A national survey of 500 U.S. workers who use AI on the job: 42% have knowingly gone with an answer they suspected was wrong, 30% have already been burned by one, and only 22% have an employer rule that says check first.

Key findings

• 65% of workers who use AI on the job do not always verify an AI answer before acting on it; one in three checks only sometimes, rarely, or never.
• 42% have knowingly accepted an AI answer they suspected was wrong because it was faster or sounded confident.
• 30% say a wrong AI answer has already caused a problem at work; among those using AI for legal, financial, or compliance questions, 43%.
• 51% believe they would personally be legally responsible for acting on a wrong AI answer. Only 12% think the AI company would be.
• 52% did not know that AI companies' terms of service disclaim accuracy and generally bar lawsuits.
• Only 22% work under a written employer policy requiring AI output to be verified.
• 78% say lawyers, accountants, advisors, and doctors should be required to disclose when AI was used in their work.

Earlier this year The New York Times hired Oumi, an AI testing company, to answer a basic question: how often are Google's AI Overviews right? The panels sit on top of the search results that billions of people see every day, so it seemed worth knowing.

The answer Google prefers is 91 percent. That is what Oumi measured for the current Gemini 3 model across 4,326 searches, up from 85 percent with the previous model a year earlier. It sounds reassuring until you do the arithmetic. Google handles something like five trillion searches a year. Nine percent of that works out to tens of millions of wrong answers every hour.

What struck me was not the error rate. It was a second finding buried in the same study, and a third from a different study altogether. Both are about people, not software.

Nobody is checking

In April, researchers at the University of Pennsylvania's Wharton School published the results of three experiments involving 1,372 participants. They wanted to know what people actually do when an AI hands them an answer. Mostly, it turns out, they take it. When the AI was confidently wrong, 73 percent of participants went along with it anyway. Having the AI in the loop made people more confident in the wrong answer, not less.

The researchers called this "cognitive surrender." It is not the same thing as handing a task to a machine on purpose and checking the output afterward. That is delegation, and there is nothing wrong with it. Surrender is when the checking stops altogether because the answer came back fast and read smoothly.

The Penn study was run under lab conditions. I wanted to know how much of this has migrated into actual workplaces, where a wrong answer has consequences. So on September 1 we commissioned a national survey of 500 employed American adults who use AI tools for work, fielded through the Pollfish research platform.

What 500 working AI users told us

These are not occasional users. Seventy-two percent reach for AI daily or several times a week. Nearly eight in ten use it to look things up, and about half use it to draft documents and email. Twenty-three percent said they have used it for legal, financial, or compliance questions. That last number is the one I keep coming back to.

We asked how often they independently verify an AI answer before acting on it or passing it along. Thirty-five percent said always. The other 65 percent do not: a third said usually, a quarter said sometimes, and 6 percent said rarely or never. So roughly one working AI user in three checks only sometimes, at best.

Often they know better at the time. Forty-two percent admitted they had gone ahead with an answer they suspected was wrong, either because it was quicker or because it sounded sure of itself. A quarter of respondents said they had done it more than once.

And it is already costing them. Thirty percent said a wrong AI answer has caused a problem in their work: a mistake in a deliverable, a bad decision, lost time or money, a client or legal issue. Four percent described the problem as serious. Another 24 percent could only say "not that I know of," which is not the same as no.

The people using AI for the riskiest questions have the worst record. Of the 113 respondents who use it for legal, financial, or compliance matters, 43 percent said a wrong answer had already cost them something. Sixty-four percent of that same group still do not always verify.

One more result deserves a mention, because it undercuts the comfortable version of this story. Among the 173 people who told us they always verify, 40 percent also admitted accepting an answer they suspected was wrong, and they reported work problems at the same rate as everyone else, 32 percent against 30 percent overall. People believe they check. The evidence says they check less than they think they do. That is more or less what the Wharton team found in the lab.

The citations don't hold up either

There is a second problem hiding inside Google's 91 percent. Oumi also found that 56 percent of AI Overviews are "ungrounded," meaning the source the AI cites does not actually support what the AI said. A year earlier that figure was 37 percent. Accuracy went up; the citations got worse.

Think about what that does to the diligent user. She clicks the footnote, sees a real article from a real publication, and reasonably concludes the answer is backed up. More than half the time, it isn't. Verification is getting harder at the same moment fewer people are bothering with it.

Workers already know who pays

This is the part that matters to a business litigator.

We asked: if you acted on a wrong AI answer at work and it caused financial harm, who do you think would be legally responsible? I expected a lot of people to point at the AI company. They didn't. Fifty-one percent said "me personally." Eighteen percent said their employer. Only 12 percent named the company that built the AI, and 15 percent were unsure.

