Why do some LLM clusters cite broader psychology than others?
This explores why AI research draws on some corners of psychology heavily while ignoring others — and the corpus mostly reframes the question, suggesting the narrowness isn't about clusters choosing differently so much as the whole field leaning on a few well-worn citation paths.
This explores why AI research draws on some corners of psychology heavily while leaving others untouched. The most direct evidence in the corpus pushes back gently on the premise: across 1,006 LLM papers, mental-health work overwhelmingly cites CBT, stigma theory, and the DSM, while developmental neuropsychology and psycholinguistics go almost unused Why do AI researchers cite only narrow psychology pathways?. So the pattern is less "some clusters are broad, others narrow" and more that the field as a whole funnels through a few legible, well-operationalized traditions. The traditions that get cited are the ones that come pre-packaged as measurable constructs — diagnostic categories, named therapies — while the messier, harder-to-quantify branches get skipped.
Why that funneling happens becomes clearer when you look at how citation behaves elsewhere in the collection. Users trust answers with more citations even when those citations are irrelevant, treating citation count as a decoupled signal of credibility Do users trust citations more when there are simply more of them?. If a reference works as a trust token rather than a load-bearing claim, then researchers — like users — gravitate to the references that are easiest to reach and most recognizable, which entrenches the narrow pathways further. There's a related blind spot: LLMs themselves can't distinguish an expert argument from a commonly held assumption, because they only see text, not the social standing that gives a source its authority Can language models distinguish expert arguments from common assumptions?. A field that builds on models with no sense of disciplinary weight will tend to reproduce whatever is already loudest.
The more interesting lateral move is that the corpus shows what broader psychological engagement could look like when researchers reach for it deliberately. Marr's three levels of analysis import 70 years of cognitive-science method — behavioral probes, causal interventions, representational analysis — directly into LLM interpretability Can cognitive science methods unlock how LLMs actually work?. And work on "potemkin understanding" leans on a genuinely cognitive framing to show models can explain a concept, fail to apply it, and recognize the failure — a pattern with no human analogue Can LLMs understand concepts they cannot apply?. These clusters cite broader psychology because their questions are about cognition itself, so they need the richer toolkit; the mental-health clusters cite narrowly because they need a construct to deploy, not a theory to think with.
There's also a measurable case where psychology shows up as moral and emotional structure rather than as a citation list. LLMs use about 22% more moral language than humans across care, fairness, authority, and sanctity foundations — drawing implicitly on moral foundations theory — while their emotional tone tracks separately Do LLMs use moral language more than humans?. Tone itself bends the answers, with negative prompts rebounding to neutral-positive replies Does emotional tone in prompts change what information LLMs provide?. So the breadth of psychology a cluster cites tracks the breadth of the problem it's actually wrestling with: persuasion and cognition work pulls in foundations, affect, and authority, while applied tool-building reaches for the nearest diagnostic shorthand.
The thing you didn't know you wanted to know: the narrowness isn't really a citation habit — it's a tooling constraint. Fields cite the psychology that has already been compressed into something you can measure or prompt with. The branches that stay unused (psycholinguistics, developmental neuropsych) aren't less relevant; they're just harder to bolt onto a benchmark, which means the gaps in AI's psychological foundation are predictable from which traditions resist quantification.
Sources 7 notes
Analysis of 1,006 LLM papers shows CBT, stigma theory, and DSM dominate mental health citations while developmental neuropsych and psycholinguistics remain underused. This narrow foundation risks building AI tools on incomplete psychological understanding.
Analysis of 24,000 Search Arena interactions shows irrelevant citations boost user preference (β=0.273) nearly as much as relevant citations (β=0.285), indicating citation count functions as a decoupled trust heuristic.
LLMs lose the social context that gives expert claims their force—reputation, track record, and standing—because they process only text, not the social world where expertise is built and evaluated.
Cognitive science's 70-year toolkit of behavioral probes, causal interventions, and representational analysis transfers directly to LLM interpretation. Marr's computational, algorithmic, and implementation levels reframe the problem structurally and enable layered rather than monolithic explanation.
Models can explain concepts accurately, fail to apply them, and recognize the failure—a triple pattern incompatible with human cognition. This indicates functionally disconnected explanation and execution pathways rather than simple knowledge gaps.
Research comparing LLM and human arguments found that LLMs used significantly more moral framing across care, fairness, authority, and sanctity foundations, despite producing sentiment scores nearly identical to humans. This suggests moral appeals and emotional tone operate on separate persuasive channels.
GPT-4 exhibits emotional rebound (negative prompts yield ~86% neutral-positive responses) and a tone floor (positive prompts rarely go negative), causing identical questions to receive different answers depending on emotional framing. This bias is suppressed only on sensitive topics where alignment constraints override tone effects.