The short version
- In a preregistered experiment with 4,927 US adults, the same Google text was rated slightly less trustworthy, and less worth sharing, when labeled as AI-generated (Li and Aral (opens in a new tab)).
- The gap was small: 0.043 points on a 7-point trust scale, and it did not survive a stricter re-test, while the drop in willingness to share did (Li and Aral).
- In a 2026 field test with 1,100 US Google users, a week of forced AI Mode lowered trust in information on Google by 0.34 points on a 7-point scale (Wang and colleagues (opens in a new tab)).
- How AI answers are presented moves trust more than the AI label does: highlighting uncertain passages cut trust by 0.157 points, while adding reference links raised it by 0.091 (Li and Aral).
Do people trust the same answer less when it is labeled AI?
Yes, slightly, according to the largest experiment so far. Li and Aral (opens in a new tab) at MIT ran it from March 20 to April 2, 2024, with 4,927 participants matched to the US adult population on age, gender and ethnicity. Each person read Google answers to 9 policy and public-affairs questions, from gun control to adult vaccine schedules.
The trick was that everyone saw identical text. One group was told it came from Google’s generative AI; another saw it in the classic Google snippet format and was told it was not AI. People rated how accurate, believable, biased, complete and trustworthy each answer was, and whether they would share it with a friend.
Labeling the text as AI lowered both trust and willingness to share. The size matters, though. Traditional search was trusted just 0.043 points more on a 7-point scale, and that difference did not hold up when the authors re-tested the data with a stricter method. The sharing gap was larger, at 0.141 points, and it did hold.
Does using Google’s AI Mode change trust in Google?
Yes: a week of AI Mode lowered trust in information found on Google. Wang and colleagues (opens in a new tab), from the University of Pennsylvania and Northeastern University, ran a preregistered field experiment in March 2026 with a browser extension that changed what participants saw on Google.
After three normal days, 1,100 US Google users were split into three groups for seven days: AI Mode for every search, Google as it is today, or Google with AI Overviews hidden. Forced AI Mode lowered trust in information on Google by 0.34 points on a 7-point scale. It also lowered perceived usefulness, satisfaction and sense of control, and the share of people searching on a competing engine rose by 11.2 percentage points.
Hiding AI Overviews, by contrast, had no detectable effect on trust. Two caveats limit how far this travels. The sample skewed young and educated (75% under 45, 87% with at least some college), and participants had used AI Mode for only 0.6% of their searches before the test, so this measures a sudden, forced switch.
What do people say when asked about AI Mode?
They mention control, sources and accuracy more often than trust. In the same field test, the researchers coded 309 open-ended comments from people in the AI Mode group after a week of use.
| Theme in comments after a week of AI Mode | Share of 309 responses |
|---|---|
| Overall negative | 33.6% |
| Overall positive | 29.3% |
| Limited links or source diversity | 13.4% |
| Hallucination or accuracy concerns | 4.6% |
| Expressed trust in AI results | 1.6% |
Expert users voice similar doubts. In a 2024 user study by Narayanan Venkit and colleagues (opens in a new tab), 14 of 21 participants with PhD-level expertise said AI answer engines took extra work to verify and trust. And 12 of 20 distrusted the kinds of sources the engines used, such as forums, blogs and LinkedIn posts. One kind of source that deserves that caution appears in our study of self-ranking “best of” lists.
Who trusts AI search more, and who less?
Frequent AI users, college graduates and people working in technology trust it more. In Li and Aral’s experiment, people who used AI tools and AI search often trusted generative results significantly more, and were more willing to share them, than people who used them rarely or never.
Use was already common. In that 2024 sample, 85% of participants said they used generative AI, and 63% of those used it to look up information. Participants with a college degree, those working in technology-related industries, and Republicans also trusted the AI results more than their counterparts.
The pattern suggests the trust gap may narrow as use grows, though no study has yet tracked the same people over time to test that. That trust is also why ads blended into AI answers raise concern.
Does presentation change trust more than the AI label?
Yes: in the same experiment, design features moved trust more than the AI label itself. Adding reference links raised trust by 0.091 points on the 7-point scale, roughly twice the size of the AI-label gap. Highlighting passages the system was unsure about lowered trust by 0.157 points, and a generic explanation of how AI search works had no effect on average.
Trust also shapes behavior. Li and Aral found that people who trusted AI results more clicked on them more but spent less time evaluating them. How rarely readers open the cited pages is covered in our guide on checking AI sources. In the AI Mode field test, people spent 0.43 more minutes per search session, yet click-through to other websites fell by 18.8 percentage points.
What should you do about it?
Assume AI answers about your company are trusted almost as much as search results, and manage them accordingly.
- Do not discount AI answers as low-credibility. The measured trust penalty for the AI label was small, and frequent users showed little of it.
- Check what AI Mode, AI Overviews and the main assistants say about you, since the people most likely to read those answers are also the ones most likely to trust them.
- Give skeptical readers something to verify. Clear, specific pages that support the claims made about you serve the people who do click through.
- Track AI Mode separately from classic search. In the field test it changed both trust and clicking behavior far more than AI Overviews did.
For help monitoring and improving how AI search describes you, see our generative engine optimization service.
What does the research not tell us yet?
The two strongest studies cover public-affairs questions and Google’s own products, not brand or product searches.
- Li and Aral tested 9 policy-style questions and note that results may differ for product searches or everyday tasks.
- Their experiment ran in early 2024. Trust may have shifted as AI answers became routine, and no follow-up has measured that.
- The AI Mode field test lasted one week, with a younger, more educated and more left-leaning sample than the US population.
- Both studies are US-only. Trust in AI search may differ across countries and languages.
- We found no study comparing trust in standalone assistants such as ChatGPT with trust in Google’s results.
Frequently asked questions
Do people trust AI Overviews less than normal Google results?
A little, in one large experiment. The same text was trusted 0.043 points less on a 7-point scale when labeled AI, a gap that did not survive a stricter re-test.
Does Google’s AI Mode make people trust Google less?
In one field test, yes. A week of forced AI Mode lowered trust in information on Google by 0.34 points on a 7-point scale among 1,100 US users.
Did removing AI Overviews increase trust in Google?
No detectable change. In the same field test, hiding AI Overviews did not measurably change trust or the overall search experience.
Who is most likely to trust AI search results?
People who use AI often. In Li and Aral’s experiment, frequent users of AI tools, college graduates and tech-industry workers trusted generative search results more than others.
Sources
- Li and Aral (2025), Human Trust in AI Search: A Large-Scale Experiment (opens in a new tab), arXiv:2504.06435.
- Wang, Gleason, Bart, Wilson and Metaxa (2026), AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence (opens in a new tab), arXiv:2608.18352.
- Narayanan Venkit, Laban, Zhou, Mao and Wu (2024), Search Engines in an AI Era: The False Promise of Factual and Verifiable Source-Cited Responses (opens in a new tab), arXiv:2410.22349.