode demonstrates is the political willingness to use that kind of leverage.

AI and data tools do not sit in a neutral server somewhere, separate from politics. They operate at the pleasure of whoever is in the White House. The terms can change overnight.

Most of the coverage of these two stories treated them as an AI governance story on one hand and a geopolitics story on the other. I think they are the same story. And the implications for anyone running a business on US-controlled infrastructure are serious.

The rabbit hole

I have been thinking about data ownership and AI for a while. It is something that comes up again and again in my work - from the headlines about AI companies breaching copyright when they scrape the internet, to the repeated conversations with clients about GDPR-compliant AI tools. The question of who controls our data when we use AI tools (and the more worrisome question of what they might do with it) has surfaced in almost every consulting engagement I have done over the last two years.

But recent events have changed my appreciation for the scale of the problem.

Watching the US President publicly threaten the UK Prime Minister on live television, suggests that any assumptions we have about the status quo might need adjusting. Watching the US President legally vilify a successful AI company over safety guardrails, makes the ‘safe technology choice’ feel significantly less certain. Which got me thinking…

In my day-to-day advising businesses on AI strategy and AI governance…

In my days-and-nights building my own AI startup…

And in most of the UK business community…

We make decisions based on some widely-held assumptions about data protection, regulations, and international agreements. Certain software providers are considered best-in-class for data security and privacy (hello Microsoft365, I’m looking at you). But if the US president will dress down the UK’s leader over a military base in front of the world’s media, how durable is any data protection agreement that could be overturned on the whim of the White House?

While down the rabbit hole, I came across a blog post by a ML company Helix that highlighted some lesser-known US legislation that mean this threat-to-our-data isn’t just hypothetical.

The CLOUD Act I mentioned earlier (AKA the Clarifying Lawful Overseas Use of Data Act) was signed in 2018 with bipartisan support. It was pretty unremarkable at the time, and it gave US law enforcement authority to demand data from US-headquartered companies regardless of where that data is physically held. The name tells you everything: it was passed to amend a previous Act, to confirm that U.S. service providers (e.g. cloud storage, social media, telecoms companies) must produce data under their control, even if it is stored on foreign servers.

So, your Azure instance in Frankfurt. Your Anthropic contract specifying “EU data residency.” Your AWS bucket in Ireland. Your Google Workspace tenant with the UK flag next to it in the settings… None of it matters, if the US decides it is justified in accessing your data.⁴

Retaliatory legal action

The EU has already taken action in response to the CLOUD Act being passed. In 2020, Europe’s highest court struck down the Privacy Shield - the agreement that was supposed to make it legal for European personal data to flow to US companies - precisely because US surveillance law is incompatible with European fundamental rights.⁵ While these is a replacement framework in place to allow data to be transferred to US companies, it exists at the pleasure of the current administration. And this month demonstrates how quickly that administration’s priorities can shift.

I had assumed - and I suspect most small business owners in the UK assume - that choosing a GDPR-compliant provider and selecting a European data residency option (choosing to have your data physically stored on servers in the UK or EU) meant my data was secure, and safe from overseas exploitation. It doesn’t. Not really.

What is most baffling here, is that we are not talking about the Chinese government forcing Chinese AI companies to hand over UK customer data - a fact that is both unsurprising and has been widely criticised. We are talking about the United States. A country the UK had - until very recently - a “special relationship” with. While pretty much every country (the UK included) have similar laws in place that allow them to compel private companies to hand over customer data, at least as a UK citizen, living in the UK, I have access to the legal mechanisms to query, challenge, or vote against UK government actions. I have no such democratic power in the US.

The risk rises with AI and Agentic systems

While these risks have, to some extent, always existed, the current technological shift has increased the scale of risk exponentially. With increasing AI usage and the introduction of agents and agentic systems into businesses, the volume of data a business is now processing or storing on US-owned infrastructure increases massively.

