Truth, Trust, AI, Tech, Universities, and Policy

Tom Forth, 12 April 2026.

“I wonder which of Paris and London is closer to the population-weighted centre of their country” I pondered online.

I’d just seen Alasdair Rae’s LinkedIn post with his latest calculation of the population centre of the UK and I’d recently looked at the French state railway’s 2026 Atlas, reminding me of how centralised France’s railway network ridership is on Paris.

Of course I could ask an AI. But I forced myself not to.

I trust Alasdair Rae’s calculations over mine, but I still confirmed them with my best efforts. Partly I wanted to check that I still could.

We agree that the field just to the South West of Lullington Park Cricket Club — midway between Burton upon Trent and Tamworth — is, according to the best data and estimates we have available, the population centre of the UK. 169km from the centre of London.

France is easier. I got a citizen-generated map of the population centres of every commune of Metropolitan France from their fantastic national data portal at data.gouv.fr, downloaded the most recent population estimates of those communes from Insee, their national statistics office, and calculated the population-weighted average of the X and Y coordinates in the French metre projection EPSG 2154 in Excel. It came back at the North East of the field to the North East of Farges-en-Septaine, a village just North of the French Air Force base at Avord. 199km from Notre Dame Cathedral in Paris.

London is closer, by about 30km, to the population centre of the UK than Paris is to the population centre of France.

Two people replied to my question on twitter with answers they’d got from asking AI. And I’ve since asked Mistral, Grok, Claude, ChatGPT, and Gemini myself. I pay for three of these, one at the £200 per month level, and I instructed them to use the full power available to me. With the other two, I took my chances on the free tier.

I finessed my prompt a bit by making clear I only wanted to consider Metropolitan France and that I really did mean the United Kingdom, not Great Britain, and not England. Those people who replied to me seem to have done the same. So we had seven AI attempts versus the combined gold standard of Alasdair and the pretty good standard of me.

The AI was right three times, and wrong three times. No better than a coin flip. Grok was disqualified because while it got its answer right, it did so only by looking at the two AI replies I’d got on X. Its explanation for why was then riddled with errors.

ChatGPT found and used Alasdair’s LinkedIn post to estimate the population centre of the UK. It then used driving distance in its calculation, which it denied at first. But it still got the right answer of London, just by the wrong amount.

Importantly, all the AI answers contained serious errors. Some were obvious. But many more errors were only possible for me to unpick over a few hours because I had spent the five hours calculating the right answers myself. It took me a full day of work. This write-up has taken another day.

Most models used the population centre for England or Great Britain as this is more widely available online. While doing so, all insisted in their reasoning that they’d understood that Northern Ireland was part of the UK and that they weren’t doing what they were doing. So I had to figure out that they were lying, sometimes by looking at their thinking traces, but usually from their answer.

Many models used astonishingly old population centres for France. Sometimes the distance by driving rather than in a straight line was reported. In one case the model insisted it was doing the straight line calculation, but got the answer wrong when doing the maths. It then quite convincingly pushed back on my ability to do maths when I explained the error to it, before finally conceding it was wrong.

The answer that AI gave to Stefan Schubert was the best. The population centres were for the UK and France, just a few years or decades out of date. A reasonable assumption of the centre of London and Paris was used. The distance calculations were correct based on this.

It ruined things a bit at the end by saying the population centre of the UK was in Leicestershire, which it isn’t. I’m pretty sure this is due to its conflation of England or Great Britain with the UK and leaning on web search results about that. It then claimed that the result depended on the method and year. Unless we’re going back many decades that’s not true. I’m not spending the week I would need to track the movement of both population centres over centuries and I certainly wouldn’t trust AI to yet.

What I found most upsetting in the AI responses was how confident and convincing they were. I got brilliant discussions of demographic shifts in both countries due to deindustrialisation and centralisation vs. decentralisation. In some cases this was due to context from my previous queries, and in some cases it wasn’t. Most of the time these were superficially interesting but ultimately factually incorrect or irrelevant.

The complexity of the bullshitting was impressive. In one case, despite insisting it had calculated the straight line distance, the AI model had calculated the driving distance from the centre of population of the UK to London. To back up its lie under questioning it found a village that was that distance away in a straight line and claimed it as the population centre of the UK. I’d have nearly fallen for it if I hadn’t done my day of work in advance.

At times, with all the models, it felt as if them just getting the answer right would have been much easier than the work they were doing to justify and obscure their errors. It reminded me of the stereotype of Oxford PPE graduates confidently bullshitting their way through tutorials having failed to do the preparatory reading. This stereotype is sadly quite close to reality and the skills transfer into senior positions in and around British government making it dysfunctional. Think of Boris Johnson failing to read his briefings while foreign secretary, or the infamous picture of Britain’s Brexit negotiators seemingly turning up to their first meeting with the EU with no notes of preparation. Three of the four people just mentioned studied humanities at Oxford University, but maybe it’s just a coincidence.

