A tough pill to swallow: AI is making outputs better and capability development worse.

Let me describe something that is happening where you work right now, at this very minute. You may not have noticed it yet, and that's kind of the point.

Someone on your team needs to write a tricky email. There's a difficult stakeholder, or a disappointed client, whatever. One of those messages where the words matter and getting it wrong has consequences. In previous years, this would have required them to sit with the discomfort for a bit, draft something, wince at it, redraft it, send it, and- in doing so- get incrementally better at the genuinely difficult skill of written communication under pressure.

Now they open Claude, describe the situation, get something back in sub-thirty seconds that is honestly pretty okay, maybe tweak a word or two, and send it.

The email is fine. Better than fine, probably. Maybe it's even really good. And the person who sent it has learned absolutely nothing.

Welcome to the quiet productivity paradox nobody is talking about yet: AI is making outputs better and capability development worse, simultaneously, in ways that are almost completely invisible on a dashboard. We are outsourcing the cognitive struggle that builds skill while keeping the output that makes it look like everything is fine. It is, as a strategic choice, a bit like taking the escalator every day and wondering why your legs aren't getting stronger.

There's a body of research in cognitive psychology called the generation effect, first documented by Slamecka and Graf in 1978, which has been quietly devastating to our assumptions about learning ever since. The finding: we remember and retain things significantly better when we generate them ourselves than when we simply receive them. The struggle- the effort of finding the word, constructing the argument, working out what you actually think- is not the inefficient bit you want to optimise away. It is the bit where the learning happens.

Strip out the struggle and you strip out the encoding. The email gets sent. The next difficult email is just as hard as the last one. Over time, the gap between "outputs this person produces" and "things this person can actually do" gets wider and wider, and nobody measures it because the outputs keep looking fine.

Philosophers Andy Clark and David Chalmers have a framework for this called the extended mind thesis- the idea that our cognitive processes don't stop at our skulls but extend into the tools we use. A notebook, a calculator, a smartphone. These aren't just aids to thinking; they become part of how we think. Which is wonderful, until you consider the flip side: when the tool is removed, the thinking it was doing doesn't automatically come back to you. The capability lives in the system. If the system goes down, or gets expensive, or changes its terms of service, you find out very quickly what you can and can't actually do.

AI is the most powerful extended mind we have ever had access to. It is also, if we use it carelessly, how an entire generation of knowledge workers ends up structurally dependent on a tool for capabilities they never fully developed.

The philosopher Michael Polanyi spent a lot of time thinking about the kind of knowledge that resists being written down. His most quoted line is: ‘we know more than we can tell.’ What he meant was that real expertise- the kind that makes someone genuinely good at their job- is substantially tacit. Built through experience, and expressed in judgment calls that the expert often can't fully explain, even to themselves. You know how to read a room in a way that cannot be reduced to a checklist. You know when something feels off in a strategy document in a way you'd struggle to articulate in a performance review.

Tacit knowledge cannot be prompted or offloaded to a model. It develops through repeated exposure to real situations with real stakes, and it is precisely the kind of knowledge that separates someone who is competent from someone who is genuinely capable.

The risk with AI-assisted work is not that it helps people do things they couldn't do before- that bit is the exciting part. The risk is when it helps people produce outputs that look like they require tacit knowledge they don't yet have, in ways that prevent them from ever developing it. A junior strategist who uses AI to produce polished strategic thinking before they've learned to think strategically is not getting a head start. They are getting a very convincing disguise. And here is the uncomfortable bit, courtesy of Kruger and Dunning's 1999 research on metacognition: the people least likely to notice this are the people it's happening to. The skill required to do something well and the skill required to evaluate whether you've done it well turn out to be the same skill. Which means the people most likely to accept an AI output uncritically are the people who most need to be developing the ability to critique it.

This isn’t a character flaw, but rather a structural feature of how expertise develops, and it is precisely the feature that AI adoption strategies are currently ignoring.

None of this is a reason to slow down on AI adoption. I want to be clear about that because this is the kind of argument that gets co-opted by people who were going to resist AI anyway and are just looking for academic cover. The technology is not the problem; it is genuinely remarkable and the organisations that figure out how to use it well will have a meaningful advantage over the ones that don't.

The problem is the assumption that better outputs automatically mean better people, which, in the AI age, they absolutely don’t. Better outputs mean… well, better outputs. And sometimes that's enough! Not every task is a development opportunity, and not every email needs to be a character-building exercise. The question is whether we are being intentional about which cognitive struggles are worth preserving and which aren't, or whether we are just optimising for immediate output quality and hoping the capability takes care of itself.

Spoiler alert: It won't. Capability takes care of itself about as reliably as fitness does when you stop exercising. Which is to say: not at all, and the decline is gradual enough that you don't notice until you really need the thing and it isn't there.

The practical implication is that AI adoption strategies need a capability dimension that most of them currently lack. Track what people are using AI for, yes. Track efficiency gains, yes. But also ask: what are people developing or not developing as a result of how they're using it? Are they becoming more able to evaluate outputs critically- which requires building underlying domain knowledge- or less so?

Ultimately we’re asking, can they do the thing when the AI isn't there?

If the answer to that is no, then I’m afraid to tell you that you’re building your house on someone else’s IP, and whilst I don’t know a lot about construction, that sounds pretty unstable to me.

Alice Veitch is a Learning & Development Lead with a background in behavioural science. She has led learning strategy across EMEIA and is a respected disruptor in the field.