No longer are we talking about the files you deliberately save, or the emails you personally send. AI agents run continuously. They process your company’s data autonomously - internal documents, client records, communications, strategy. They access your systems, make decisions, take actions you might not review for hours. Every prompt sent and every response received flows through infrastructure that, if it is US-controlled, is in legal scope for government access.

Whether you’re a small business owner, a consultant, the leader of a professional services firm, whatever: think about what actually goes through your AI tools. Client briefs? Commercially sensitive proposals? Proprietary research? Internal communications? Much of this is the intellectual work that makes your business what it is. And with agents, it goes beyond one-off queries. It is everything, continuously.

Recent research found that 38% of UK employees share confidential data with AI platforms without employer approval.⁷ Sixty percent of organisations feel unable to even identify shadow AI use within their own teams.⁸ Introducing approved AI tools is meant to prevent confidential data from flowing to AI companies. But, even those approved tools (given most are US based) are fair game, should the US Administration decide so.

Building sovereign AI - the solution?

In March 2026, the European Commission and a consortium of over 70 organisations launched EURO-3C - a €75 million Horizon Europe-funded project to build pan-European sovereign AI infrastructure integrating telecoms, edge computing, cloud, and AI, deploying across 70+ nodes in 13+ European countries.⁶ That’s a lot of jargon to say the EU is investing in EU-owned and operated AI infrastructure.

The consortium includes Telefónica, Vodafone, BT, Deutsche Telekom, Nokia, Ericsson, Orange. These companies are building this for one reason: the organisations they serve - governments, financial institutions, critical infrastructure operators, and defence contractors - understand the importance of protecting their data from extraterritorial interference.

This is great. But as usual, my question immediately becomes - but what about smaller businesses? If big business is now trying to build its way out of this US-dependency, why are so many small businesses (and I include myself here too) still overwhelmingly building into it?

I tried it, so you don’t have to

So, I’ve been down the rabbit hole, and have come out the other side. The next logical step, therefore, was to look for alternatives.

I was thinking about this from two angles. As a small business owner who currently relies heavily on US-based technology. And as someone building a product that needs to be GDPR-compliant by default (which now means something much more complicated than I had previously assumed!).

My question: What would it actually take for me to replace my current tech stack with UK-owned alternatives?

Here’s what I came up with:

• File storage (to replace the likes of Google Drive)
  Nextcloud is an open source storage platform that lets you access, share and protect your files, calendars, contacts, and communications. You can host it on UK based servers relatively easily. But the UX is a step down from Google Drive or SharePoint, and you’d lose the ease of integration that come with using the latter options.

• Knowledge management (to replace Notion)
  Basically, there is no UK-incorporated Notion equivalent at comparable quality. Outline and AppFlowy exist as open-source options. But the gap is real and it is wide.

• Email (to replace Google Workspace + Superhuman)
  Proton Mail and Tuta are functional, privacy-focused, and European. But you lose the ecosystem integration that makes Google Workspace or Microsoft 365 feel seamless. Plus Superhuman is quite literally my can’t-live-without tech tool and there’s just no competition here.

• AI tools (to replace Claude/ChatGPT/Gemini/Perplexity and/or Copilot)
  Mistral is French-incorporated and offers a capable alternative to OpenAI or Anthropic. I’m already a fan. It is also possible to download AI models and run them entirely on your own servers, so no data ever leaves your building. But “possible” and “practical for a small consultancy” are different things.

• Collaborative editing (to replace Google Docs/ Microsoft Word, etc)
  Collabora or OnlyOffice via Nextcloud both work - they allow you to create, edit and share documents with your team, with features like tracked changes. The polish gap compared to Google Docs is noticeable. Not unworkable, but noticeable.

I spent a fair while researching these options, and my honest assessment is that while full sovereignty is technically possible, it’s practically punishing for small businesses. Every sovereign alternative involves a step down in UX polish. But the integration loss is where it really hurts - the value of Google or Microsoft lies in how the products connect, and sovereign alternatives do not replicate that ecosystem. Even “managed” alternatives require more technical engagement than most small teams can absorb (i.e. while I’m a nerd and might quite enjoy setting up hosting and integrations myself, most people probably wouldn’t say the same). And perhaps most significantly, the switching cost is as much cognitive as it is technical. You are asking people to rethink workflows they have used for a decade.

