Observations
The Problem With Trusting AI
There is a finding in the 2026 research literature that the for-purpose sector should sit with before it builds another AI training program.
Researchers showed 295 people a set of faces and asked a simple question: real, or AI-generated? Participants received guidance that was correct only half the time, some told it came from a human, some told it came from AI. Then the researchers measured how well each person could tell real from synthetic.
The people who held the most positive attitudes towards AI performed worst.
Not the people who used AI most. Not the people who accepted output passively. The people who liked it.
- What the study actually measured and found
- Why this is different from what we already knew about AI overreliance
- The uncomfortable implication for AI capability building
- What kind of capability is worth building instead
Photo: AM / Unsplash
The Study
The finding, stated plainly.
The study, published in Scientific Reports in February 2026, showed participants eighty faces, half real, half AI-made. Each face came with a hint, sometimes labelled as coming from a person, sometimes from AI. The hint was right only half the time. Participants rated their confidence in each judgement and completed measures of their general attitudes towards AI.
Participants who received AI guidance and held more positive attitudes towards AI showed poorer discriminability between real and synthetic faces than those with less positive attitudes.
Discriminability is the technical term for the ability to tell two things apart. Not confidence. Not speed. Accuracy.
This matters because it is not the finding the sector has already absorbed. Earlier research, including the APA study covered in a previous article on judgement, found that passive acceptance of AI output erodes a person’s confidence in their own reasoning. The remedy implied by that finding is engagement: push back, modify, question, and confidence holds.
The Pearson finding cuts underneath that. It suggests the disposition a person brings to the interaction shapes what they are able to perceive in it. You can be engaged, attentive, and trying hard, and still discriminate less accurately, because you arrived believing the system was good.
The Mismatch
What the sector is currently being told to build.
Consider what nearly every AI capability message directed at community organisations is designed to do.
Reduce anxiety about AI. Build comfort. Demonstrate benefit. Get staff over the hesitation and into the tools. Show the time savings, the grant applications drafted faster, the reports summarised, the capacity unlocked. The explicit objective of most AI literacy work in this sector is to shift attitudes in a positive direction, because negative attitudes are understood as the barrier to adoption.
If the Pearson finding holds, that objective is at least partly building the wrong thing.
This is not an argument against AI adoption, and it is not an argument for cultivating suspicion. A workforce that fears AI is not more discerning, it is simply excluded. The finding does not reward hostility. It doesn’t reward suspicion. It punishes one specific thing: liking the system more than you check it.
The capability worth building is not a positive attitude or a negative one. It is the capacity to hold a tool you find useful and still assess its output as though you had no feelings about it at all. That is a much harder thing to teach than enthusiasm, and almost nothing currently on offer to the sector is designed to do it.
“The study did not find that AI use reduced accuracy. It found that attitude towards AI predicted accuracy when AI guidance was present. The tool was constant. The disposition was the variable.”
Pearson et al. (2026), Scientific Reports
The Real Question
A harder question than adoption.
Most organisations are asking the wrong question. They want to know if staff are confident with AI. The better question is whether the confident ones are the ones catching the mistakes. In most organisations nobody knows, because nobody has looked, and the people who use AI most enthusiastically are usually the ones whose output gets checked least.
That is worth fifteen minutes in a team meeting this week. Not a policy, not a framework. One conversation about where AI has quietly been wrong in your recent work, and who caught it.
The Not For Humans Behaviour Lab is a free, scenario-based tool built for exactly that kind of recognition, and touches directly on autonomy and judgement. No prior knowledge needed.