The real effect of AI: from my desk to everyone else's future

A few months ago I said, in print, that the very idea of anything produced by "vibe coding" getting into production was frankly terrifying to me.

I stand by that. Mostly.

But I've also spent a chunk of my evenings this year building a backgammon app with an AI agent doing essentially all of the typing, because my wife and I like playing backgammon casually, the backgammon app I already have on my phone is riddled with all-too-frequent ads, and the rules are simple enough that it seemed like a reasonable field test. I've mapped out a detailed enhanced feature set and a faintly ridiculous revenue model for it too, obviously, because that's what I do — but the point is that both of those things are true of me at once.

I think the tooling is genuinely dangerous if you trust it blindly, and I use it constantly, and I think both of those positions are correct.

That contradiction is basically the seed of everything that follows. I started writing about it at the scale of "should I let AI write this function", and by the time I'd finished I was writing about who gets to own the next fifty years of the global economy. That's not me getting carried away with my own cleverness (well, not only that). It's that the same four forces keep showing up at every scale I look at, and once you've spotted the pattern it's very hard to stop seeing it wherever you look, especially when that happens to increasingly be in the news.

Those four forces, for reference, because I'll keep coming back to them (I promise not to turn this into a PowerPoint or listicle though):

• capability is what's actually possible now (and the people that capability resides in)
• agency is about who gets to decide how it's used (which is not as open as you might think)
• speed is how fast it's all moving (really fast in some places, glacially in others, but we only really hear about the fast bits from the AI boosters and Tech Bros)
• time is how much runway anyone has to adapt (which is vanishingly little in a number of key areas)

It's worth having that framing in your pocket, because it's the same frame whether we're talking about my ability to ship an app feature by Friday or a country's ability to keep its economy functioning as AI continues its "HULK SMASH!" rampage through the collective consciousness.

Speed isn't the same thing as quality​

I resent that I have to keep saying this, but here's where it starts, at the most boring and most personal level: a COO on LinkedIn recently posted that developers who spend three days writing "clean" code that someone else would ship in four hours with AI are no longer being rewarded for their diligence, they're being penalised for their slowness.

Friends, I had (and still have) opinions.

Being fast to market can absolutely be a good thing — first-mover advantage is definitely "a thing".

It is not automatically a good thing, especially if there are unrecognised security or performance issues in that code (which the AI probably also wrote the tests for, tests which therefore also don't catch the problem), and an outage or a data leak results in the kind of reputational and financial consequences that make "we shipped it in four hours" look like a spectacularly bad trade.

"Clean" versus "fast" gets treated as a value judgement about virtue, when it's actually just a set of questions nobody's bothered to ask properly: do your customers need it in four hours rather than three days? What do they actually need, as opposed to what they'll say they want if you ask them? What's "value" for this feature, for this client, for the (presumably paying) consumer at the other end who's going to have to use it?

I'm not against the tools. I couldn't be — I use them constantly, including for the things I'm building for fun in my spare time. What I'm against is using speed as a proxy for judgement, because it very obviously isn't one.

My fundamental — and evolving — advice is to use AI the way you'd use a team of very enthusiastic employees of variable quality, experience, and attention to detail: useful, occasionally brilliant, and in need of a manager who actually understands the work well enough to know when they've gone quietly off the rails.

Which brings me to the bit that actually worries me, because it's not really about speed at all...

It's all about judgement​

The thing that's disappearing when we start talking about shifting team headcount because of AI is judgement, and I don't think people have clocked how much of a problem that is.

For the last few decades, being good at a job like mine has mostly come down to two things: what you know, and how well you can apply it. AI is putting pressure on both of those simultaneously — knowledge is getting more accessible and compressible, and execution is getting cheaper and more automated. So what's left, once both of those are commodities?

Judgement is framing the problem properly, deciding what actually matters, and knowing when the output in front of you is just plain wrong, or subtly wrong, or technically correct but a terrible idea. That doesn't make the old skills irrelevant, but it changes their job — they stop being the differentiator and start being table stakes.

Here's the bit that I think gets missed in the "will AI take my job" conversation, which is otherwise a bit boring by now: everyone frets about junior roles disappearing as though that's simply a headcount problem.

Junior roles are the entry point into the whole system — they're how you build judgement in the first place. You don't arrive at "good instincts" by reading about it; you arrive at it by doing the boring, repetitive, slightly-too-hard-for-you work for a few years under someone who occasionally tells you you've cocked it up and coaches you on how to fix it.