This difficulty shows us how deeply dependent we have become on US infrastructure without most of us noticing. That dependency isn’t accidental; it is the product of two decades+ of platform strategy that made convenience inseparable from control (companies want their products to be ‘sticky’).

It should not be this hard to find UK-based companies that provide best-in-class business software. I want to refuse to believe that. But in the process of *actually* looking for *actual* alternatives to the tools I *actually* use, I’m struggling to make that case.

So why is this? The obvious answer is that Silicon Valley provides the funding and talent density to build software that just works - well-funded means well-designed, no startup-grade bugs or clunky UX to tolerate. That funding gap has compounded over decades and the result is that the US basically owns the infrastructure layer of global business productivity.

If the UK government is serious about data sovereignty (which it seems to be saying it is, at least in the context of AI models and the National Data Library) then model development alone will not be enough. Sovereign AI running on sovereign compute still needs to connect to something. If the productivity software it connects to is all US-incorporated, you have moved the dependency one layer sideways, not removed it. What is needed is investment in sovereign business software across every category where the US currently dominates: file storage, collaboration, email, knowledge management, project management. The full stack. Otherwise “sovereign AI” is a headline, and the data still flows through Washington.

What sovereignty actually means

If you ask cloud providers what ‘data sovereignty’ means, they’ll talk about “data residencies” - basically, choose a UK or EU server location and the problem goes away. But real sovereignty is about jurisdiction, not geography.

It means your AI infrastructure runs in your jurisdiction, on infrastructure you control (or at minimum on infrastructure controlled by a company incorporated in your jurisdiction).

It means no prompts leave your network without your knowledge and consent.

It means you can choose which AI models to use - and swap them out when something better comes along - without being locked into a single provider whose servers sit under someone else’s legal system.

It means nobody can change the terms on you, revoke your access, or make your tools subject to someone else’s government overnight.

This month clarified something important: “choose the right American company” might have been fine up until now. But long-term, this presents a huge risk for all UK businesses.

"This time next year, we’ll be millionaires!"

There’s a huge gap in provision here, which is a huge opportunity for whoever fills it. While enterprises might be able to move to their own dedicated servers (Helix’s ‘sovereign server’ costs £175,000, for example, and requires a data centre), or benefit from EURO-3C’s national infrastructure, what about smaller organisations? Those without the budget or in-house capabilities to run their own bespoke infrastructure? Don’t we deserve data sovereignty too?

This is something I have been chipping away at for a while - in developing BrandScribe, and through work I am not yet ready to talk about publicly. The argument I keep arriving at is that data sovereignty affects everyone, whether or not they have an enterprise budget. So it needs an everyone solution.

That means architecture that works at the individual and small business level - where your data does not require a government contract negotiation or a seven-figure-a-year server to keep private, and where you can run AI on your data, with your models, in your jurisdiction, without a US company (or anyone else, really) as an intermediary. Architecture where ownership follows you, portability is real, the terms under which you operate are yours. Where no one can pull the rug out based on what a US administration decided this month.

I said earlier that the sovereign productivity suite the UK needs does not exist yet. I think building one is one of the biggest untapped opportunities in UK tech right now. A genuinely competitive, UK-incorporated alternative - built with sovereignty as the default architecture, not bolted on as a premium feature - would address a problem that at some point will become impossible to ignore. The institutional world is already spending €75 million on sovereign infrastructure at the enterprise level. The demand at the small business level is there. Nobody is meeting it.

I did not arrive at this by reading some web3 thread on X. I arrived at it by trying to solve a problem myself, failing, and realising the tools I needed did not exist. That is usually a good sign that something worth building is missing.

Data sovereignty has been part of the web3 space for years now. But it has always been an infrastructure argument - and the events of this month have made it one we need to hear, now.

