Open Questions for Human-AI Communication

Dan Kahan's research found that the most analytically skilled people are often the most biased on emotionally charged topics, not the least. Their skill gets used to build a more persuasive case for what they already believe, not to find the truth.

Two salespeople pitch the same product. One is mediocre at their job, the other is brilliant. If both are equally willing to stretch the truth, the brilliant one is more dangerous, not less, because they are simply better at making the stretch sound convincing.

If capability helps: a more advanced AI system might get better at catching its own errors, cross-checking itself, and flagging uncertainty, the way an experienced professional double-checks their own work out of habit.

If capability doesn't help: a more advanced AI system might just get better at sounding right, making its mistakes harder to catch, not easier, mirroring exactly what Kahan found in skilled human reasoners.

Genuinely unresolved. The human version of this question already has a documented answer. Whether the same pattern applies to AI systems as they improve is still untested.

A person's reasoning gets bent by wanting to protect their identity or beliefs. An AI system doesn't have that kind of personal stake in any particular answer. That doesn't automatically mean it has no bias at all.

A translator who has never visited either of two feuding neighbourhoods might seem perfectly neutral. But if all the training materials they learned from happened to describe one neighbourhood more favourably, they could still translate in a lopsided way, without ever feeling biased, because the bias came from what they were taught, not from personal loyalty.

If it's a real advantage: an AI system without personal stakes may genuinely be freer to follow the evidence wherever it leads, compared to a person defending a long-held position.

If it just relocates the problem: whatever shaped the AI's training material, and whatever a search process chooses to surface or leave out, could introduce a different kind of skew, one that is harder to notice precisely because there is no visible motive driving it.

Genuinely unresolved, and probably not a single yes or no answer. AI likely avoids one specific type of bias while remaining vulnerable to other types that need to be checked separately.

Jonathan Haidt's research suggests people often rely on a felt sense that something has been thought through enough, not just a formal rule. An AI system doesn't have that felt sense. The natural fix people suggest is to give it an explicit checklist instead.

A home cook often just knows a sauce needs more salt, without measuring anything. A recipe that instead says "taste after every half teaspoon added" is more explicit and repeatable, but it might also miss subtle cues an experienced cook would have caught by feel, or create false confidence that following the steps guarantees a good result.

If a checklist works: it gives a consistent, visible, checkable process that doesn't rely on an invisible internal feeling nobody else can verify.

If it doesn't fully work: a checklist can create a different problem, false confidence. Something marked as checked can feel finished even if the check itself wasn't actually thorough enough for the situation.

Partially answered. Explicit checklists are clearly useful and more transparent than relying on an unverifiable feeling, but whether they fully substitute for the flexible judgement a felt sense provides is still open.

Ditto and Lopez found that people apply a lower bar of scrutiny to information that already confirms what they believe. The open question is whether an AI source makes that effect stronger, because AI can carry a sense of neutral authority a human source doesn't automatically have.

If a friend tells you a rumour, you might reasonably wonder about their motives. If a search engine or reference book tells you the same thing, it can feel more official, even if the underlying reliability is actually similar. The open question is whether AI answers get treated the same way, as if they are inherently more neutral just because they didn't come from a person with an obvious agenda.

If it's true: AI-sourced claims may spread and get accepted more easily precisely because they don't look like they're coming from someone with an agenda, even when the AI got something wrong.

If it's not a major effect: people may already discount AI answers appropriately, treating them the same as any other source that needs checking.

Genuinely unresolved. A reasonable, evidence-informed concern extending known research about humans, but not yet specifically tested for AI as the information source.

Lee and See's trust research found that people build trust in unfamiliar systems through observed behaviour over time, not through the system's own claims about itself. The open question is what this means practically for someone who can't watch a system's behaviour over months or years, like someone having a single long conversation with an AI.

You can't fully verify a new contractor is trustworthy in one conversation. You would normally check references, watch how they handle a small job first, and see how they respond when something goes wrong. A single conversation doesn't give you that track record.

If testing is the only option: people are stuck doing the equivalent of a spot check in every single conversation, since there is no accumulated history to rely on.

If something else could help: systems could be designed to make their own consistency more visible over time, through transparent records, consistent behaviour across sessions, or some other mechanism that gives a person more than just how it responded today.

Testing in the moment currently appears to be the main practical option. Whether better mechanisms for showing consistency over time could reduce this burden is an open design question.