On Trust

Researchers at the University of Arizona have studied a remarkable effect: people who disclose that they used generative AI in their work are, on average, trusted less.

Oliver Schilke and Martin Reimann examined this across 13 experiments in a range of contexts—involving managers and employees, professors, analysts, creatives, and even organizations. The effect proved remarkably consistent: disclosing the use of AI reduced trust in the person or organization making the disclosure. The authors call this a “trust penalty.” ¹

At first, this seems tragic. The very act of being transparent comes at a cost.

But I also find the effect understandable.

Whom do I believe?

We cannot independently verify every claim we encounter. So we have learned to look for other signals.

Who is saying this?
What experience does this person have?
How deeply have they engaged with the subject?
Are they competent?
Do they have something to lose if what they say turns out to be wrong?

These are not bad questions. On the contrary, they are necessary shortcuts in a world where no one can know or verify everything for themselves.

So the source of a statement is not irrelevant to us.

When a scientist reaches a conclusion after ten years of research, we respond differently than when someone who first encountered the subject yesterday makes exactly the same claim.

But this can easily blur two very different questions:

What is an insight worth?

and

What is the achievement or credibility of the person conveying it worth?

Suppose two people articulate exactly the same valid and relevant insight. One arrives at it after ten years of research. The other encounters it after twenty minutes of intensive dialogue with an AI system.

Is the second insight worth less?

The intellectual achievement is clearly different. The path by which they arrived there is different. There may also be good reasons to initially place greater confidence in the first person.

But that does not make the statement itself any less true.

AI disrupts a familiar shortcut

Generative AI is changing the relationship between effort, expertise, and results.

Today, someone with limited expertise can arrive at a remarkably good question or insight with the help of AI. An expert with decades of experience can still make a false claim. Both have always been possible. What AI changes is the scale and speed at which they can happen.

As a result, some of our familiar heuristics become less informative.

“Who said this?” still matters.

But the answer may tell us less about the quality of the statement than we are used to assuming.

More recent research by Schilke and Reimann helps explain why this unsettles us. Disclosing the use of AI affects perceptions of typicality, personal commitment, and authenticity—and, through them, trust. ²

The immediate reaction, “AI wrote that,” is therefore not simply irrational.

It becomes problematic at a different point:

when it ends our engagement with what was actually said.

“That came from AI”

I have encountered this problem in my own work.

AI systems are deeply involved in my work and in my thinking. I make no secret of that. On the contrary, I try to be as open about it as possible.

But that openness also makes it surprisingly easy to explain away whatever someone reads afterward:

That came from AI.

Very different processes can disappear into this single category.

Perhaps I asked an AI for its opinion and simply adopted its answer.

Perhaps I already held the view and only asked for help expressing it clearly.

Perhaps everything began with an observation of my own, which led to a question through dialogue.

Perhaps the AI challenged something I had previously believed.

Perhaps I explored several possible explanations and rejected them again.

Perhaps new sources or observations caused me to revise my original position.

Perhaps a belief emerged only after many conversations and repeated corrections.

From the outside, all of these can be described with the same sentence:

“He does it with ChatGPT.”

And I have a problem with that.

Not because I insist that the thoughts are somehow “really mine.”

But because the information “AI was involved” tells us far too little about how a belief was formed.

Perhaps we are disclosing the wrong thing

The obvious response to the trust penalty would be to make the use of AI less visible.

I want to move in the opposite direction.

More transparency.

But a different kind of transparency.

Not just:

AI was used here.

But:

This is how I came to believe this.

What did I initially observe?

What question emerged from that observation?

What possible explanations did I consider?

What supported them, and what spoke against them?

What sources did I consult?

Where did a conversation—with a person or with an AI—challenge my previous view?

Which assumptions did I discard?

What ultimately convinced me?

And what remains uncertain?

This also allows the role of AI to be described more precisely. It can be a source, a conversation partner, a critic, a research tool, a writing aid, or the catalyst for a question. Sometimes it may be several of these at once.

The point is not to minimize its involvement.

The point is not to confuse its involvement with the process of forming a belief.

From trust to traceability

Perhaps the growing role of AI brings us back to a much older question:

Why do I believe that a statement is true?

Quite often, the honest answer is: because I trust the person who said it.

There is nothing wrong with that. We could hardly function without trust.

But AI may force us to look more closely at the situations in which trust has simply been a necessary shortcut.

At the same time, AI gives us tools that can sometimes make that shortcut less necessary. We can ask for counterarguments. Find sources. Clarify concepts. Break down arguments. Explore alternative models. Look for contradictions.

Perhaps, then, we do not need to learn to place more trust in AI-generated statements.

Perhaps we can learn to need less trust in some situations.

Not by becoming generally more suspicious.

But by making more things open to scrutiny.

For my own work, this leads to something very concrete.

It is not enough to disclose that AI was involved.

If I want others to be able to seriously examine what I believe, I need to make the process by which those beliefs emerged more visible as well.

Not as proof that I am right.

Not as evidence that, despite the involvement of AI, a thought is somehow still “mine.”

But as an invitation:

This is what I believe today.
And this is how I got here.

What, if anything, you find convincing is still for you to decide.

Sources

Postscript

How this text came to be → Read the conversation