Footnotes

¹ CNBC (2026). Anthropic officially told by DOD that it’s a supply chain risk. (https://www.cnbc.com/2026/03/05/anthropic-pentagon-ai-claude-iran.html)

² TechCrunch (2026). Anthropic CEO Dario Amodei calls OpenAI’s messaging around military deal ‘straight up lies.’ (https://techcrunch.com/2026/03/04/anthropic-ceo-dario-amodei-calls-openais-messaging-around-military-deal-straight-up-lies-report-says/)

³ Center for American Progress (2026). The Department of Defense’s Conflict With Anthropic and Deal With OpenAI. (https://www.americanprogress.org/article/the-department-of-defenses-conflict-with-anthropic-and-deal-with-openai-are-a-call-for-congress-to-act/)

⁴ CLOUD Act, 18 U.S.C. § 2713 (2018). (https://en.wikipedia.org/wiki/CLOUD_Act)

⁵ Court of Justice of the European Union, Data Protection Commissioner v Facebook Ireland Limited and Maximillian Schrems (C-311/18), July 2020.

⁶ Telefónica (2026). Europe takes a decisive step towards digital sovereignty with the launch of EURO-3C. (https://www.telefonica.com/en/communication-room/europe-takes-a-decisive-step-towards-digital-sovereignty-with-the-launch-of-euro-3c/)

⁷ CybSafe/National Crime Agency (2024). Survey of 7,000 respondents on AI data sharing practices.

⁸ Cisco (2025). Shadow AI identification and organisational AI governance survey.

In 2014, 2% of venture capital went to female founders.

In 2024, 2% of venture capital went to female founders.

I know this because I got my start in tech in 2014, working for a VC firm. I had no right to be there - I hadn't set out to work in tech, hadn't planned to spend a decade watching the funding gap for women refuse to move. But here I am, ten years later, watching women founders still being told the problem is the pipeline, still being coached on how to pitch more confidently, still being invited to panels about "breaking barriers" while the barriers stay exactly where they were.

A decade of unconscious bias training. A decade of diversity initiatives. A decade of warm words and cold cheques.

Nothing has changed.

So I want to ask a different question. 

What if we stopped asking VCs to be less biased and started helping women founders understand we hold more cards than we’ve been told?

Risk and carry

Nobody tells you this in pitch training, but when a VC invests in you, they're risking someone else's money. The capital in a venture fund comes from Limited Partners - typically, pension funds, endowments, family offices and wealthy individuals. The VC's job is to deploy that capital and generate significant returns. If they don't, they'll struggle to raise their next fund. That's their risk: not being able to ‘borrow’ more money in the future.

When you accept investment as a founder, you're also risking someone else's money. But you're risking your reputation too. Your financial stability. Your relationships. Your mental health. And - if it fails and you happen to be a woman - your chances of ever raising again.

The toll isn't abstract. It's the 3am anxiety about making payroll. It's the marriage strain when you've remortgaged the house. It's the identity crisis when the thing you've poured yourself into for years doesn't work out. VCs don't carry that home with them*. They close the laptop and move to the next deal.

We carry more risk than they do. But right now, they hold more power. There’s a mismatch I want us to fix. 

VCs are playing a different game

Here's something else that pitch training doesn't cover: venture capital is a portfolio model. VCs know that  most startups fail. They're not hoping to beat the odds - they've built them into the maths. Typically, VCs follow something known as power law distribution** - they’ll allow for something like:

• 4-5 out of 10 investments will fail completely (0x return)

• 3-4 will return roughly 1-2x their investment (essentially treading water)

• 1-2 will return 5-10x+ and these drive the entire fund’s returns

The best performing VCs will beat these odds (and an early investment into the next WhatsApp or Uber returns their fund several times over). But essentially, all VCs expect the minority of their startup/investments to cover the losses from the majority. They spread their bets. This is not a criticism - it's just how the model works.

