Some 42% of US workers surveyed who use AI on the job admit they have accepted an AI answer they suspected was wrong, while 30% say inaccurate AI output has already caused problems at work and just 22% are covered by a written employer policy requiring verification, reveals new research.
More than four in 10 US workers surveyed who use artificial intelligence (AI) at work have knowingly accepted an AI-generated answer they suspected was wrong because it was faster or sounded convincing, according to new research from Kolmogorov Law, raising fresh questions about human oversight, workplace accountability and responsible AI.
The national survey of 500 employed US adults who use AI for work, commissioned by California business litigation firm Kolmogorov Law, P.C. and conducted through Pollfish, found that 42% had gone ahead with an AI answer they suspected was wrong. Meanwhile, 65% do not always independently verify AI answers before acting on them or passing them along, while 30% say a wrong or inaccurate AI answer has already caused a problem at work.
Yet just 22% said their employer has a written policy requiring AI output to be verified before it goes into work product. The findings expose an important weakness in one of the most frequently cited safeguards for responsible AI: putting a human in the loop does not necessarily guarantee meaningful human oversight if that person fails to check the answer – or suspects it is wrong and uses it anyway.
VARIED AI VERIFICATION PRACTICES
The research found widespread use of AI among those surveyed. Some 72% use AI daily or several times a week, while nearly eight in 10 use it to look up information and around half use it to draft documents and emails. Almost one-quarter – 23% – said they have used AI for legal, financial or compliance questions. But verification practices vary considerably. Just 35% said they always independently verify an AI answer before acting on it or passing it along. Overall, 65% do not always verify AI output, with 32% verifying only sometimes, rarely or never.
More strikingly, 42% admitted accepting an AI answer they suspected was wrong because it was faster or sounded confident. The finding echoes wider concerns about employees becoming increasingly reliant on AI while simultaneously questioning whether they can trust its outputs.
Previous research found that nearly half of employees believe they rely too heavily on AI at work, while 39% fear that overdependence is weakening their skills and making them “less intelligent”. That study also found 80% believe workers are not being properly trained to use AI tools responsibly. Taken together, the findings raise a bigger question for employers: is requiring human oversight enough if employees do not have the skills, time, incentives or clear responsibility to challenge what AI produces?
AI ERRORS CAUSING PROBLEMS AT WORK
The Kolmogorov Law survey suggests the verification gap is already having consequences. Some 30% of respondents said a wrong or inaccurate AI answer had caused a problem in their work. Those problems ranged from errors in deliverables and poor decisions to lost time or money and client or legal issues. The risk appears greater when AI is used for higher-stakes work.
Among the 113 respondents who use AI for legal, financial or compliance questions, 43% said an incorrect answer had already caused a problem. That finding comes amid growing evidence that AI errors are moving beyond hypothetical risks. As recently reported, one in four executives say internal audits have detected AI errors that reached boards or external audiences. The global Workiva research revealed a striking disconnect: 84% of executives were at least somewhat confident in AI-generated output without human review, despite just 11% believing their organisation’s data quality was sufficient for AI use.
Some 26% said internal audits had detected AI errors that had already reached external audiences or board members, while 89% of institutional investors expressed concern about AI accuracy in corporate disclosures. The two studies examine very different populations and should not be directly compared. But together they highlight what can happen when confidence in AI – whether among employees or executives – runs ahead of organisations’ ability to verify what the technology produces.
AI TRUST PRESENTS A GROWING WORKPLACE PARADOX
The findings point to a wider contradiction emerging around workplace AI: people can question the technology while simultaneously becoming increasingly reliant on it. The issue is therefore not simply whether employees trust AI. The more important question may be whether they trust it appropriately. Too little trust can lead workers to reject useful AI recommendations, repeatedly recheck reliable outputs or avoid tools that could improve their work.
Too much trust creates a different risk: automation bias, reduced critical thinking and unquestioning acceptance of plausible but inaccurate information. The Kolmogorov findings expose a third possibility – people may use an AI answer even when they do not actually trust it. Some 42% of respondents admitted accepting an answer they suspected was wrong because doing so was faster or because the answer sounded confident.
