Don’t 'never-skill' yourself with AI

If you rent your abilities from AI companies, other people can rent them too

The other evening, my wife and I wanted to watch a noughties TV comedy show. But we couldn’t, because we didn’t have a subscription to the one streaming service that was showing it. It was frustrating: both of us had previously owned DVDs of this exact show, which we’d got rid of years ago.

In other words, we’d have to pay to do something that previously we could have done immediately and for free.

For many, this is an increasingly common feeling in the AI era. Previously, I’d make plots manually in Python or R, by adapting my existing templates. But in the past year or so, it’s felt easier to just tell Claude or Codex to make it for me. Each time, I pay some tokens and get an output – but with no real improvement in my own skill.

If you rent your abilities, other people can rent them too

There is growing speculation that AI will eventually take over most jobs. Often, this view is promoted by companies that have a strong incentive to make investors believe that AI will eventually take over most jobs. And without much reflection when predictions about the end of human radiology, software engineering etc don’t come to pass.

But that doesn’t mean there is no risk. The more people outsource their skills, knowledge and abilities to AI, the more they will be making themselves replaceable.

A recent opinion piece in Nature Medicine warned of the risk of ‘never-skilling’, where medical trainees fail to develop the fundamental skills and judgement required to progress later in their careers:

trainees who rely on AI during the early formative years of clinical education may fail to develop the foundational reasoning skills that safe, independent practice requires. We refer to this as ‘never-skilling’, distinguishing it from deskilling in experienced clinicians and from mis-skilling, in which uncritical acceptance of AI errors leads trainees to internalize flawed clinical knowledge as fact.

The same is true of other industries. If your job application, or PhD proposal, or Substack article is a cut-and-paste from ChatGPT, why would anyone commit resources to support your future work? It’s much cheaper and easier for them just to go directly to the source you’re renting from.

You may have seen the below graphic recently. It shows exam results from a class at Brown University. The midterms were take-home exams (and hence vulnerable to AI); the final exam was in person. Bar a handful of students, it appears that many were renting their knowledge and skills. The final exam is likely to be the first of many situations where this will incur a cost.

What’s your moat?

In the world of startups, where I’m spending most of my time these days, people often talk about the ‘moat’. What is it that you can do that others can’t? What if Fable 6 or GPT-6 appears and is able to build any software or discover any science or develop any method, thereby destroying any advantage you may have been able to offer?

The grandiose claim that an all-powerful AGI model is just around the corner is not especially scientific, because it’s near impossible to falsify. If models can’t reliably handle governance, workflow, auditability, or domain nuance, the response is often ‘well, they’ll improve’. If models struggle with reliability or stability at scale? ‘Well, just wait for the next generation of agents/recursion’.

After all, remember GPT-5?

But that doesn’t mean the concept of an AI-proof moat is pointless. Startups need to think about it, and so do others, from students to scientists. A recent piece by John Drake at the University of Georgia discussed what the democratisation of certain skills with AI might mean for competitive advantage in the field of ecology:

When analytical capability is abundant and freely shared, comparative advantage will shift to those who control the means of data production: the experimentalists, the field ecologists and the researchers embedded in long-term monitoring programs... It does not require that AI tools fail or stall. But requires only that they succeed so thoroughly that computation ceases to be a distinguishing capability.

Controlling the data and the understanding of that data – especially for complex, nuanced topics – is one area where the brute force AI invasion is a long way off. As I’ve written about previously, some areas of science and research are less susceptible to scale than others:

A protein structure from one location can generally be compared like-for-like with a protein structure from somewhere else. But what it means to identify a disease case, or to intervene successfully, or to have a social interaction, can vary wildly between different settings. Heavily customised analysis and contextual understanding becomes essential.

The experience, relationships and intuition required to deliver on complex high-stakes tasks remain qualities that are difficult to encode. And this is the risk of never-skilling with AI. It’s less about a brilliant artificial mind swooping in and doing all the hard jobs; almost every week we hear stories of where attempts to remove human specialists have failed. Instead, the risk is that trainees will be tempted to rent their understanding, and with it lose ownership of their future trajectory.