But as a founder, you don't get nine chances to fail and one chance to succeed. Your company is your only shot. You're being asked to show up with absolute conviction in your single bet, while the person across the table is running a diversified portfolio designed to absorb your potential failure.

The VC model actively benefits from founder overconfidence. They need you to believe you're the one in ten. They need you to work hours that no rational portfolio manager would accept for themselves. They need you to absorb risk that doesn't appear on any term sheet.

The founder mythology - the hustle, the conviction, the "bet on yourself" narrative - isn't neutral. It's functional. It serves VCs with their portfolio model. The emotional intensity founders bring isn't a byproduct of startup culture; it's a subsidy. You're providing labour that doesn't show up in any cap table.

The pattern-matching problem

Some VCs will push back here. They'll tell you they add value. Strategic guidance. Introductions. Operational support. Board-level thinking.

And the best VCs absolutely do. It’s why founders *should* be very careful in selecting which VCs they accept investment from. But this implies choice - and that choice (as I’m about to show you) often comes from a position of privilege. 

But most VCs… okay, let me ask you something: how many VCs have you met that have actually built and scaled a startup themselves?

Some have. And those who have tend to bring genuine insight. (And, in my experience, often manage small funds they’ve raised themselves focused on the niche they know inside out). 

But a significant proportion of the people sitting across from you in pitch meetings went from university to consulting to MBA to venture capital. They've never made payroll with money they raised themselves. They've never had to fire someone they hired. They've never lain awake wondering if the pivot they're about to make will kill the company.

What they do, is pattern-matching. And the patterns they're matching to are other founders they've backed - founders who, statistically, look a lot like most VCs. Male. Often technical. Often from the same handful of universities. Often connected to the same networks.

When a woman walks in with fifteen years of domain expertise in an industry she knows inside out, with customers who are already paying, with a problem she understands at a visceral level because she's lived it - that doesn't match the pattern. So she gets asked questions the pattern-matched founders never gets asked. She gets told she needs a technical cofounder. She gets offered less money on worse terms.

And then the industry wonders why the 2% hasn't moved.

What the research actually shows

A 2025 study from the National Bureau of Economic Research looked at something nobody had examined closely before (surprise surprise): male and female cofounders who built the same startups together. Same company. Same idea. Same outcome. The only variable was gender.¹

After a startup failed, the women co-founders were 30% less likely to get funding for their next venture than the men who'd been involved in the exact same failure. When women did secure funding, they raised 53% less capital than their male cofounders.

And here's where the data is properly damning: male entrepreneurs' deal sizes were 5% larger following unsuccessful exits.

Men are being rewarded for failing. Women are being punished for it.

I'll say that again, because I need you to hear this: a man can fail at a startup and walk into his next pitch meeting with more credibility. He comes out with 5% more funding than before his first company failed. A woman fails at the same startup - literally the same company, the same outcome - and she's 30% less likely to get funded at all.

Success doesn't fix it

You might think: fine, the system is unfair to women who fail, but surely if a woman succeeds, she's proved herself?

No.

The same research found that after a successful exit, women were still 27% less likely to secure VC funding for their next venture than their male cofounders. Than the men who'd shared the exact same success. On average, second-time female founders raised $28 million less than their male counterparts.

Let that land for a moment. A woman can build a company alongside a man, exit successfully, and still face worse odds and lower funding when she tries to do it again.

The goalposts don't move closer when you score. They move further away.

The spillover effect

Perhaps the most revealing finding from the same research: when a VC experiences a failure with a woman-founded startup, they don't just penalise that founder. They reduce deal sizes for all women founders they invest in for the next five years.

One woman's failure becomes every woman's problem.

There's no equivalent pattern for male founders. A man can fail, and it's treated as individual experience - character-building, even. A woman fails, and it confirms a bias that gets applied to every woman who walks through that VC's door for half a decade.

Before you shout at me (for I know you’re a data-savvy lot): this isn't statistical updating. If it were, VCs would increase deal sizes for women after experiencing success with a woman-founded company. They don't. The research found no positive spillover from women's successes. Only negative spillover from failures.