That suggests responsible AI requires what might be described as calibrated trust: employees need to understand when AI can reasonably be relied upon, when an output requires independent verification and when human judgement must take precedence.
TRUSTWORTHY AI
That distinction is particularly significant in light of other recent AI research. As Fair Play Talks recently reported, organisations with strong trustworthy AI practices are 15 times more likely to report strong or high returns from their AI investments.
The global SAS report, featuring research insights from International Data Corporation (IDC), found that organisations with stronger AI governance, data quality, explainability and accountability were substantially more likely to report strong returns. It also exposed the importance of human trust: 97.2% of users override AI-generated recommendations in at least some circumstances, with the inability of AI to explain how it reached a decision emerging as the number-one reason for doing so.
The SAS findings and the latest Kolmogorov research approach trust from different directions. The first suggests that trustworthy and explainable AI is associated with stronger business outcomes.
The second demonstrates why the human side of that equation matters: workers need to know when an AI output deserves their confidence – and when it needs to be challenged. Responsible AI therefore does not require blind trust in technology. It requires informed and appropriately calibrated trust.
VERIFICATION AND HUMAN OVERSIGHT
One of the most interesting findings appears in the fuller analysis accompanying the survey. Among the 173 respondents who said they always verify AI output, 40% also admitted accepting an answer they suspected was wrong.
They also reported problems arising from incorrect AI answers at a similar rate to the overall sample – 32% compared with 30% overall. The apparent contradiction suggests that what people understand by “verification” may vary considerably – and that self-reported confidence in checking AI output does not necessarily translate into consistently challenging it.
That matters because businesses increasingly rely on human review as an important safeguard against AI hallucinations, inaccurate information and inappropriate recommendations. If human oversight becomes little more than routinely approving AI-generated output, having a person nominally “in the loop” may provide much less protection than organisations assume.
POOR AI OUTPUT AND HIDDEN PRODUCTIVITY COSTS
Inadequate verification can also create hidden productivity costs elsewhere in an organisation. Previous research showed that two-thirds of employees spend up to six hours each week correcting poor-quality AI-generated work or “workslop”. The wider problem is that AI can appear to save one employee time while creating additional work for someone else who has to identify, check and correct the output. That makes verification a productivity issue as well as a governance one.
If speed is rewarded while accuracy and verification are treated as secondary considerations, organisations could inadvertently encourage exactly the behaviour identified in the Kolmogorov research: accepting a questionable AI answer because doing so is quicker.
LACK OF FORMAL AI VERIFICATION
Despite those risks, the latest survey suggests many employers have yet to formalise basic verification expectations. Only 22% of respondents said their employer has a written policy requiring AI output to be verified before it goes into work product.
Another 59% said no such policy exists, while 19% did not know. The findings add to evidence that workplace AI adoption may be moving faster than organisational governance.
As previously reported, workers are adopting AI faster than many employers can govern it. Across 72 organisations analysed in that research, employees actively used almost 50 different AI applications, highlighting the challenge businesses face in maintaining consistent governance, training and oversight as different teams adopt different tools.
The growth of this so-called shadow AI creates an additional verification problem. Organisations cannot establish meaningful safeguards around AI-generated work if they do not understand which tools employees are using, for which tasks, what information is being shared with them and where AI-generated outputs are influencing decisions.
UNCLEAR AI RULES CAN DRIVE USE UNDERGROUND
Governance is not simply about publishing more policies. Employees also need to understand the rules and feel able to talk openly about how they are using AI. Recent research found that half of Gen Z workers feel guilty using AI, while more than four in 10 are hiding their use from employers. That creates another challenge for employers.
If organisations respond to AI mistakes purely through punishment, employees may become less willing to disclose their use of the technology or admit when something has gone wrong. But the opposite approach – allowing AI use without clear accountability – creates its own risks. As previously explored, AI transparency increasingly depends on psychological safety in the workplace.
