Should you tell your boss how much of your work is AI?

AI Makes Workers Look More Skilled Than They Are: New Data

Disclose when AI materially shaped the work. A Use.AI survey of 9,684 workers, reported September 4, 2026, found 64% used AI to complete work they couldn't do alone and 52% say it makes them look more experienced than they are. Dorie Clark's The Long Game names the risk: reputation compounds like capital, and an undisclosed skill gap is borrowing against it at interest you will pay during your next live test.

I build AI systems for a living and coach executives who use them daily, and one number this week still made me stop. Use.AI, a platform that bundles several large language models into one subscription, surveyed 9,684 working adults across the US, UK, Canada, the EU and Latin America; Euronews reported the results on September 4, 2026. Sixty-four percent said they had used AI to complete work they could not have done independently. Fifty-two percent said AI makes them appear more experienced than they actually are. Thirty-nine percent had submitted AI-assisted work without mentioning it, and 30% had accepted praise for work the technology substantially produced. Nineteen percent said AI-assisted work contributed to a promotion.

The number I keep returning to is quieter: a quarter of respondents worry their employer now believes they are more capable than they really are. That is not a productivity statistic — that is a reputation running ahead of a skill, and Dorie Clark's The Long Game gives the cleanest account of why it ends badly. Clark describes careers as waves you move through in sequence: learning, creating, connecting, reaping. The compounding that makes a career durable happens in the learning wave, during what she calls the deceptive phase, when effort is invisible and progress looks like zero. AI collapses that sequence. It lets you reap — the polished analysis, the praise, the promotion — while skipping the invisible accumulation underneath. The survey's 19% who converted AI-assisted work into promotions are reaping a wave they never surfed. On Clark's seven-year horizon, that debt always comes due, usually at the worst possible moment: a live board Q&A, an outage, a negotiation where no model is within reach.

My own disclosure rule matches the one Use.AI's chief executive Ihor Herasymov proposed in the Euronews piece, and I held it before I had his words for it: disclose when AI materially shaped the substance of the work. Not every autocomplete, not every rephrased email — that would be theater, and Herasymov concedes a blanket rule would collapse as these tools embed into ordinary software. But if a model generated a significant part of an analysis, a recommendation, a deck, or code, I say so, and I stay responsible for defending it. In three years of doing this with clients, disclosure has cost me exactly nothing. What it buys is that nobody ever discovers a gap I hid.

The practice I give coaching clients is a monthly twenty-minute review I call the Reputation Gap Audit. First, list every output you shipped this month that you could not reproduce without AI — the survey's 64% test, applied to yourself, and the step is done only when the list is written down rather than estimated. Second, for each item, ask whether you could explain the reasoning and catch a wrong answer in front of a skeptical audience; mark the items where the honest answer is no. Third, disclose the marked items' AI assistance in your next one-to-one, framed as tooling and verification rather than confession. Fourth, pick one marked item and rebuild that capability deliberately over the following month — Clark's version of this is 20% time, protected learning hours invested from strength rather than desperation — so the gap closes instead of compounding.

Name the limits honestly. This is vendor research: Use.AI sells AI subscriptions, and a company with that incentive publishing a survey that normalizes heavy AI use deserves a raised eyebrow even when the findings ring true. The data is self-reported, which in my experience understates embarrassing numbers rather than inflating them. And disclosure norms are not symmetric: I have seen organizations where admitting AI assistance quietly moved someone down the promotion list. If yours is one of them, the culture is the problem to fix before you volunteer as its test case. Economist Amar Bhidé argued in a Project Syndicate essay republished on September 1, 2026 that AI dependence could end up reducing output per worker — organizations that punish honesty about AI use are exactly the ones that will find out the hard way.

The reason this matters more every quarter is that the autonomy curve is steepening. The same week the survey landed, OpenAI began rolling out GPT-6 Astra on September 3, 2026, pitching it explicitly on delegating professional work with greater confidence in the model's judgment. The better the models get, the less finished output reveals about the person who signed it. What still differentiates you is what Herasymov calls problem framing — defining the right question, challenging an assumption, catching the model when it is wrong — and those are precisely the muscles the hidden-gap loop atrophies. Play the long game: close the gap on purpose, or the tools will close it for you, in the wrong direction.

Sources: Euronews — Workers say AI makes them look more skilled than they are, Amar Bhidé — AI Addicts Won't Make Better Workers, OpenAI — GPT-6 Astra announcement.