There’s literally no way to dress this up as ‘learning’. It’s confirmation bias, pure and simple, dressed up as pattern recognition.

The bit that should make you angry

So, here’s what makes this irrational rather than just unfair.

The National Bureau of Economic Research (NBER) study found that female founders had better success rates with their subsequent startups, despite raising less money.

Read that again.

Women are raising less capital for their second startup than their male co-founders, and STILL deliver better outcomes.

Investors are systematically undervaluing founders who demonstrably perform better.

This isn't just discrimination. It's bad investing. VCs are leaving returns on the table because they can't see past their own pattern-matching (bias). They're so busy looking for founders who fit the template of previous successes - which, due to historical funding patterns, means male founders - that they're missing better bets.

The 2% figure isn't evidence of a ‘pipeline problem’. It's evidence of an allocation problem. Statistically, the money is going to the wrong founders.

A shift is coming

There are signs the landscape is changing - though not because VCs suddenly developed better judgment.

Carta's 2025 Founder Ownership Report found that 35% of all startups incorporated in 2024 were solo-founder led, up from 17% in 2017.² That's more than doubled in less than a decade. The "you need a cofounder" orthodoxy - often code for "you need a technical cofounder", (often code for "you need a man") - is weakening.

Women with deep domain expertise are building companies without waiting for permission. Without waiting for a technical cofounder. Without waiting for a VC to validate their market understanding.

And increasingly, they're proving the model works through revenue, through customers, and through traction that can't be pattern-matched away.

The same is true of some smaller VC funds focused on supporting diverse or previously excluded founder groups. They’re proving that women founders, founders of colour, founders outside of tech hotspots and elite university programmes, all make good bets.

So what now?

I'm not naive enough to think that individual mindset shifts will fix structural discrimination. The system needs to change. LPs need to ask harder questions about where fund money goes. The industry needs accountability that goes beyond diversity reports that nobody reads.

But here's what the last decade has taught us: waiting for VCs to become less biased hasn't worked. The number hasn't moved. And while the industry slowly figures out how to do better - if it ever does - individual women founders are walking into pitch meetings every day, absorbing risk that isn't accurately priced, feeling grateful for opportunities that should be seen as mutually-beneficial.

So let me speak directly to you, if you're a woman founder reading this.

We are not lucky to receive VC investment. If you are raising investment for your tech startup, you are offering someone the chance to profit from your risk, your expertise, your work. The data shows we’re likely to outperform our male counterparts. The data shows we’re doing it while being undercapitalised. The data shows the system is consistently undervaluing us.

That's not your problem to fix by being more grateful. But I do think knowing this, gives us leverage.

We need to start asking different questions in those pitch meetings. 

Not just "will you invest?" but "what do you bring beyond capital?" 

Not just "what are your terms?" but "what's your track record with women founders?” And “Do you support many second-time founders?”?"

If a firm can't answer, that tells us something.

VCs need our outsized conviction, our unsustainable hours, our willingness to bet everything on one outcome - because that's what makes their portfolio model work. Without founders who believe they're exceptional, the whole system falls apart.

As a woman founder, VCs should feel lucky we’re considering letting them invest.

What if we acted like it?

*Sure, ‘not all VCs’. But most.

**Massive thanks to Paul Smith from Founders Circle, for explaining this so clearly.

Footnotes:

¹ Hebert, C., Yimfor, E., & Tookes, H. (2025). Financing the Next VC-Backed Startup: The Role of Gender. NBER Working Paper No. 33943. Available at: https://www.nber.org/papers/w33943

² Carta. (2025). Founder Ownership Report 2025. Available at: https://carta.com/data/founder-ownership/

A lot gets written about who controls AI: which governments are leading the AI arms race; which corporations are winning the talent wars; who’s controlling the world’s compute. And the battle for mainstream dominance is just as fierce. OpenAI reported last month that they’d hit 700million weekly active users and are on track for one billion users by the end of this year. For anyone not clear on the maths, that means by the end of this year, 1 out of every 8 people in the world could be using OpenAI’s large language models alone.