Employees need to feel able to disclose AI use, question an AI-generated answer and report mistakes without fearing embarrassment or disproportionate punishment. Responsible AI therefore requires both psychological safety and accountability. People need permission to speak up – but they also need clear responsibility for the quality of consequential work.
WHO IS RESPONSIBLE WHEN AI GETS IT WRONG?
The survey also explored workers’ perceptions of legal responsibility. Asked who they believed would be legally responsible if they acted on an incorrect AI answer at work and it caused financial harm:
- 51% said they personally would be responsible;
- 18% named their employer;
- 12% said the company that made the AI; and
- the remainder selected other responses or were unsure.
Crucially, these figures represent respondents’ beliefs about legal liability. They do not establish who would actually be legally liable in any particular case. Responsibility can depend on the jurisdiction, circumstances, employment and contractual arrangements, applicable professional duties and the nature of the loss or harm.
Pavel Kolmogorov, founding attorney of Kolmogorov Law, believes the findings nevertheless demonstrate that many workers recognise that relying on AI does not necessarily remove their own responsibility for professional work. “Workers have figured out something most of their employers have not written down yet,” said Kolmogorov. “Half of them know the liability for a wrong AI answer is theirs. They are still not checking. When that answer ends up in a contract, a filing, or a client deliverable, ‘the AI said so’ is not a defense anyone has won with.”
His interpretation of the legal position should be distinguished from the survey finding itself, which establishes what respondents believe rather than determining their legal liability.
AI ACCURACY DISCLAIMERS
The survey also identified a knowledge gap around the terms governing AI tools. Some 52% did not know that the terms of service for major AI products disclaim responsibility for the accuracy of their answers, according to the research.
Kolmogorov argues that professionals and businesses therefore cannot assume responsibility transfers to an AI provider when they use its output. “The AI companies have disclaimed accuracy in their terms of service, so there is no one upstream to sue,” Kolmogorov said. “That leaves the professional and the business, which is exactly where the duty of care has always been. The tool changed. The rule did not.”
Again, the precise allocation of legal liability will depend on the relevant facts and law. But the broader workplace point remains important: using AI does not remove the need for organisations and professionals to establish who is accountable for the quality and accuracy of the work they produce.
MOST PROFESSIONALS TO DISCLOSE AI USE
On one issue, respondents showed considerably greater consensus. Some 78% said professionals such as lawyers, accountants, advisers and doctors should be required to disclose to clients when AI was used in their work.
Among respondents who were themselves client-serving professionals, support reached 84%. The finding suggests transparency could become increasingly important as clients seek to understand not simply what advice or work they receive, but how it was produced.
However, disclosure alone does not establish that AI has been used responsibly. The more important questions are whether the output was appropriately verified, whether confidential information was protected, whether human judgement remained meaningful and who was ultimately accountable for the work.
HUMAN OVERSIGHT TO CHALLENGE AI
The findings are particularly relevant as businesses increasingly describe “human-in-the-loop” systems as a safeguard against AI risk. Human oversight can play an important role. But simply inserting a person somewhere between an AI system and the final output does not guarantee that the output receives meaningful scrutiny.
The latest findings suggest employees may sometimes accept AI recommendations even when they have doubts about them. That complements the SAS findings showing that explainability influences whether users accept or override AI recommendations, as well as the Workiva findings showing the consequences when inaccurate AI output progresses far enough to reach boards and external audiences.
The emerging lesson is that organisations need to move beyond the simple instruction to “check AI”. They need to establish what checking actually means.
WHAT EMPLOYERS & BUSINESS LEADERS SHOULD DO NEXT
Define when independent verification is mandatory
Not every use of AI carries the same level of risk. Employers should identify higher-risk areas – including legal, financial, regulatory, safety, employment and client-facing work – where AI-generated facts, citations, calculations or recommendations require independent verification before use.
Define what “checking” actually means
Telling employees to check AI output is insufficient if everyone interprets the instruction differently. Verification might require returning to an original source, independently checking calculations, confirming that citations exist and actually support the claim made, or obtaining specialist review for consequential decisions.
