Nearly half of American workers using AI on the job admit they have accepted an answer they suspected was wrong, according to new research that exposes a troubling gap between responsible AI policies and actual workplace behavior. The findings raise urgent questions about whether simply putting a human in the loop actually prevents AI errors from causing real damage.

What's Actually Happening When Workers Use AI at Work?

A national survey of 500 employed US adults who use AI for work, commissioned by California business litigation firm Kolmogorov Law and conducted through Pollfish, paints a picture of widespread AI adoption paired with inconsistent verification practices. The research found that 72% of respondents use AI daily or several times a week, with nearly eight in 10 using it to look up information and around half using it to draft documents and emails.

But here's where the problem emerges: just 35% of workers said they always independently verify an AI answer before acting on it or passing it along. Meanwhile, 65% do not always verify AI output, with 32% verifying only sometimes, rarely, or never. Even more striking, 42% admitted accepting an AI answer they suspected was wrong because it was faster or sounded confident.

The consequences are already materializing. Some 30% of respondents said a wrong or inaccurate AI answer had caused a problem at work, ranging from errors in deliverables and poor decisions to lost time, money, and client or legal issues. The risk appears significantly higher 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.

Why Doesn't Human Oversight Actually Work?

The research reveals a troubling contradiction at the heart of responsible AI governance. Just 22% of workers said their employer has a written policy requiring AI output to be verified before it goes into work product. Yet even among the 173 respondents who claimed they always verify AI output, 40% also admitted accepting an answer they suspected was wrong. They also reported problems from incorrect AI answers at a similar rate to the overall sample, 32% compared with 30% overall.

This apparent contradiction suggests that what people understand by "verification" may vary considerably. Self-reported confidence in checking AI output does not necessarily translate into consistently challenging it. The problem is that businesses increasingly rely on human review as an important safeguard against AI hallucinations and inaccurate information. 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 organizations assume.

How Can Organizations Build Better AI Verification Practices?

• Establish Clear Written Policies: Only 22% of workers have written employer policies requiring verification before AI output enters work product. Organizations need explicit, documented standards that specify when and how AI answers must be independently checked before use or distribution.
• Provide Targeted AI Literacy Training: Previous research found that 80% of workers believe they are not being properly trained to use AI tools responsibly. Training should focus on when AI can reasonably be relied upon, when outputs require independent verification, and when human judgment must take precedence.
• Create Accountability Mechanisms: Workers need clear responsibility for challenging AI output and consequences for accepting answers they suspect are wrong. This includes tracking which employees are approving AI-generated work and monitoring error rates by department or role.
• Implement Verification Checkpoints for High-Stakes Work: Given that 43% of workers using AI for legal, financial, or compliance questions reported problems, organizations should require mandatory secondary review for AI output in these domains before it reaches clients, boards, or external audiences.
• Develop "Calibrated Trust" Frameworks: Rather than asking workers to either trust or distrust AI completely, organizations should help employees understand the specific contexts where AI is reliable versus where it requires scrutiny, based on the type of task and the model's known limitations.

The research points to 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 judgment must take precedence. Too little trust can lead workers to reject useful AI recommendations or repeatedly recheck reliable outputs. Too much trust creates automation bias, reduced critical thinking, and unquestioning acceptance of plausible but inaccurate information. But the Kolmogorov findings expose a third possibility: people may use an AI answer even when they do not actually trust it.

What Does Trustworthy AI Actually Look Like?

Recent research suggests that organizations with strong trustworthy AI practices see substantially better returns on their AI investments. A global report from SAS, featuring research insights from International Data Corporation (IDC), found that organizations with stronger AI governance, data quality, explainability, and accountability were substantially more likely to report strong returns. The research 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.

This finding aligns with broader concerns about AI in professional settings. Research from Bangladesh's legal sector, which examined how AI should be governed in courts and law firms, found that stakeholders recognize AI's potential to reduce inefficiency and improve legal services but strongly emphasize human oversight and accountability. Judges stressed that AI should be advisory and not replace judicial discretion, while lawyers highlighted issues of accuracy, trust, bias, confidentiality, and professional responsibility.

The Bangladesh study argues that responsible AI governance should be grounded in principles of equality, due process, fair trial, and judicial independence. Rather than transplanting foreign regulatory models, countries should adopt a gradual, risk-based, and context-specific framework that integrates human oversight, transparency, professional ethics, data protection, institutional accountability, and AI literacy.

The broader pattern emerging from these findings is clear: responsible AI does not require blind trust in technology. It requires informed and appropriately calibrated trust, backed by clear policies, adequate training, and accountability mechanisms that ensure human oversight is meaningful rather than merely nominal. Without these safeguards, the human in the loop may provide much less protection than organizations assume.