# I Think You Might Be Fooling Yourself With AI

Louwrentius makes two claims and, in a post-script added after the Hacker News thread, states them plainly because almost nobody engaged with either. First: your perception of being more productive with AI may be wrong, and you have no way to check. Second: even granting the gain, the tool may not be worth what it costs once vendors stop subsidizing it. The externalities section that closes the post is labeled as his disclosed bias, not as the argument.

The frame is Feynman's line from the 1974 Caltech commencement address on cargo cult science: the first principle is that you must not fool yourself, and you are the easiest person to fool. Louwrentius's point is that maintaining that kind of integrity on a personal level is unusually hard here, because the tools feel unreasonably capable and using an LLM *feels* more productive while you are doing it.

## Why you can't check

The structural argument is short and does not depend on any study. You cannot go back in time and do the same task without the tool, so there is no counterfactual to compare against. Experiments are blind or double-blind precisely to keep this kind of bias out of the result, and blinding yourself against a tool you are actively using is not something you can arrange. This is why he is unmoved by personal testimony, including testimony about apps that would never have been built otherwise: the testimony is exactly the thing the method exists to exclude.

The one measurement he leans on is METR's study of experienced developers, where participants believed they had completed tasks faster with AI and were measured at roughly 19% slower. He is careful about it in the post-script, and the care is the interesting part. He acknowledges METR's February 2026 uplift update, which found more productivity, and passes along the authors' own qualifications — they paid participants considerably less, and they ran into developers who did not want to code without AI. What he keeps from the original study is not the direction of the effect but the gap between what developers thought and what was measured. They were measurably fooling themselves, which is the claim, and the follow-up does not touch it.

He also refuses one popular piece of evidence outright. Mass layoffs attributed to AI do not demonstrate productivity gains; Sam Altman himself called many of them AI washing. His reading is that these are ordinary corrections after over-hiring in an economy that is not doing well, and that some layoffs are specifically freeing up budget to buy AI resources — firing people to fund a technology that may never replace them, driven by fear of being left behind.

## The price you are not paying

The second claim is about what happens when the subsidy ends. Citing a TechSpot piece, he notes that a $200 subscription starts losing the vendor money at around 11% utilization, and that using it fully would consume roughly $14,000 worth of API-priced tokens. His question is whether the perceived productivity gain would still feel worth it at $2,000 a month rather than $200. A $200 bet is small enough that few people interrogate it; the interrogation is what he thinks is being deferred.

## The disclosed bias

The last section is explicitly labeled as his own bias against AI, which is a structural courtesy the rest of the piece earns. He is an AI skeptic, influenced by Ed Zitron's reporting, and expects AI to run out of money and be switched off. His challenge to the optimistic case is a question about arithmetic: close to a trillion dollars invested, and where does the money to recover it come from. If a hyperscaler absorbs a frontier lab the economics do not change, they just move onto a bigger balance sheet.

He then lists the externalities he says he can't look past — energy usage against climate change, data centers imposed on poor communities, large-scale appropriation of copyrighted work, the enshittification and "enslopification" of the internet, memory hoarding driving up prices. His conclusion is that the tool cannot be separated from these, so he does not use AI, on moral and ethical grounds.

Read strictly, this section does not support the two claims and does not need to. The productivity argument stands or falls on whether you can validate your own experience, and the cost argument on subsidy arithmetic. He put the ethics last and named it as bias rather than mixing it in, which is more than the genre usually manages.

The post-script records what happened next: the HN comments were mostly dismissive without engaging, and nobody offered anything beyond further personal anecdotes, which is the exact evidence class the post rejects.

## Where it fits

This is the epistemic entry in the vault's skepticism cluster rather than the engineering or economic one. [[anti-llm-discourse]] maps that cluster; louwrentius straddles the economic slot (the subsidy argument, in the same family as Zitron's) and something narrower that the map doesn't yet have a slot for, which is the claim that self-reported productivity is not admissible evidence about a tool you are using.

The self-assessment failure has an empirical counterpart already in the vault. [[cognitive-debt]] is the same gap measured on comprehension rather than speed: MIT found 83% of LLM-assisted writers could not quote a line of what they had just produced, and Anthropic's randomized trial found AI-assisted engineers finishing at the same speed with markedly worse follow-up comprehension. [[dont-outsource-learning]] collects the converging studies and adds the finding that posture, not tool access, drove the split. In all of these the artifact looks fine and the person's estimate of their own state is wrong, which is louwrentius's thesis arriving from the learning side.

On the money, [[local-ai-is-not-opus]] is the practitioner's version of the same worry from someone who does use the tools: Alex Ellis points at Copilot's move from flat pricing to tokens and at Uber's $1,500 per-developer-per-tool cap as evidence that the subsidy is real today and not promised tomorrow. He draws the opposite operational conclusion, buying hardware rather than abstaining, from an almost identical reading of the economics.

Post: [louwrentius.com](https://louwrentius.com/i-think-you-might-be-fooling-yourself-with-ai.html), with the [HN thread](https://news.ycombinator.com/item?id=49021843) that prompted the post-script.