TL;DR
Ask employees whether the technology makes them look better at their jobs and most say it does. In a poll of 9,684 working adults spanning Britain, North America, continental Europe and Latin America, 52% reckoned the technology now makes them seem further along in their careers than is really the case. Nearly two-thirds had leaned on the technology for tasks they could not have finished unaided. The commissioning caveat matters: Use.AI, which sells bundled access to language models, paid for the research.
Results running ahead of ability
The figures describe a widening gap between what people deliver and what they can do. Roughly 43% had accepted duties they felt underqualified for. About 35% would find parts of their present role difficult to perform unaided. A quarter privately suspect their employer rates them too highly.
Their managers are largely unaware. Around 39% had handed in AI-assisted material without flagging it, and 30% took the credit when the praise arrived. Very few workplaces require anyone to declare it, so no rule is being broken — but appraisal is happening blind to how much of the work was actually delegated. One in five went further, saying such output played a part in getting them promoted.
The company’s own prescription
Ihor Herasymov, who co-founded Use.AI and runs it, does not favour declaring every interaction — unworkable, he argues, once these tools sit inside everyday software. His threshold is substance: if the machine wrote a substantial share of the code, the slide deck, the recommendation or the analysis, say so.
“The purpose should not be surveillance. It should give managers enough context to judge both the work and the human contribution to it,” he said, adding that the person putting their name to a piece of work remains accountable for grasping it and standing behind it.
On assessment he is more useful still. A finished deliverable, he suggests, no longer tells you what it used to about its author. What employers actually need to establish is whether someone can walk through their reasoning, catch an error in a machine-generated answer, make a call on partial information, and react when the system returns something wrong. He picks out problem framing as the hardest thing to read off a polished document: can this person frame the question correctly, push back on an assumption, and explain why one option beats another.
Looking forward
Herasymov concedes the uncomfortable finding deserves attention — dependency starts, he says, where someone produces an answer but cannot reliably tell when it is wrong. No settled model exists yet for grading any of this. UK employers can read it against the Yorkshire figures published the same week, where 47% of firms reported an AI skills gap in entry-level hires. Both point at the same redesign: interviews and appraisals now have to probe understanding, because the output itself no longer separates candidates.