The majority has it right. OpenAI, Google, and Anthropic all disclaim responsibility for the accuracy of their output in their terms of service. You cannot bring a malpractice claim against a chatbot, and there is no errors-and-omissions policy sitting behind a free browser tab. Whatever duty attached to the work before AI came along (competence, accurate statements to clients and investors, contractual warranties) still attaches to the person who did the work. The tool has nothing to do with who answers for the result. "I got it from AI" is an explanation, not a defense.

So workers, by and large, understand that the exposure is theirs. It has not changed how they behave. And a good many are still missing the specifics: 52 percent did not know that the AI vendors' terms of service say this in writing and, in practice, foreclose a lawsuit.

Employers have not filled the gap. Only 22 percent of respondents said their employer has a written policy requiring AI output to be verified before it goes into work product. Fifty-nine percent said there is no such policy. Nineteen percent didn't know, which for practical purposes is the same thing. Among client-serving professionals in law, finance, healthcare, consulting, and real estate, the figure was about one in three.

The courts are starting to weigh in

On May 28, 2026, the Regional Court of Munich issued a preliminary injunction holding Google directly liable for false statements its AI Overviews made about two German businesses. Google argued that it tells users to double-check AI results. The court was not persuaded. An AI summary, it reasoned, is Google's own statement, composed in Google's own voice, not a link to something someone else wrote. Search engines have relied on the opposite characterization for about thirty years. Google is appealing.

A German ruling does not bind an American court. But the reasoning travels, and it lines up with a view many U.S. scholars already hold: Section 230, which protects platforms from liability for what third parties post, probably does not reach content the AI itself generates. If a platform can be answerable for what its AI says, a business can certainly be answerable for what it does with the answer. (We keep a running FAQ on AI liability questions for business owners.)

The one thing almost everyone agreed on

Seventy-eight percent of respondents said professionals (lawyers, accountants, advisors, doctors) should be required to tell clients when AI was used in their work. Twelve percent disagreed. Among respondents who are themselves client-serving professionals, support ran to 84 percent. The people using these tools want the people they hire to be candid about using them.

What to do with this

I use AI tools in my own practice and I am not going to tell anyone to stop. As a first draft, a way to organize a problem, or a place to start research, they earn their keep. The trouble starts when the output gets treated as finished work.

For the individual professional the rule was settled long before the technology arrived: a source is a lead, not a conclusion, and you verify before you rely. For an employer the picture is more concrete. Two-thirds of your AI-using staff don't consistently check. Close to half have knowingly run with a doubtful answer. Fewer than a quarter are operating under any written rule about it. That is not a technology problem. It is a claim that hasn't found its facts yet.

Google's 91 percent tells you something about the machine. The numbers above tell you something about us, and they are less flattering. Most of us know the bill lands on our own desk. Most of us still don't check.

Methodology

This article is based on a survey of 500 employed U.S. adults, ages 18 to 64, conducted September 1, 2026, through the Pollfish online research platform. Respondents were screened to full-time, part-time, and self-employed workers. All 500 reported using AI tools for work at least occasionally; 72 percent use them daily or several times a week. The questionnaire consisted of one occupation question, one usage-frequency question, eight substantive questions, and one attention check, which all 500 respondents passed. Results are unweighted. The margin of error is approximately plus or minus 4.4 percentage points at the 95 percent confidence level for the full sample. Subgroup figures (client-serving professionals, n=79; users of AI for legal, financial, or compliance questions, n=113; respondents who say they always verify, n=173) carry wider margins.

Sources

Original survey data: Kolmogorov Law / Pollfish, September 1, 2026, n=500 employed U.S. adults who use AI for work. AI Overviews accuracy and grounding: New York Times-commissioned Oumi study (SimpleQA benchmark, 4,326 searches; 85% Gemini 2 / 91% Gemini 3 accuracy; ungrounded responses 37% to 56%), as reported by The New York Times, Search Engine Land, TechRepublic, and Oumi (April 2026). "Cognitive surrender" and verification behavior: University of Pennsylvania / Wharton study (3 experiments, 1,372 participants; 73% accepted confident-but-wrong answers), April 2026. Platform liability: Regional Court of Munich preliminary injunction holding Google liable for false AI Overviews (May 28, 2026; on appeal). AI terms-of-service accuracy disclaimers: OpenAI, Google, and Anthropic. Tip and framing courtesy of Kay Rubacek, Human Edge Research.

Related reading from Kolmogorov Law

• Nearly 2 in 5 U.S. workers have put company information into personal AI accounts, and most don't know it can be illegal
• AI notetakers have sat in on 1 in 3 U.S. workers' meetings, but only a third say they were asked first
• The AI Layoff Defense: 87% of Americans want a human to sign off before a machine cuts their job
• The ChatGPT subpoena revolution: when your AI conversations become court evidence
• An AI read your brief before the judge did, and most Californians have no idea

Kolmogorov Law, P.C. is a California business litigation firm in Irvine. If a wrong AI answer has become a contract, fraud, or professional-liability dispute for your business, contact us or call (909) 235-6420.