If AI eats that layer of work, you're not just cutting cost, you're switching off the mechanism that produces the senior people you're going to need in ten years. That's not an engineering problem. It's a labour-market problem, and it's one that compounds quietly until the day you notice you don't have anyone left who knows how to make the hard calls.

I say this as someone who is, by any reasonable definition, one of the "senior people" this dynamic currently benefits. Which is exactly why I don't trust my own instinct to shrug it off.

The shape of the team is changing​

Most engineering organisations I've seen — including the ones I've helped run — are still shaped like a pyramid: lots of juniors doing execution at the base, a layer of mid-level people translating intent into delivery, a thin band of seniors making the trade-offs, and leadership on top setting direction. It's a good model, because it scales and it creates a pipeline.

It also depends entirely on there being enough entry-level work to justify that wide base, and that's the assumption AI is currently kicking the legs out from under — boilerplate, first-pass implementations, routine transformations: exactly the stuff that used to onboard junior engineers.

One version of where this goes is a diamond instead of a pyramid: a small intake, a dense, highly capable core, continued emphasis on senior judgement, fewer people overall doing more with AI as a multiplier. It's closer to how specialist units in other fields operate — recruit selectively, invest heavily, expect people to take on real responsibility earlier.

I don't love the "elite squad" framing, if I'm honest, because it ends up sounding like it was written by someone who's read too many Andy McNab novels and self-improvement management handbooks without ever having to actually manage the resulting attrition, but the structural logic holds up.

It still doesn't solve the actual problem though, it just relocates it. Even if you're relying on someone else to build the basic skillset so that you can suck them into your elite squad of coding badasses through your gruelling, SAS-style selection process, you still need a way to grow the next generation of people who can make good judgement calls, and with fewer junior roles and less exposure to real complexity early on, that path gets less obvious, not more.

When I think about the small team that built Google Web Fundamentals in four weeks flat back in 2014 — seven people, one of them a tech intern who ended up as our second most active contributor on the whole project — I think about how much of that came from someone with almost no experience being handed real, meaningful, slightly terrifying responsibility on day one, with people around them who could catch it (and them) if it went wrong. That's an incredibly hard thing to manufacture deliberately. It mostly used to just happen, as a side effect of having enough entry-level work to go round. Lose that, and you have to go and design the thing you used to get for free.

That's a bit of a problem for many businesses, because it means investing in things that don't deliver an immediate, tangible ROI. Capability development is slow and accretive. Training juniors is slower still. Mentoring takes time away from "doing", and creating structured learning environments requires effort and budget. None of it shows up immediately in delivery metrics, so it gets deprioritised.

Optimising for immediate output at the expense of future capability essentially means that the pipeline narrows, and basic physics tells us that if the pipe gets narrower while you try to force the same volume through it, pressure increases. In this particular plumbing-based analogy, the dependency on a small number of highly capable individuals increases, and at some point, something is going to go "BANG" in an unpleasant way.

Time is not on (y)our side​

Being "an old" now, the pace of change also worries me precisely because it is a fast-moving topic that affects very slow-moving processes, including education. Kids today are choosing their academic or vocational direction based on the future of several years ago.

If you’re 15 or 16, doing your GCSEs and thinking about A-levels, you are implicitly making decisions on an eight-year horizon once you factor in those A-levels and a subsequent university degree. Even a shorter path still implies a two- to five-year bet on where the world is heading.

That is a very different problem to the one previous generations faced, and they are not dealing with workplace uncertainty - they are making personally significant bets against a background of rapid, compounding change.

The world at the end of that eight-year window could be materially different from the one we see today, particularly if current rates of progress hold, so the conversation needs to broaden — beyond engineering workflows and startup tactics, to labour markets, education systems, and economic structure.

On top of that, the clocks in many organisations have also stopped agreeing with each other.

There's a version of this that's less about people and more about pacing, and it took me an embarrassingly long time to notice it because it's not the kind of thing that shows up on a roadmap.

Different parts of a modern organisation are now operating on genuinely different timescales.

Tooling and AI capability shift week to week. Delivery works in short, iterative cycles. Capability — actual skill depth in your people — builds up slowly, over months and years. Strategy moves slower still, revisited quarterly if you're lucky. Historically those horizons were close enough together that the whole system stayed roughly coherent. They're not any more, and the gap is what's actually behind a lot of the low-grade organisational chaos that gets blamed on other things.

You end up with strategy built on assumptions that quietly expire before you've finished executing against them. Teams investing in skills that turn out to be the wrong shape for the environment they land in. Output going up while alignment goes down — everyone's shipping more, but that's not the same as shipping the right things faster. The uncomfortable trade-off AI actually introduces isn't "we can go faster", which is the bit everyone's excited about — it's "we can go faster in the wrong direction with more confidence than we used to be able to". Speed goes up. Certainty goes down. Nobody puts that in the slide deck.

