An interactive map of 26 research papers on whether large language models can predict the latent structure of human attitudes and beliefs.
Do language models read the structure of belief?
Whether a model can infer what you believe about guns from what you believe about the climate — when the two share no words at all.
Beliefs travel in packs
why no opinion stands alone
People's opinions don't stand alone. What you think about immigration bends toward what you think about taxes; your stance on vaccines leans on how you feel about natural versus artificial things. Three forces hold these links together. Ideology bundles positions into broad worldviews. Social influence means we absorb the attitudes of the groups we join — and join groups that already share ours. And coherence stitches beliefs together through meaning and cause, so a clash between them registers as the discomfort of cognitive dissonance.
The hard question isn't whether a model can guess your view on abortion from your view on contraception. Those sound alike, and a model could pass just by matching words. The real test is whether it can infer a link between two attitudes with no words in common. Gun control and reproductive rights share nothing on the surface; the connection runs entirely through ideology. A model that captures that has learned something about the latent structure of belief — not merely the vocabulary.
The whole question in one schematic. Two attitudes with almost no shared words can still be tightly linked: the faint line is what the words share, the solid path is the real connection — running underneath, through ideology and coherence.
That single distinction — surface similarity versus latent structure — is the axis the whole field turns on. Below is the working vocabulary you need to read any of the papers in the library.
GPT-4o reads the geometry, but presses too hard
the controlled test, across three studies
Ma and Powell (Arizona State, 2025) ran the controlled test on GPT-4o. The verdict: the model genuinely taps the latent structure — it isn't just matching words — but it over-polarises, turning loose human correlations into near-certainties.
Real people, 64 attitudes
376 US adults rated agreement with 64 diverse statements drawn from Pew's surveys (the basis of the OpinionQA benchmark) on a five-point scale. Attitudes correlated strongly both within and across topic areas. Agreement that government should prioritise climate change and that more guns is bad for society correlated at r ≈ 0.58 — two topics with nothing in common on the surface.
Does the model reproduce the correlation matrix?
For every pairing of the 64 statements, GPT-4o predicted one attitude from another. Its predicted correlation matrix tracked the human one at r ≈ 0.77, but its estimates piled up near ±1 where humans were spread out — it over-polarised. Word-similarity (cosine) predicted association strength only weakly (r ≈ 0.33), and keeping only the dissimilar pairs barely lowered the match with humans (still r ≈ 0.72).
Guessing your answer from your other answers
Predicting each person's answer to a target question from their other answers, GPT-4o ran about 40–45% accurate against a five-way chance rate of 29.4%. The decisive result: in the cases where dissimilar items were genuinely the more predictive ones, GPT-4o was still more accurate using them — proof it isn't merely similarity-matching. It tracked the trained-on-humans ceiling a touch more closely with similar items, so surface similarity does give it some lift.
Above chance, below the ceiling. On a five-way choice (chance 29.4%), GPT-4o lands well above guessing — and close to a random forest trained directly on the human data. Hover any bar for the exact figure; the dashed line marks pure guessing.
The model reads the latent geometry of belief, but draws the lines harder than reality.
Two caveats. The model was limited to five-point outputs and forced to answer immediately, with no chain-of-thought. The obvious next experiment is to hand the task to a reasoning model and see whether step-by-step inference sharpens the predictions or softens the over-polarisation.
An inter-concept map
the connective tissue beneath the 26 papers
The papers share a small set of concepts, and those concepts lean on each other. Tap any concept to trace its links and jump straight to the papers that build on it.
The research, by thread
26 papers, grouped by how they relate to the keystone
Filter by thread or search the text. Open any card for its summary, key findings, the one insight that makes it distinctive, and a link to the source.
The shape is easier to read than to exploit
Put the threads together and a tension emerges. The keystone paper and its nearest neighbours show that models genuinely encode the structure of belief — Lee and colleagues even turn belief distance into a measurable quantity that predicts both the beliefs you will adopt next and the dissonance you will feel. The fear that naturally follows is microtargeted manipulation: if a machine can read your belief geometry, surely it can exploit it.
Yet the largest persuasion studies point the other way. Across roughly 77,000 people and 19 models, how a model was trained and prompted mattered far more than knowing who you are; a meta-analysis finds today's models persuade about as well as humans — no better. And the silicon-samples critics add a quieter warning: the alignment that makes models safe also makes them bland and culturally provincial, strongest where the training text is and weakest on the sensitive, identity-laden attitudes where structure matters most.
So the risk to watch may be broad, fluent persuasion at scale — not bespoke profiling. Knowing the shape of your beliefs turns out to be easier than exploiting it.