Aim for calibrated trust
The goal should not simply be to make employees trust AI more. Workers need the skills to recognise when AI can reasonably be relied upon, when an output requires further scrutiny and when their own professional judgement should take precedence.
Make accountability explicit
Employees should understand who owns the final output when AI contributes to their work. The principle should be straightforward: AI can assist with work, but accountability for consequential decisions cannot simply disappear into the technology.
Don’t equate human involvement with human oversight
A person clicking “approve” does not necessarily constitute meaningful review. Organisations should examine whether workloads, deadlines, productivity targets or excessive confidence in AI encourage employees to approve outputs without properly challenging them.
Train people to recognise wrong AI
AI literacy should go beyond teaching employees how to write better prompts. Workers need to understand hallucinations, fabricated citations, weak sourcing, outdated information and the danger of assuming polished or confident language indicates accuracy.
Address shadow AI
Businesses should understand which AI tools employees actually use and provide appropriate approved alternatives. Rules that bear little relationship to everyday employee behaviour are unlikely to provide meaningful governance.
Create safe routes to report AI mistakes
Employees should be able to say “I think the AI got this wrong” or disclose that they relied too heavily on an AI output before the problem reaches a client, regulator or board. Psychological safety and accountability should reinforce one another rather than being treated as opposites.
Review the incentives behind AI use
The finding that workers sometimes accepted questionable AI output because it was faster deserves particular attention. Employers should examine whether productivity targets, workloads or deadlines unintentionally reward speed over accuracy and careful verification.
Document AI governance
With just 22% of respondents reporting a written verification requirement, organisations should ensure expectations are documented, communicated and understood rather than relying on informal assumptions about responsible AI use.
WHY THIS MATTERS FOR RESPONSIBLE BUSINESS
Businesses frequently respond to concerns about AI accuracy by pointing out that a human remains “in the loop”. The latest findings demonstrate why that reassurance may not be enough.
A human being present in the process is not the same as meaningful human oversight. If employees are under pressure to work quickly, trust confident-sounding AI answers, do not understand how to verify them – or use outputs even when they suspect they are wrong – the human can become little more than the final link through which an AI error reaches a client, customer, regulator or board.
Across recent Fair Play Talks reporting, a consistent picture is emerging. AI errors have already reached boards and external audiences. Workers are adopting AI faster than employers can govern it. Poor-quality AI output is creating hours of additional work. Employees are questioning whether reliance on AI is weakening their own capabilities. And some employees are concealing their use of the technology altogether.
TRUSTWORTHY AI DELIVERS STRONGER RETURNS
At the same time, the SAS research suggests organisations with stronger trustworthy AI practices are considerably more likely to report strong returns from the technology. Together, these findings suggest the challenge is not simply whether businesses and workers trust AI. It is whether they can trust it appropriately. Too little trust can undermine adoption and productivity. Too much can weaken scrutiny and encourage automation bias. And employees using AI even when they suspect it is wrong exposes a third risk: convenience overriding judgement altogether.
Responsible AI therefore requires calibrated trust – supported by good data, explainability, AI literacy, clear verification requirements, meaningful human judgement, transparent accountability and workplace cultures where people feel able to challenge what AI tells them.
The question for employers is no longer simply whether they have put a human in the loop. It is whether that human has the skills, time, authority and responsibility to challenge what the machine produces.
RELATED ARTICLES
AI may be transforming workplace productivity, but a growing number of employees fear the technology is making them less intelligent in the process, according to survey.
Despite 84% of executives expressing confidence in AI-generated output without human review, just 11% believe their organisation’s data quality is sufficient for AI use, exposing a potentially dangerous gap between confidence, governance and reality.
Two-thirds of employees say they spend up to six hours a week fixing low-quality, AI-generated work-slop, according to a new study.
AI transparency policies will only succeed when employees feel psychologically safe to disclose how they use AI at work.
Half of Gen Z workers feel guilty using AI and more than four in 10 are using it without their employer’s knowledge, despite businesses increasingly demanding AI skills from new recruits.
Employees are embracing AI faster than many organisations can govern it, according to new research.
