45% of Workers Skip AI Checks? The Report Says 60%

The Line That Got Away

One line made its way through the French business press: 45% of workers don't check whether an AI tool's output is actually correct. It comes from a press release KPMG France put out in April 2025, summarizing "Trust, attitudes and use of AI," the study the firm ran with the University of Melbourne, surveying more than 48,000 people across 47 countries.

In that study, the number listed for France next to this exact question is 60%.

Both figures come out of the same house. The 45% is in the press release. The 60% is in the report. In France, it was the press release that got read.

Five Lines, Five Numbers

The France page of the report fits on a single infographic. One block is labeled "Complacent use of AI," and it lines up five behaviors self-reported by workers who use AI on the job, each with its own percentage.

In order: used AI inappropriately (46%), made mistakes because of it (52%), used it in violation of internal rules (45%), relied on an output without checking its accuracy (60%), passed off AI-generated content as their own (53%).

So the 45% is real, it's right there on the page. It's just attached to a different question, the one sitting directly above.

What the 60% Line Actually Measures

The wording of the item matters, because that's exactly what got lost along the way. The question asked was "how often, in your work, has it happened that you...," and the report counts anyone who answers "sometimes," "often" or "very often" toward that 60%. So the threshold captures a habit that recurs, not a constant failure. And the base is workers who already use AI at work, not the entire workforce.

In practice, that looks less like a scandal and more like a Thursday afternoon. A meeting summary an assistant wrote up and forwarded without a second read. A table copied straight out of a chatbot answer. A translation sent to a client because it sounded about right. None of that leaves a trace, and that's exactly what the study is trying to catch.

Zoom out to the whole world and the same item lands at 66%, a number the report spells out directly: "Two-thirds of employees report having relied on AI output at work without critically evaluating the information it provides (66%)." France sits a bit below the global average.

How We Could Be Sure

An infographic doesn't read like a table. The numbers float next to the labels, and nothing in the file states which number belongs to which line. It's a document where the layout does half the work of making sense of it, which means it can just as easily undo that work. You can be off by one row and never notice.

So we didn't trust the read. We pulled the same block for all 47 countries in the report, then looked for one country whose numbers are also published elsewhere, in plain text, to check whether the pairing held up.

The same five values show up in a second document: the country factsheet KPMG International publishes separately. Two different files, the same pairing.

The UK serves as the control. Our extraction gives it 38, 54, 39, 58, 47. On its own page, KPMG UK writes: "More than half of workers using the technology (54 per cent) say they made mistakes in their work due to AI and 58 per cent report that they have relied on AI output at work without evaluating its accuracy." Both numbers land exactly on the labels we expected.

The UK Office Read the Actual Report

That check has a side effect, and it's the important one. KPMG UK published its own non-verification rate, 58%, under the correct label, spelled out in its own press release.

Same study, same chart, same spring, two national offices of the same firm. One went and pulled the number from the report. The other misread the chart. So the research itself isn't the problem here: what broke is a communications link, and it's local.

The Trail Left in the Press

From there, you can actually trace the path a number took, the way a printing error gives away which edition a book came from. An article carrying the 45% read the press release. An article carrying the 60% opened the report.

Le Monde du Chiffre writes: "In France, only 39% of workers take a critical approach to AI, and 45% don't check the accuracy of the results they get." That's the press release's signature, word for word: this exact pair of numbers, in this exact order, in a single sentence, appears nowhere in the report.

IT Social goes a step further: "45% of users don't check the results from ChatGPT, Gemini and the rest." The base quietly widens along the way, from workers who use AI to all users, and the study's abstract "AI output" turns into three products the study never actually names.

None of the French coverage we read carries the 60%. There's nothing to hold against those newsrooms: reporting on a major firm's press release is a newsroom's normal job, and nobody opens a hundred-page PDF to fact-check a number they have no reason to doubt. The mistake happened upstream, and it traveled intact.

What Else Never Made the Cut

The 60% isn't the only thing from the report that stayed put. On feeding sensitive company information or protected content into a public AI tool, the study finds, globally, that the behavior is most common at organizations that have banned generative AI outright (67%) or set a usage policy (56%), versus 33% where there's no rule at all.

The authors' takeaway: outright bans probably don't work, and a policy doesn't guarantee compliance either. Neither Le Monde du Chiffre nor IT Social mentioned this finding. It wasn't in the press release.

The Caveats That Matter Too

Fieldwork ran between November 2024 and mid-January 2025. Eighteen months ago, and on AI use at work, that's a long time. The number describes a moment, not where things stand today.

KPMG has skin in the game here. The firm sells AI governance services, and a study showing workers left to their own devices doesn't hurt that business. The scientific leadership is academic and the methodology is public, which softens the concern without erasing it.

The item is self-reported, too. It measures what people say they do, not what they actually do.

The Real Number Is the Uncomfortable One

What the correction actually changes is the opposite of reassuring. It's not 45% but 60% of French AI-using workers who say they've relied on an output without checking its accuracy. Three in five, not a bit under half.

The number that circulated in France was the comfortable one.