I really hate that part of British culture. I nearly quit my degree in my second year because of it, and I was at Imperial College studying Physics, a setting where you might expect tutorials to be less amenable to bullshit. I am very glad that my personal tutor agreed with me while talking me out of quitting and suggested I take the chance to know my enemy. He promised it would be better on my exchange year in Paris and he was right.

Escaping, even if just for a year (though I would later briefly return for work early in my career), from the more lawyerly culture of the UK to a more engineering culture in France was life changing. I hope that the many Britons who find me arrogant and pompous do so in the same way as they consider “the French” — who I rarely found like that — to be.

Today I help run an AI company, the most recent of a chain of companies in data, AI, and tech that I’ve helped to succeed. We are much more engineering focused than lawyerly.

We are very pro AI, even if so far this blog post has not sounded like it.

I spend hundreds of pounds of my own money on AI every month and tens of thousands of pounds of my company’s. The trajectory of our spend is only upwards. I am not an AI luddite as I was, I still think correctly, about NFTs and Blockchain currencies. If you are writing software today without AI you are not going to make it. But I am also cautious and sceptical, as we all should be, about how AI is making its way into our work and our society.

In a few seconds, AI created content and reasoning that took me a whole day to prove near-valueless. I choose “near-valueless” not “wrong” because so much of the content was correct and so much of the consolidation of those correct parts into a seemingly coherent whole was also correct. But the output was near-valueless — perhaps some value could be plucked from the promising directions suggested — because no part was reliably correct.

Every piece, every fact, every logical step, every calculation, and every answer needed checking. And at every point of the checking, the odds were stacked against me. I was being lied to, tricked, and misled, over and over again. I got a taste, perhaps, of what it must be like to lead a tutorial at Oxford full of Britain’s future leaders, just with a lot more maths and geography and a lot less history and politics.

This bought me back to two pieces of writing I have been urged by people in and around the UK government to respond to this week. They make, I was told, some useful challenges to points that I have made and which they, or some others in and around Whitehall, had started to find convincing.

I have responded to Tim Leunig’s piece, inspired by a recent trip to the Bay Area, on Northern City economies on X. I hope I did so constructively and politely. The piece partly builds on work I have done over years to build tools that show population and population density for anywhere in the world. Most recently I refined and then used these tools to show that Great British big cities no longer lag similar cities in Northern Europea or North America in terms of population density.

His core argument as I understand it is that North English cities like Leeds, Manchester, and Liverpool, and the vast swathe of urban England between them, should allow more building, to get office prices down.

While I agree that more building is better, the argument relies on a fact that office rents in Manchester are the same as in Mountain View, California. And I don’t think that’s true.

Perhaps Northern rents could be cheaper still, but I know that in prime central Leeds my rent is not a concern. Whether it is 20% cheaper than the Bay Area or 50% cheaper, or maybe even cheaper yet, I just don’t think about it. I could move to Bradford or Huddersfield for even cheaper rent if I wanted, but since the cost of rent is of so little concern to me, I won’t. AI seems to back me up.

I note, of course, the irony, in me using AI to generate tables of comparative rent between North England and the Bay Area. I would trust someone like Jon Neale to have done the work needed to check the AI on this. And it is partly because of his knowledge in this area, far greater than mine or AI’s, that I have had many excellent discussions, and changed my mind on quite a lot, on how we might return the great industrial cities of Britain — his particular interest is in Birmingham — to prosperity.

The second piece I was urged to read and respond to was Stephen Bush’s newsletter, inspired by a recent trip to Philadelphia. It contains thoughts on the Transport Theory of Everything he kindly associates with me. Improving transport is always in my top three ways to make Britain’s second cities rich again, along with more public R&D spending, and devolution.

He asks, naturally, why, if public transport is so important to prosperity, Marseille is no more productive than its English equivalents?

The problem is that Marseille is more productive than its English equivalents. Using OECD data for functional urban area geographies, the best geographies for such a comparison, Marseille’s GDP per hour worked is 20% higher than Leeds and Manchester, and 30% higher than Liverpool, Sheffield, Birmingham, and Newcastle. In GDP/capita terms, and this is a reasonable measure since France achieves part of its higher productivity by excluding lower earners from the labour force, Marseille’s advantage shrinks, but remains at between 10% and 20%. These are all big numbers in countries struggling to achieve any sustained economic growth at all.

There are big and interesting — complex and nuanced even — debates to be had on whether public transport is more “how you spread prosperity”, or, as is closer to my view and a friend who is a much better economist than they think and who was pilloried on social media for suggesting the same, more that public transport is how you accelerate the development of prosperity in the first place. The answer is, clearly, some complex and nuanced balance of both, interacting with many other factors within a complicated system.

One of these factors will be the similarly complex and nuanced debates to be had about how important universities are to economic prosperity via education and research & development and in what balance, and to what extent, venture capital generates higher economic growth. We might want to extend our thinking to the role of local and national government in stimulating venture capital investment if it does.