The shift this actually demands is from planning, which assumes a reasonably stable target, to something more like continuous recalibration, which assumes the landscape is moving under you, and somebody is also moving the goalposts that the team is running towards. That's a much less comfortable way to run a business than "did we hit the roadmap", but I think it's the honest one.

People have always made the difference​

I've talked about skills, training, and the people pipeline problem ¹, and I think there's a parallel shift in how we think about individual capability. People have never been identikit pieces that can be swapped in and out of a machine at will — it's always been far more complex than that.

The idea of the T-shaped individual has been useful for a long time — it's a model that I have recognised in myself for many years, as well as many of the most capable and interesting people that I have worked with. Cultivating a broad base of understanding across multiple areas with deep expertise in one or more specialisms still has merit, but it's not enough on its own any more.

AI is putting pressure on knowledge-based capability, and breadth and depth are both under pressure in different ways. AI can substitute for true breadth of knowledge, because LLMs can provide rapid access to information and synthesised knowledge across multiple domains. Shallow depth is easier to replace, because routine expertise can be automated away.

What is emerging is a different shape, albeit one that isn't handily encapsulated in the same western typographic metaphor. I will pick up on this in more detail in another essay, but to me it looks more like a tree - strong, wide roots supporting a trunk and expansive, branching knowledge and capability on top.

The exact metaphor is less important than the underlying idea that individuals need both deeper expertise in selected areas and sufficient breadth and adaptability to operate effectively across systems.

Judgement is what brings it all together and also serves as the key differentiator of individual value. Judgement is the connective tissue between areas of expertise — knowing which tool to use, when to trust the output, when to override it, how to balance competing constraints, and how to make decisions under uncertainty. None of these are new skills, but they become more central as other aspects of the work change.

Let's zoom out a bit​

Take a deep breath, because this is where it stops being just about my job and gets a bit loftier and vertigo becomes a real possibility.

I studied Economics and Politics at university, a very long time ago², and for most of my career that's felt like a slightly expensive but pleasant piece of trivia rather than something I actually use. Lately it's been getting much more of a workout than I expected.

Everything I've just described — capability concentrating, agency shifting, speed outrunning structure, time running out to adapt — doesn't stop being true at the boundary of my inbox or Teams chat. It's the same four forces across the "real world", at a different scale, and once I noticed that I couldn't really pretend I hadn't.

There's a framing that's been doing the rounds courtesy of Mo Gawdat — MAD, Mutually Assured Destruction, versus MAP, Mutually Assured Prosperity — the idea that AI could deliver something genuinely new in the form of a shared upside rather than the old Cold War logic of shared catastrophe.

It's a tidy idea, and that's why I don't believe it. Certainly not as a default outcome, because it assumes a symmetry and mutuality that isn't supported by anything in the way these systems actually get built and deployed.

The more honest question isn't "does AI lead to disruption or prosperity". It's who experiences the disruption, and who captures the prosperity — because those are demonstrably not the same group of people, and the gap between them is not narrowing.

Let's start with the fact that AI capability is not even remotely evenly distributed.

AI is often discussed as if “the future” is a shared destination — something we are all heading towards, more or less together — but that has never really been true, and it's even less true now.

At this point, it feels almost mandatory to quote William Gibson³:

“The future is already here — it’s just not evenly distributed.”

It’s one of those lines that risks being overused precisely because it is so consistently accurate.

The frontier models come out of a small number of organisations, mostly in the US and China, backed by capital and compute and political interest that most of the world simply doesn't have access to.

Economics is political​

For decades, developing economies have had a rough but real ladder to climb — industrialise, move into services, integrate into global markets, use lower labour costs to build capability over time. AI is aimed squarely at the rungs of that ladder: back-office work, customer support, routine analysis, data labelling. If that work automates away before those economies can build alternative capability around it, the ladder doesn't get harder to climb, it disappears.

And it's not just a between-countries problem — it plays out inside them too.

It’s easy to talk about “jobs being automated” in abstract terms, but it's much harder to think about what that means in practice. If AI automates away significant portions of the work that was building the ladder, the impact isn't just technological, it's social, economic, and deeply personal. These are not just "jobs" as numbers on a graph, they are economic foundations that support families, local economies, education, and upward mobility across generations.

