--- title: "How often do people check AI output before using it? (41–44%) | The AI at Work Fact Book" description: "About 41 to 44 percent of AI users use the output without routine checking, and 56 percent have made a mistake at work because of AI. Self-reports and observed behavior." url: "https://evy.so/facts/checking-ai-output/" last_updated: "2026-10-16" --- F-22 · What Teams Trust # How often do people check AI output before using it? About four in ten AI users regularly use AI output without checking it, by their own account: 44 percent of AI-using employees across 47 countries do not verify its accuracy most of the time, and knowledge workers report no critical thinking on 41 percent of their AI-assisted tasks. 56 percent of AI-using employees say AI has caused a mistake in their work at least once. 41–44% of AI users, or of their AI tasks, go without routine checking (self-reported) - Medium confidence - 4 primary sources - 2023 to 2026 - Last checked October 14, 2026 ## Cite this EVY (2026). How often do people check AI output before using it? The AI at Work Fact Book, F-22. https://evy.so/facts/checking-ai-output/ (last checked October 14, 2026). CC BY 4.0. Copy citation [Data as JSON](https://evy.so/facts/checking-ai-output.json) [Markdown](https://evy.so/facts/checking-ai-output.md) ## How to read it Most of this is self-reported, and people likely flatter themselves. KPMG counts people, Lee and colleagues count tasks. Checking depends on how easy it is: developers can run code, so only about 10% use AI answers as-is, while KPMG finds that people who trust AI more check less. The only observed studies are a 2023 writing experiment and a small 2025 trial of 16 developers, who accepted under 44% of AI code suggestions. ## Every study on one scale Shaded: 41–44% One figure A range in one study Chart values as text - Scale: % used without routine checking. Shaded range: 41–44%. - KPMG and Melbourne (2025): 42–44% of AI users - Lee et al., Microsoft and CMU (2025): 40.7% of tasks - Stack Overflow 2026 (2026): 9.6% use as-is - Noy and Zhang, experiment (2023): 68% unedited (experiment) **Why medium confidence:** Two independent creators with very different methods land on almost the same figure, one counting people and one counting tasks. Both are self-reports, which likely understate careless use, and the observed studies are small or older. ## Sources Sources for F-22 | Study | Creator | Year | Sample | Method | Figure | Link | | --- | --- | --- | --- | --- | --- | --- | | Trust, attitudes and use of artificial intelligence: A global study 2025 | University of Melbourne and KPMG International | 2025 | 32,352 employed people in 47 countries, November 2024 to January 2025 | Online survey of employees who use AI at work, self-reported frequency | 56% verify AI accuracy most of the time or always, so 44% do not; 42% rely on output without evaluating it sometimes or more often (66% ever) “Two in three report relying on AI output without evaluating the information it provides” | [mbs.edu ↗](https://mbs.edu/-/media/PDF/Research/Trust_in_AI_Report.pdf) | | The Impact of Generative AI on Critical Thinking (CHI 2025) | Hao-Ping (Hank) Lee and colleagues, Microsoft Research and Carnegie Mellon University | 2025 | 319 knowledge workers using AI at least weekly, describing 936 real tasks | Online survey; for each task, whether the worker did any critical thinking such as verifying or editing | Critical thinking reported on 59.3% of tasks, so none on 40.7%; higher confidence in AI predicted less critical thinking “Our participants self-reported to have enacted critical thinking for 555 (59.29%) of the examples” | [microsoft.com ↗](https://www.microsoft.com/en-us/research/wp-content/uploads/2025/01/lee_2025_ai_critical_thinking_survey.pdf) | | 2026 Developer Survey: AI data | Stack Overflow | 2026 | 13,162 developers answering this question, June to August 2026 | Opt-in survey; what developers usually do with an AI answer, several answers allowed | 9.6% use AI answers as-is; 76.5% run the code, 63.9% compare it with the codebase “Only 10% use answers as-is.” | [survey.stackoverflow.co ↗](https://survey.stackoverflow.co/2026/ai/data) | | Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence | Shakked Noy and Whitney Zhang (MIT) | 2023 | 444 college-educated professionals in an online writing experiment, early 2023 | Self-report plus observed activity on the platform; not participants' real jobs | 68% said they submitted ChatGPT's first output unedited, and were active for about 3 minutes after pasting it in “68% of treated participants report submitting ChatGPT’s initial output without editing it” | [economics.mit.edu ↗](https://economics.mit.edu/sites/default/files/inline-files/Noy_Zhang_1.pdf) | Mistakes Low confidence ## How many workers have made mistakes because of AI? Between about a third and a half of AI users. 