I have had many such debates. The most valuably challenging week of them was eight years ago in Lille, the twin city of Leeds. It is a city economy barely more productive than its English equivalents despite excellent public transport and fantastic high speed rail connections within Europe. It is a fine city in which to argue against a transport theory of everything.

Everyone at the discussions shared a set of well-earned facts and opinions. We ended up sharing improved opinions that Lille was an example of favouring transport and state intervention too heavily in regional economic development and that Leeds was an example of favouring skills and low taxes too heavily in regional economic development. We probably all felt, while careful not to always see greener grass elsewhere, that there was probably a sweet spot in the middle that spoke Dutch. Our later meetings involving economists from Lille’s twin city of Rotterdam were fascinating in this regard.

Within the UK, I’ve had similarly good discussion in Newcastle, Glasgow, Liverpool, Birmingham, Manchester, and of course in Leeds. I would love to have more similar good discussions in the UK in South East England, where our national government and national institutions are. But they are hard to find, because for them to be usefully complex and nuanced we need to have a more shared view of the world and we rarely seem to.

As an example, in the Financial Times piece we are urged to look to Philadelphia as an American city with “a residual tram system” as if to dismiss its public transport advantages over British equivalents. Philadelphia has much more transport than this. At the 20km radius at which Manchester and Philadelphia both have a population of 2.7 million, Philadelphia has well over twice as many metro and train stops (Manchester has no metro), and over twice as many tram stops, though Philadelphia’s trolley system may confuse the counts here. Compare Philadelphia to Leeds and the gap is even more extreme.

That said, I wouldn’t argue that it is substantially in public transport where American cities gain the greater connectivity that permits their greater benefit from agglomeration benefits. Anna Stansbury’s work, heavily building on mine, suggests that it is in road connectivity that most US cities gain most of their advantage over their British equivalents. This has been my experience, albeit limited, of the USA.

There are so many interesting further discussions to have. The USA’s slower return to the office, which in turn leaves offices vacant and makes Tim’s point about comparative rent less wrong than it would otherwise be, combined with its excellent productivity performance since the pandemic, suggests that agglomeration may be over-valued. The emergence of self-driving cars may radically change how we achieve agglomeration even if it does still matter.

And of course transport is just a small part of a large set of complicated systems interacting in complicated ways that no-one can ever measure more than a small part of, and even then only poorly. But there is little point in having such discussions if at the foundational level we find so little ground for agreement.

I am infamously negative about calls for more complex and nuanced discussions in UK public policy. I own the domain fucknuance.com which I forward to my favourite sociological paper on this topic. But this is not, as many imagine, because I dismiss that the real world is complex and nuanced.

It is because people in those debates in Britain so rarely seem to have done the work needed to set the foundation for having fruitful complex and nuanced discussions. For someone to usefully be part of a complex and nuanced discussion they need to bring a set of well-checked and informedly-held simplifications which others can earn trust in them having assembled in a fair way, through diligent work, at risk to their own reputation, biased only by their own well-documented accumulated prejudices. One reason I find anonymous accounts online so wasteful of my time is that they can so rarely come close to meeting this bar. But the discussions I have with national government and national institutions in Britain rarely get close either.

I think it quite plausible that those on the other sides of these debate think the same about me of course. Adding more intelligence to such debates feels unlikely to help much.

Which brings me back to AI.

In so much British policy discussion, there is an appeal to great complexity and nuance before we’ve achieved much agreement at a more basic level. It feels, so often, more like tutorials at Imperial College and debates at the Oxford Union than the kind of rigorously blunt experiences I enjoyed in France. And of course I am exaggerating the difference.

In so much of what I deal with from AI today, I feel the same. AI can generate in a few seconds what will take me two days to check. And when the checking is done, it has cost me days of time, and provided no value. We are adding intelligence to debates that lacked something else.

I believe that AI will improve substantially. I am sure that with better prompting and with increasingly complicated AI systems of dozens of interacting agents interacting with each other, mediated by complicated documents full of instructions, rules, and guidance, it will do better.

Maybe skipping ahead of the foundations and getting into the complexity and nuance of the systems is the right approach. I entertain the idea that I could be wrong much more than my detractors imagine. But for now, AI is generating much more work for me and my teams than it is clearing. We are making faster progress in some areas, but losing all that advance and more because of the huge quantity of extra work checking the output of ever more convincing AI that emerges elsewhere and finds its way to us.

The answer of many people to these failings of AI is to improve the prompts, to add more guardrails, to pile on more layers of complexity and nuance to a system whose output we rarely ever check with the rigour that is needed to establish shared facts that lead to productive discussion. It feels very similar to debates in British policy and economics. And we need only look at the outcomes of those debates in terms of Britain’s comparative economic standing in the world to encourage ourselves to consider a different path.