How Human Psychology Can Inform Better AI Systems

When AI researchers want to know whether a system "understands" people, they often reach for tools that psychology built to study people. That instinct is sound, because psychology has spent a century learning how to measure things you cannot observe directly. But the tools can be misused, and the field's own history carries warnings. This article describes four things psychology can offer AI, and four cautions about using them.

What psychology offers

1. Tasks that isolate a specific ability

The false-belief task, in which a person must predict what someone will do based on what that person believes instead of what is true, isolates one component of social reasoning [1]. The same family of tasks was used with autistic children, to ask whether a specific difficulty with attributing mental states exists [2]. For adults, the "Reading the Mind in the Eyes" test was developed as a measure of mentalising, and its revised version was published to fix psychometric problems in the first [3]. Each of these gives AI researchers a clear target and a human baseline.

2. A theory of what is being measured

The key concept in psychological measurement is construct validity: the question of whether a test really measures the theoretical thing it is named after. Cronbach and Meehl described how a test is tied to a construct through a network of evidence [4]. Messick described validity as the meaning of scores and the justification for inferences drawn from them [5]. Computer scientists have begun to import the idea. Jacobs and Wallach proposed measurement modelling as a framework for exposing the assumptions in turning unobservable constructs into metrics [6], and Raji and colleagues argued that popular AI benchmarks are often treated as general measures of progress without the construct-validity case being made [7].

3. Frameworks for emotion and social skill

Salovey and Mayer's framework describes emotional intelligence as skills of appraising and expressing emotion, regulating it, and using it to plan and act [8]. It gives AI work a vocabulary that separates perceiving emotion from managing it, which matters because a system might do the first without being good at the second.

4. Evidence that people treat machines as social actors

Nass and Moon's review shows that people mindlessly apply social rules to computers, from politeness to stereotypes [9]. That is useful for designers: the social effect of a system's behavior exists whether or not the system "understands" anything.

Four cautions

1. Human tests do not transfer unchanged

Psychological tasks were designed for human development and cognition, with assumptions about how people arrive at answers. Language models may reach a correct answer by other routes, and may be sensitive to features irrelevant to humans. Researchers who stress-tested models on theory-of-mind tasks cautioned against using human-designed psychological tests to draw conclusions about models [10], and showed that trivial alterations to a task could reverse apparent successes [11].

2. "Human" is not one reference population

Henrich, Heine and Norenzayan reviewed the comparative evidence and argued that research samples from Western, educated, industrialised, rich and democratic societies are frequent outliers compared with the rest of humanity [12]. Benchmarks that adopt tests and norms from that literature inherit that limitation. Why cultural context matters in AI looks at its consequences.

3. There may be no single right answer

Psychology has long dealt with variation between people. AI annotation practice often assumes a single ground truth. Aroyo and Welty argued that human interpretation is subjective and that annotation should capture the range of reasonable interpretations [13]. For social reasoning, where qualified people can legitimately disagree, a model scored against one "gold" label is being scored against an artefact.

4. Anthropomorphism cuts both ways

Because people are quick to see minds in machines [9], researchers must be careful about their own readings, and about how results are described to the public. "The model has theory of mind" and "the model passes this theory-of-mind test" are different claims.

What this means in practice

A sound approach to evaluating social understanding in AI borrows psychology's discipline, not just its tasks:

• State the construct and what is out of scope.
• Use families of items with perturbations, not single scripts.
• Compare with humans, and say which humans.
• Seek evidence that scores measure the intended construct, not format familiarity.
• Report disagreement instead of hiding it.

This is the logic behind our proposed framework in Research Report 002, which is a proposal and has not been validated. It is also why HUTE draws on psychology and linguistics as well as machine learning. For the wider picture, see why AI needs to understand human behavior.