In my last post I said that ChatGPT couldn't simply decide to become a builder. But a huge amount of work today is knowledge-centric and unconstrained by the physical world; law, accounting, writing, analysis, software engineering, domains in which AI is currently trying to assert dominance. So the question many of us are asking is: Is AI coming for my job? As a developer, it seems most sensible that I tackle the question from that angle.

Big Claims

It's difficult to get a handle on what AI is actually capable of, let alone what it could be capable of in the next ten years. Part of that difficulty is inherent to a fast-moving industry with very clever people doing very clever things. A bigger part is that the loudest voices in the debate are also the ones with the strongest incentives to exaggerate.

And that's putting it mildly.

Anthropic's CEO told us AI would automate software development within 12 months [1]. A year later, he said it again [2]. Microsoft's AI chief said all white-collar work would be automated within eighteen months [3]. Headlines warn of a white-collar "bloodbath" [4].

The claims don't end there.

Anthropic has a model called Mythos that it claims to be so capable at hacking that it's too dangerous to release to the public [5]. At face value this is a very virtuous sentiment, especially for a company that desperately needs to find revenue to add to its $200 million contract with the US military [6].

Small Evidence

The evidence suggests that developers aren't actually any more productive using AI tools and that they overestimate the gains they do achieve [7]. This is not limited to developers [8]; economists call it the productivity paradox [9].

(There is tentative evidence that this could be reversing as people start to use the tools more effectively but the gains are not yet massive [10].)

The investment riding on this is enormous: the five largest cloud providers have committed roughly $690 billion to AI infrastructure in 2026 alone [11]. The brute-force formula that produced the leap from GPT-2 to GPT-4 (more data, more parameters, more compute) appears to be hitting diminishing returns [12]. Progress hasn't stopped, but the gap between "what the money needs AI to become" and "what the evidence says it currently is" remains wide.

The Expert Gap

One thing that's become clear from watching people actually use these tools is that expertise still matters. A developer already knows what a prompt's output should look like. They can spot when the model tries to use code that doesn't exist, elegantly solves the wrong problem, or confidently asks to delete the production database. A non-developer doing the same thing has no reliable way to distinguish a working solution from a plausible-looking one. The model's conviction is indistinguishable from its correctness unless you already have the competence to judge.

I think this is what the replacement question misses. The tools are genuinely helpful for boilerplate, for exploring unfamiliar APIs, for rubber-ducking, for generating first drafts of tedious code, but the key point here isn't what it can do, it's that it's a tool. The limiting factor in software engineering was never typing speed. The true challenges were understanding the problem, navigating the codebase, making design decisions, and knowing which solution can scale. AI accelerates the mechanical parts of the job and they were rarely the bottleneck. The relevant question isn't whether AI can produce code (it obviously can) but whether anyone should trust that code without someone who understands the domain standing behind it. And if you still need that person, what exactly have you replaced?

Could we replace all programmers with AI? In my opinion, absolutely not. Could we replace many devs with fewer AI-assisted devs? Possibly, though even Anthropic's own CEO has invoked Jevons paradox to argue that cheaper coding might increase demand for developers rather than reduce it [13]. And that's only if the increasing costs of AI don't start outweighing the cost of another dev [14].

The Broader Scope

The same pattern plays out beyond software. AI can draft contracts, summarise case law, and generate legal memoranda that read convincingly. But a junior lawyer using these tools still needs a senior partner to catch the precedent that doesn't apply. The tool accelerates the drafting. It doesn't replace the judgement that decides whether the draft is right.

What about the more industries that rely on more creativity? Machines can move faster and lift more than humans, and we still watch the Olympics. But spectacle isn't really the point here. Writing, design, and art are supposed to provoke thought. Can AI really do that on its own? If you find genuine insight in a LinkedIn thought-leadership post, maybe.

But for all my apparent confidence that AI isn't replacing us, I do still worry. What if my boss or my boss' boss believe the hype? What if it's not just hype? After Kasparov lost to Deep Blue in 1997, he proposed Advanced Chess: humans and computers playing as a team [15]. For a while, it worked. "Centaur" teams outperformed both grandmasters and engines playing alone. But computers kept getting stronger and today the human's contribution is limited to shuffling the pieces on behalf of the machine. What if that's where we are heading with AI at work? What if the next model really is the game-changer?

I was lucky enough to be included in a trial for Codex. Their product looked exactly like Cursor and I couldn't see what they thought they were doing differently. It turns out their USP was 'pets'. Effectively a clippy that could write code for you. So I'm not worried every day.