Can You Trust AI

AI has a trust problem.

Even the people building the world's most powerful AI systems are telling the public to be cautious. Anthropic CEO Dario Amodei recently called for slowing the development of frontier AI. OpenAI CEO Sam Altman agreed that the industry needs to "pace the frontier," and SpaceXAI CEO Elon Musk backed Amodei's proposal. 

The public was already skeptical. A 2026 Pew Research Center survey found that 40% of Americans expect AI to have a negative impact on society over the next 20 years, compared with only 16% who expect a positive impact. About six in ten lack confidence that U.S. companies will develop and use AI responsibly. 

So what would make people trust AI?

My research reveals a surprising answer: other people using it.

The AI Trust Paradox

Years before the arrival of ChatGPT, my lab studied why people would trust algorithms. Participants completed timed mazes to earn money and could pay for assistance from an imperfect algorithm. They knew the algorithm sometimes made mistakes and were free to follow or ignore its recommendations.

The choice was important because my research showed that trust is a behavior, something that is much stronger and more actionable an answer to a survey question. Trust occurred when someone actually relied on the algorithm's advice. 

We then gave subsets of participants different information about the algorithm.

Some received no information. Some were told the algorithm was 75% accurate. Others were told that a percentage of other people had used it.

Telling participants that the algorithm was 75% accurate did not significantly increase its use compared to the no-information condition. But, telling them that other people used the algorithm did. The statistical model showed that social information increased the odds of adoption 4.5-fold. 

That creates a conundrum for AI.

People trust algorithms when they see other people using them. But widespread use requires people to trust them in the first place.

Don't Ask People What They Like

There was another important result in our study.

People who used the algorithm without information put less cognitive effort into monitoring it.  In other words, they paid less attention to what it was doing and just relied on its recommendations. Since they did not monitor the algorithm, their performance suffered and they earned less money. By contrast, people given social information about the algorithm had a heightened neurophysiologic response showing they were monitoring the algorithm.  Those who monitored it performed better. 

Trust, then, is not the same thing as liking something. Nor is trust necessarily revealed by asking someone whether an algorithm is useful, safe, or trustworthy.

Behavior is the test.

This distinction matters enormously for AI.

AI companies routinely solicit explicit feedback. Thumbs up. Thumbs down. Ratings. Surveys. Did this answer help? Did the user "like" this response?

But liking is not the same as being valuable.

People say they like healthy food and eat potato chips. They say they want to exercise and stay on the couch. They say an advertisement did not influence them and then buy the product.

Behavior reveals what people value; self-reports often do not.

Watch What People Do

The path to trustworthy AI is therefore not simply convincing people that AI is accurate. My lab's published research shows that accuracy information alone is insufficient to induce trust.

AI needs to demonstrate that it understands what matters to people.

That requires shifting attention from what people say they like to what they actually value. The distinction is fundamental. A person can like an AI response without acting on it. Another response will influence behavior without receiving a thumbs-up.

The first produces an opinion. The second reveals value.

The same principle applies outside AI. Surveys and ratings capture conscious reports. Behavior reveals value.

This is why ImmersionLive was built around the objective measurement of value. This easy-to-use kit has wearables and software that continuously measure a neurophysiologic signal called Immersion, identifying second by second what the brain values during an experience.

No survey is required. No rating is required. No one has to explain why something mattered.  The data are granular.

The brain continuously reveals the value of an experience.

The best way to know what is important to people is to measure what they value.

See how ImmersionLive measures value by scheduling a demo today.

FAQs

What makes people trust AI?

Research suggests that accuracy information alone may not be enough. In Dr. Paul J. Zak’s study, participants were significantly more likely to use an algorithm when they learned that other people had used it.

Why aren’t surveys enough to measure trust?

What people say they like or trust does not always predict what they will actually do. Behavioral and neurophysiologic measures can provide additional insight into whether an experience is valuable enough to influence action.

How can neuroscience help measure what people value?

Immersion continuously measures a neurophysiologic signal associated with value, allowing researchers and organizations to see how people respond to an experience moment by moment rather than relying only on feedback collected afterward.