The AI experts worth listening to are the ones who make it easiest to catch them being wrong. I don't want a prophet. I want the working, the limitations and the bit they'd rather leave out of the demo.
That's my filter for a subject where certainty can be sold separately from competence. If someone knows exactly what AI will do to every business, they're either remarkably gifted or selling a workshop. The workshop usually has better availability.
My default assumption is that the official story is the version somebody paid for. AI labs want subscriptions. Consultants want transformation projects. My suspicion is that some professional sceptics have discovered an equally comfortable business selling permanent disappointment. Different sermon, same collection plate.
None of that makes their arguments wrong. It makes evidence rather more interesting than confidence.
For a founder who isn't technical, the temptation is to outsource judgement to whoever sounds most fluent. I think that's precisely backwards. We're looking for people who expose enough of their reasoning that we can disagree intelligently, not people who make disagreement feel like being left behind.
The useful bit is the embarrassment
Simon Willison describes adding whole features to software with AI without reviewing how they'd been built. It worked surprisingly well, but he found himself getting lost in his own projects. Each extra feature became harder to reason about because he no longer had a clear picture of how everything fitted together.[1]
That's useful. Not because it proves AI coding is rubbish, but because it identifies a cost that a successful demonstration can hide: the thing works, and the human understands it less.
The term is cognitive debt. In plain English, you've acquired software faster than you've acquired an understanding of it. Willison connects his experience to that problem rather than pretending that working output settles the argument.[1]
I find that a much better reason to pay attention to someone than a screenshot of a miraculous app. A building can look finished before anyone's checked the plumbing. The opening ceremony is a poor time to discover the toilet flushes into reception.
There's a business question underneath the joke. If an AI-built tool becomes important, who can explain it, change it and recover when it breaks? A demo doesn't answer that. Someone willing to describe losing the plot gives us a way to ask it.
Ethan Mollick's experiment with two AI systems is useful for the same reason. Claude prepared an email draft, while ChatGPT actually sent an email to his colleagues. He'd previously given ChatGPT permission to send; Claude had been told to ask first.[2]
The interesting difference wasn't eloquence. It was permission.
I don't read that as proof that one brand is safer than another. It's one person's experiment, and the settings explain the outcome. That's exactly why it's worth reading: we can see what happened and what would need to change. The useful lesson isn't that AI has become your colleague. It's that your colleague can now send the email before you've finished admiring the draft.
A number isn't a verdict
METR, which evaluates AI systems, provides an awkward counterweight to impressive personal experiments. In an earlier controlled study, 16 experienced developers worked on 246 real tasks in software they'd contributed to for years. With the older AI tools in that experiment, tasks took 19% longer. The developers still thought AI had made them faster.[4]
Before anyone turns that into a victory lap against AI, the original study page now explicitly says those historical results no longer reflect the current impact of the tools.[4]
That update matters as much as the number.
In its follow-up, METR found that the measurement itself had become troublesome. Some developers didn't want to participate if it meant working without AI, and some withheld tasks they didn't want to do unaided. The newer results suggested speedups, but the researchers judged their data an unreliable guide to the current effect.[3]
In other words, the people and jobs you manage to measure can change the answer. You don't need to understand the mathematics to understand the problem: a restaurant review is less useful if everyone who enjoyed dinner refuses to fill in the form.
I rate that willingness to question the measuring stick. It doesn't hand us a satisfying universal verdict. It does stop an old result becoming a permanent belief simply because it suits our side.
Mollick's example shows what happened in a particular setup. Willison's account exposes a hidden cost. METR tests a narrower question systematically, then explains where the test struggles. They're different kinds of evidence, not contestants for the position of Official AI Oracle.
That's the mix I think deserves attention. Enthusiasm with something inspectable underneath it, scepticism that can change its mind, and an honest distinction between a demonstration and a dependable business process.
Call me cynical, but the beneficiary of unquestioning confidence is usually the person selling the next step. The beneficiary of visible working can be the person paying for it. I'd rather our attention went to the latter arrangement.
The best AI expert isn't the one who makes you feel certain. It's the one who makes certainty earn its keep.
Who gets paid when nobody asks for the working?
Sources
[1] https://simonwillison.net/2026/feb/15/cognitive-debt
[2] https://www.oneusefulthing.org/p/an-opinionated-guide-to-which-ai-b22
[3] https://metr.org/blog/2026-02-24-uplift-update
[4] https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study