Historically, technological advancement and automation has been disruptive, but it has often been accompanied by transition — old jobs disappear, new ones emerge. People move, adapt, or retrain. These days we see scarce few farriers shoeing horses, and a lot more mechanics fixing cars — although they too are being affected by electrification of vehicles and a reduction in what can be done by physical labour and the physical tools that have remained relatively unchanged for a long time.

The bigger concern with AI is that it compresses that transition period — the jobs of today end up disappearing faster than new categories of work for the jobs of tomorrow are created, and faster than people can realistically retrain. That creates a gap, not just in employment, but in identity, stability, and opportunity.

There is also a darker possibility, not that labour disappears entirely, but that it becomes even more commoditised. If AI handles the high-value work, what remains may be lower-paid, more fragmented, more precarious, more physically demanding, more dangerous, or more tightly controlled by platforms. In that scenario, countries that previously benefited from labour arbitrage may not move up the value chain, they may be pushed further down it, becoming more dependent and less empowered.

Governments trying to deal with this rapidly evolving landscape are not starting from a neutral position. They are dealing with existing inequality, limited resources, conflicting incentives, electoral cycles, and global competition. Even under stable conditions, coordinating economic transition at scale is hard, but under conditions of rapid technological change, it becomes significantly harder.

Going back to the geographical concentration of the companies building the fundamental AI capabilities that the world is coming to rely on, for the vast majority of the global population this means that the technology is controlled elsewhere, the pace of change is externally driven, their economic models are being disrupted from the outside, and they hav no obvious playbook for this.

For some countries and communities, the rise of AI represents a loss of economic role, a challenge to social stability, and a test of political capacity. Some countries may move faster, but others may lose the ability to move at all if they can't absorb the consequences.

There is a chance that developing economies may be able to leapfrog more developed ones — countries without legacy systems may be able to adopt new economic and technological models faster. We’ve seen versions of this before, with mobile phones becoming uiquitous and skipping landlines almost entirely, and digital payments skipping traditional banking structures. AI could follow a similar pattern, but it still depends on access to infrastructure, compute, education, capital, and platforms. Without that, leapfrogging becomes much harder, and things get more geopolitical.

Infrastructure is not neutral — who builds it matters. China’s investment in digital infrastructure through initiatives like the Belt and Road and the Digital Silk Road is often framed as development support, but it's not investment without agenda⁴, and it's building to something much bigger and more impactful in the long term.

Infrastructure defines standards, dependencies, influence, and data flows. Countries that adopt that infrastructure are not just gaining capability, they are also entering into a system of capability and control.

Without deliberate effort, the benefits of AI are unlikely to be anything close to evenly distributed, equitable, or in any way neutral. They will flow towards capital and away from those who lack it.

You keep using that word - I do not think it means what you think it means​

There's a lot of talk about AI "democratising" software development, and I think it's mostly wrong, or at least it's not democratising it in the direction people assume. For most of my working life, all you actually needed to learn to write software was access to a library and access to a computer — you didn't need to own either.

AI coding tools have free tiers, but they run out fast, and the pricing is drifting steadily toward metered usage as the AI companies try to actually turn a profit rather than subsidise adoption. If you want to opt out of that and run everything locally, you need hardware with a price tag that's got several zeroes on it before the decimal point.

So yes, more non-programmers are producing code — but that's "democratising access" only in the sense that anyone who can pay can now produce code, which I'd call explicitly commercialising access instead, and given how comfortable we've all become with the commercialisation of democratic institutions generally, maybe that distinction is becoming harder for people to even register.

There's a name I've been using for the version of this where it all lands in the same place at once — capability, access, and dependency concentrating together — and I've been calling it Asymmetric Corporate Expansion.

Asymmetric, because the advantages compound unevenly. Corporate, because it's companies, not states or communities, best positioned to capture it. Expansion, because once a lead opens up, AI lets it compound faster than any previous technology wave — faster than competitors can catch up, faster than regulators can respond, and faster than markets can rebalance.

It's tempting to assume that at least some of the organisations building these systems will act in the broader public interest. Some explicitly position themselves that way, but even within that group there are meaningful differences. OpenAI and Anthropic, for example, both talk about safety, alignment, and responsible development, but they have taken different stances on military usage, government partnerships, and deployment constraints. Those differences are not accidental or incidental — they reflect funding structures, governance models, strategic priorities, legislative leverage (or laxity), interpretations of risk, and, ultimately, incentives.

This isn't a criticism of any individual organisation, it's a reminder that these companies are not neutral actors, and they are not insulated from the pressures of capital, competition, or national interest, operating as they do within highly competitive economic and political systems. That means the trajectory of AI is not shaped by what is best for mankind or the planet that we live on, or even by what is technically possible, but by what is economically, politically, and strategically advantageous.