56% of AI-using employees across 47 countries say they have made a mistake in their work because of AI, and 31% say it happens sometimes or more often. In a small 2026 US poll, 30% said a wrong AI answer had caused a problem in their work. Noticing errors is more common still: 66% of developers report AI answers that are almost right, but not quite. How many workers have made mistakes because of AI? | Study | Creator | Year | Sample | Method | Figure | Link | | --- | --- | --- | --- | --- | --- | --- | | Half of U.S. Workers Know They Would Be Personally Liable for a Wrong AI Answer | Kolmogorov Law, poll on Pollfish | 2026 | 500 employed US adults who use AI for work, September 2026, unweighted | Online poll commissioned by a law firm; treat as indicative | 30% say a wrong AI answer has caused a problem in their work; 65% do not always verify AI answers “a wrong or inaccurate AI answer has caused a problem in their work” | [prnewswire.com ↗](https://www.prnewswire.com/news-releases/half-of-us-workers-know-they-would-be-personally-liable-for-a-wrong-ai-answer-two-thirds-still-dont-always-check-kolmogorov-law-survey-finds-302867898.html) | | 2025 Developer Survey: AI | Stack Overflow | 2025 | 25,332 developers answering the frustrations question | Opt-in online survey, several answers allowed | 66% cite AI answers that are almost right, but not quite; 45.2% say debugging AI code takes more time “AI solutions that are almost right, but not quite” | [survey.stackoverflow.co ↗](https://survey.stackoverflow.co/2025/ai) | Reaching customers ## Do AI mistakes reach customers? Sometimes. A quarter of IT leaders (25%) say AI mistakes have already hurt customers or clients, 23% that they cost the company money, and 16% that they seriously disrupted operations. Of employees who review colleagues' AI work, 79% regularly receive output that is low quality or contains errors. Do AI mistakes reach customers? | Study | Creator | Year | Sample | Method | Figure | Link | | --- | --- | --- | --- | --- | --- | --- | | The Pulse of Work in 2026 | GoTo with Workplace Intelligence | 2026 | 1,250 knowledge workers and 1,250 IT decision-makers in 10 countries, November 2025 to January 2026 | Online survey. GoTo sells IT and communications software. | 25% of IT leaders say AI mistakes hurt customers; 79% of employees who review colleagues' AI work regularly get flawed output “around one-quarter of IT leaders say AI has made mistakes that affected customers” | [edge.sitecorecloud.io ↗](https://edge.sitecorecloud.io/gototechnol00e8-mktglobalxm07e7-mktglobalpr9498-b870/media/goto/pdfs/ebook/ebook_pulse-of-work-2026.pdf) | ## Numbers you may have seen These figures circulate widely. Here is what the original source says. - **Two thirds of employees use AI output without checking it (KPMG).** 66% have done so at least once, including rarely. Doing it sometimes or more often is 42%, and 56% verify accuracy most of the time. - **57% of workers have made mistakes because of AI (KPMG).** The global report says 56%, counting anyone who did so even once. Sometimes or more often is 31%. - **AI is destroying critical thinking (Microsoft and Carnegie Mellon).** The study measured 319 workers' own sense of effort and whether they thought critically. It did not test thinking skills or show a decline over time. - **77% of employees review AI work as carefully as human work (Workday).** The 77% covers only employees who use AI every day, at companies with more than $100 million in revenue, in a vendor survey. - **66% of developers spend more time fixing AI code (Stack Overflow).** 66% cite answers that are almost right; 45.2% say debugging AI code takes more time. ## More from What Teams Trust - [F-21**24–53%**How much do workers trust the output of AI tools at work?](https://evy.so/facts/trust-in-ai-output/) - [F-23**17–45%**How often is AI output wrong on real work tasks?](https://evy.so/facts/ai-error-rates-at-work/) - [F-24**2–12%**Which work tasks do people want AI to do on its own, and which do they want to keep?](https://evy.so/facts/tasks-workers-want-ai-to-do/) - [F-25**30–44%**How many workers use AI in ways that break company policy?](https://evy.so/facts/ai-use-against-company-policy/) - [F-26**42–48%**How often do employees put company data into public AI tools?](https://evy.so/facts/company-data-in-public-ai/) - [F-27**21–29%**How many organizations have had a security incident or data breach involving AI?](https://evy.so/facts/ai-security-breaches/) [The whole chapter](https://evy.so/facts/what-teams-trust/) · [Chapter poster](https://evy.so/facts/what-teams-trust/sheet/) · [All of The AI at Work Fact Book](https://evy.so/facts/)