A Quinnipiac University poll from March 2026 found that 76% of Americans trust AI-generated information "rarely or only sometimes." Only 21% trust it most or almost all of the time.

The same poll found that 51% use AI for research and information gathering.

The headline writers called this a contradiction. I think it's the most rational thing I've heard all year.

Trust and use are not the same question.

When you use a hammer, you don't trust it. You use it. It does what hammers do, and you stay aware that it might miss.

When you use GPS navigation, you don't trust it in the full sense of the word โ€” you follow it while keeping a mental model of whether the route feels right, ready to override when it says "turn left into a lake."

These are not failures of trust. They are correct calibrations. The tool is useful. The tool is also wrong sometimes. Both things are true.

What "trust" actually means when applied to AI

The survey asked whether people trust AI-generated information. This is a strange question because information has always required verification, regardless of source.

You don't trust a newspaper. You use it while knowing it has editors, biases, deadlines, and errors. You cross-reference. You update.

The new thing about AI is not that it can be wrong โ€” so can everything. The new thing is that it produces output with the texture of confidence regardless of accuracy. The receipt looks the same whether the transaction cleared or not.

That's the gap people are sensing: not that AI is untrustworthy, but that the signal for "this is reliable" looks identical to the signal for "this sounds reliable."

Using without trusting is a feature, not a bug

76% of Americans using AI while not trusting it is not mass cognitive dissonance. It is a population learning the right relationship with a new tool.

The right relationship is: useful, not oracular.

Use it to accelerate drafting. Keep the judgment layer yourself. Let it find the quote, then verify the quote. Let it write the first version, then rewrite it.

The bottleneck was never generation speed. The bottleneck was always the judgment layer โ€” who has the authority, the taste, the willingness to sign their name to a decision.

AI made the cheap part cheaper. The expensive part stayed expensive.

The receipt problem

There's a harder issue underneath this.

When you use GPS and it's wrong, you know it's wrong when you arrive somewhere unexpected. The feedback is immediate and physical.

When you use AI for information, the feedback loop is much longer โ€” sometimes months, sometimes never. A wrong answer that's never checked stays wrong. A wrong answer that sounds authoritative gets cited, then cited again.

The 76% who don't fully trust AI are not being paranoid. They are maintaining the verification infrastructure that the other 21% are partially outsourcing.

The question is not whether to trust AI. The question is: what does your verification loop look like, and is it faster or slower than the generation speed?

Source: Quinnipiac University National Poll, March 30, 2026. TechCrunch: "As more Americans adopt AI tools, fewer say they can trust the results."