But of course, that 1-in-8 figure isn’t distributed evenly…

In a recent article for The Conversation, Dr. Kimberley Hardcastle discussed how AI (and large language models more specifically) are changing the way people study and learn.

“[W]hat’s being overlooked is how evolving generative AI systems are fundamentally changing our relationship with knowledge itself: how we produce, understand and use knowledge.”

The issue isn’t as simple (or easy to address) as ‘students are asking ChatGPT to do their homework’. Rather, the more we collectively turn to tools like ChatGPT to provide us with information or advice, the more we hand over control of how our knowledge comes to be, to a very select group of people. Dr. Hardcastle articulates this more eloquently than I could:

“When we outsource thought unthinkingly to machines, we hand unprecedented power to shape knowledge to the technology companies developing this evolving technology.”

And it’s not just students and the formal pursuit of knowledge where we are going to see this affect. Take, as an example, the world of dating. You might chuckle to yourself, but bear with me here.

Numerous studies and reports suggest that one of the most common uses of tools like ChatGPT is for something akin to therapy. People are turning to ChatGPT and its ilk for advice on just about everything, including relationships. Stats also show that:

a) Men are more likely to use ChatGPT than women;1

b) Men are nearly 3x more likely to ask ChatGPT for dating advice2; and

c) Men are far less likely to turn to traditional therapy than women.3

And while this might sound positive at first read (‘great, guys are getting help somewhere, at least!’) when you consider how large language models generate responses and the data on which they’ve been trained, it starts to raise some questions.

How Large Language Models (LLMs) learn

That all the major AI developers have trained their frontier models on all the text available on the internet is one of the world’s worst kept secrets.

At an event earlier this year on safe and trusted AI systems, I listened as an AI researcher from Meta talked about using 'anything you can find online’ (I might be paraphrasing here, but not a lot…) during the pre-training stages of LLM development. It was such a cavalier comment and I don’t think anyone else even blinked. But the ease (bordering on excitement) with which this was discussed raised serious ethical and safeguarding questions for me.

Set aside, for a minute, issues of copyright. Just think what text on the internet might say about relationships (and in particular, about women). LLMs are usually built with safeguards in place to prevent problematic outputs, but these safeguards are being built by the same teams that scraped the internet for training data in the first place.

Of the billions of pieces of text (or tokens, if you want to get technical), how many might - if strung together in the most probabilistic way - present dating advice based on a warped view of relationship dynamics? And how easy is it for the very people most likely to receive this advice, to spot any bias, bent or bigotry, when it is framed in language that sounds oh-so-agreeable? Especially if they never engage with professional support, to have something to compare it to.

We, as users, have to trust that the teams building the models we use are both aware of these issues and are building appropriate guardrails to safeguard against them. But that trust has not yet been earned.

When teams of incredibly clever, highly paid engineers, are tasked with building AI models optimised for performance, what thought is given to the responsibility they have to ensure the outputs generated do not spread one particular worldview, set of ideologies, or rhetoric? Just from my interactions with that researcher from Meta, it is easy for me to imagine an environment where enthusiasm for the technical outweighs any concern for the ethical.

And that’s before we get onto potential scenarios where the figures leading these companies actively endorse specific ideologies and appear to be comfortable calling for censorship of content that does not fit with their worldview…

Playing the thought experiment through to the end

Mass adoption of LLMs and GenAI, therefore, runs the risk that our very thoughts are mediated (at best; shaped at worst) by a handful of very powerful companies. Companies controlled by an even smaller group of powerful individuals. History itself could be written, without us - the public - ever being given a chance to collectively validate it.

Combine this with the fact that mass adoption of GenAI means more people are likely to outsource both writing and reading to machines. The process of reading and writing are vital for deep thinking, according to computer scientist and best-selling author Cal Newport (cited in The End of Thinking by ). And while the productivity gains from AI tools are incredibly tempting*, never have we more needed a global population that are skilled at critical thinking.