We're not looking at runaway sentient AI as the risk here, whatever the marketing decks (and the more excitable corners of the internet) would have you believe. The mundane version is worse, precisely because it's already happening: a handful of boardrooms ending up in a position to dictate the shape of global productivity, not through malice, just through reach and the ordinary, grinding logic of compounding advantage fed by the kind of money that expects to yield a substantial return.

Ultimately, the risk isn't that AI develops its own agenda, it’s that it inherits ours and then scales it.

So what would actually have to change?​

Yes, I know how this sounds, but I am not simply raging against the machine for the sake of it.

None of what I have laid out so far points to one fixed, inevitable outcome. It points to real, complex systems under real pressure, and pressure doesn't resolve itself in the direction you'd like just because you'd like it to — it's more likely to find the weak points that you ignored, overlooked, or tried to hide, and do something loud, explosive, and messy to them.

AI systems are not neutral in practice — they are built, trained, and deployed by humans, both individually and collectively. That means they reflect the data they are trained on, the objectives they are optimised for, the constraints and policies imposed on them, and the commercial and strategic incentives of the organisations behind them.

Even when the underlying models are broadly capable, the interfaces we interact with are not raw intelligence, they are curated, guided, and constrained. In some cases, they are steered quite deliberately — we’ve already seen glimpses of this in how different systems behave and what they are willing to answer, how they frame responses, what they prioritise or avoid, what they permit or preclude, and where and how they apply guardrails (or don't).

That shouldn’t be surprising, in fact it should be expected, because these systems are products, and products are shaped by market incentives. If the underlying incentives prioritise growth over accuracy, engagement over truth, profit over public good, or national interest over global stability and cooperation, then those priorities don’t just influence outcomes, they become embedded in the systems themselves.

All of this is already happening at a scale we have never experienced before, and moving faster than moderation, monitoring, legislation, or existing safety nets can adapt to. If a small number of organisations control the most widely used AI systems, then their assumptions, biases, incentives, and intentions don’t just affect their own products — they begin to shape how information is generated, interpreted, and acted upon more broadly.

If you wanted a more balanced outcome — something closer to that Mutually Assured Prosperity idea, treated as something you build rather than something that arrives — then I think you can tie it back to the four fources that I have been using to frame this discussion.

Capability must be decentralized. That means moving beyond metered, commercial APIs controlled by a tiny group of tech giants. It requires real investment in open, locally deployable models, and the access to the raw compute required to run them without paying a permanent platform tax. A good chunk of that must happen at the national and cooperative, international level if we want to avoid corporate hegemony.

Agency must be reclaimed. We can't audit systems we're not allowed to look inside. True agency requires absolute transparency in how these models are steered — including open documentation of the hidden system prompts, heuristics, and safety policies that sit between the user and the raw model weights.

Speed must be matched by structural resilience. The current competitive landscape rewards shipping features faster than they can be understood or moderated. We need counterweights — whether through legislative accountability or collective professional standards — that penalise platforms when mistakes scale faster than they can be corrected.

Time must be intentionally carved out. Societies, legal frameworks, and labour markets cannot absorb immediate, compressed disruption without systemic failure. We have to create buffer zones, allowing institutions the space to stabilise and retrain people before entry-level roles are automated away entirely — or we will face social and economic shocks of a wholly different order to any that have gone before.

I'd love to tell you I think that's the likely path, but it isn't.

The market does reward velocity over safety, explosive growth over equilibrium, and lock-in over distributed capability, and none of the actors best positioned to change that have much incentive to do so unilaterally. Diagnosing the shape of the problem clearly doesn't make the underlying incentives go away.

Here's where I ended up, for what it's worth from someone whose actual expertise is "has shipped a lot of software and is still tinkering with a moderately over-engineered backgammon app":

The real question was never whether AI is going to change the world. It already has, and comprehensively.

The question that's still open is whether anyone's both willing and able to put deliberate friction into how fast — and on whose terms — that change keeps happening.

If we all just stand back and let it run, the outcome won't be neutral — it'll simply reflect the raw incentives of whoever was moving fastest, had the most money, or happened to own the infrastructure everyone else was standing on at the time.

I still think the tooling is dangerous if you trust it blindly. I'm still going to keep using it every day, including on my backgammon app, because that particular exercising of cognitive dissonance turns out to be more or less the whole story (and is probably why us squishy-brained humans are such a brilliant and simultaneously terrible species).