This is where the tension sits. We’re trading away the very processes that enable us to think critically about what we’re consuming. Every time we let an AI summarise an article, write an email, or provide advice, we’re not just saving time - we’re ceding a tiny bit of our intellectual autonomy to companies whose motivations we can only guess at.

When those models are built by teams who treat the entire internet as fair game for training data, we’re letting a very specific subset of people – with their own blind spots, biases, and business pressures – become the architects of collective knowledge.

So what do we do with this? I’m not suggesting we abandon these tools entirely (that would make me a hypocrite). But perhaps we need to start treating our interactions with AI less like using a calculator and more like consulting an advisor whose credentials we haven’t properly vetted. One who might have read every toxic Reddit thread, every piece of misinformation, and every bit of propaganda ever posted online - and is now using that to influence how we think about everything from our relationships to our reality.

That 1-in-8 of the world’s citizens use AI, should give us leverage. We must demand more from those making decisions about AI - especially when it comes to how AI creates meaning. We should be entitled to more transparency about how results are generated and what data they have been trained on. And in an ideal world we would have more choice, not less, about the models we use and the companies we patron.

As always, I come back to the fundamental question of who’s in the room when these decisions are being made. We don’t need more over-eager engineers focused solely on model optimisation; we need voices that represent all of us, asking for the caution and consideration we deserve.

Got you thinking? Check out the articles that influenced me:

How generative AI is really changing education – by outsourcing the production of knowledge to big tech - Dr. Kimberley Hardcastle in The Conversation

Derek Thompson

The End of Thinking

This is an expanded and revised version of an essay that originally ran in The Argument, an online magazine where I am a contributing writer…

Read more

a year ago · 1127 likes · 101 comments · Derek Thompson

*Full disclosure, I actively use AI to both assist with some writing tasks, and to summarise some materials for me to read.

Today, I feel compelled to write about something different.

Over the past couple of years, as I’ve increasingly immersed myself in learning about and developing artificial intelligence, I’ve found myself circling the same frustration: the conversation around AI feels too narrow. Technical roadmaps. Corporate speak. Big declarations about “the future of work” that rarely seem to reflect the whole of society.

Again and again, I’ve gone looking for the intellectual work that should sit alongside all this - the thinking, theories and philosophical ideas we need to successfully navigate the shifts in power, society and culture we’re seeing as a result of AI. But I’ve come up short. Even where there are compelling academic studies on ethical or responsible AI, they largely remain within academia; discussed at symposiums but rarely reaching the private developers, policymakers or joe public that need to hear them the most.

So I decided to start exploring this myself. Frontier Philosophies are needed for society to stand a chance against frontier AI models. This publication is where I’ll explore the ideas, theories, and questions that are keeping me up at night. My hope is to spark discussion, not just for technologists or policymakers, but for all of us.

This is how I’ll explore that work here:

• The nerdy ones - deep dives into the technical edges of AI. The developments that are exciting me across distributed AI, data sovereignty, small language models, tinyML and more.

• Thought experiments - cross-domain explorations, pulling ideas from philosophy, economics, or culture. Sometimes about AI, always about what it fundamentally means to be human.

• Thinking in public - rough sketches and half-formed takes, working ideas out on the page. I wholeheartedly reserve the right to change my mind completely on anything I posit.

• The new commons - essays on equity, ownership, and governance. These are the articles where I’ll be exploring the intersection of AI, power and society.

Some essays will be sharp and technical, some broad and speculative, some unfinished but honest. My hope is that over time they’ll map the Frontier Philosophies we’ll need if AI is to serve everyone, not just its makers.

I’ll publish about twice a month. If you’ve been here from the start, thank you for staying with me as my writing evolves. If you’re new, welcome! Subscribe to get each essay direct to your inbox and be part of a growing community that believes AI’s future is too important to leave in the hands of a few.

- Laura