Can LLMs Emulate Human Belief Dynamics?

URL SCAN: Can LLMs Emulate Human Belief Dynamics?
FIRST LINE: Can LLMs simulate how humans form and change beliefs in social networks?

THE DISSECTION

This paper performs a behavioral fidelity audit on LLMs: it replicates a canonical belief dynamics study, runs 12 LLMs through it, and finds systematic divergence from human behavior. The authors conclude with a "sharp warning" against using LLMs as human proxies in social simulations. The paper positions itself as a caution against AI hubris in social science modeling.

On its face, this is a narrow empirical exercise in computational social science. It asks: do LLMs simulate humans accurately? The answer: no, in systematic and predictable ways.

But the question worth asking—particularly through the DT lens—is: does the accuracy of LLM simulation of human belief dynamics matter for the displacement thesis?

Short answer: no. And the authors' framing obscures this.

THE CORE FALLACY

The paper's fundamental category error is conflating simulation accuracy with economic substitution capacity.

The displacement mechanism in the Discontinuity Thesis operates through a much simpler logic than behavioral fidelity:

When AI achieves durable cost and performance superiority across cognitive work, it displaces humans from economically necessary roles—regardless of whether it models human behavior accurately.

A hiring algorithm doesn't need to understand how humans form beliefs about job preferences to eliminate the need for human resume reviewers. A trading system doesn't need to emulate human belief updating to displace human traders. A diagnostic system doesn't need to simulate how doctors form clinical intuitions to replace diagnostic labor.

The paper's warning—"don't use LLMs as human proxies in social simulations"—is a narrow methodological caution for a specific use case (social network modeling). It is not a systemic assessment of AI's displacement capacity. The authors are answering a question about epistemic fidelity when the relevant question is functional replacement.

HIDDEN ASSUMPTIONS

1. That social simulation fidelity is the primary risk vector. The paper implicitly assumes the danger is LLMs being used instead of human subjects in research. The actually dangerous use case—LLMs being deployed instead of human workers in productive roles—goes unexamined.

2. That behavioral divergence from humans is a defect. The paper treats LLM conformism and failure to match initial belief distributions as failures. But from a displacement perspective, that's the point. AI doesn't need to behave like humans. It needs to be cheaper, faster, and more reliable.

3. That social science simulation is the relevant domain for evaluating AI risk. This is a self-referential trap. The paper's authors are social scientists, so the risk they can conceptualize is simulation-of-humans risk. The existential risk (mass productive displacement) operates in domains that don't require simulating human belief formation at all.

4. That "human proxy" is the failure mode. The paper warns against LLMs being used as proxies for human subjects. But the displacement threat is that LLMs become replacements for human labor—not proxies in research, but substitutes in production.

SOCIAL FUNCTION

This is partial truth packaged as a warning, functioning as prestige signaling within the academic AI ethics genre.

The paper performs the ritual of "caution about AI" without actually addressing the structural displacement mechanism. It says: "be careful, LLMs aren't good human simulators." This is true but epistemically narrow. It allows the authors to claim they've identified a risk ("don't use LLMs as human proxies") while avoiding the harder question: that AI displacement operates through functional superiority, not behavioral mimicry.

The social function is to satisfy the academic norm of AI-skepticism in a way that doesn't threaten the broader AI development project. It's a caution about simulation methodology, not about systemic displacement. It's the difference between "don't build a bad bridge model" and "this bridge will collapse."

THE VERDICT

The paper's empirical findings are likely sound: LLMs do not faithfully emulate human belief dynamics. But the DT assessment is blunt:

The findings are irrelevant to the displacement thesis.

AI doesn't need to simulate human belief updating to sever the mass employment -> wage -> consumption circuit. It needs to be cheaper and more reliable than humans at economically necessary tasks. This paper is a methodological caution for social scientists. It is not an assessment of AI's structural threat to the post-WWII economic order.

The authors have correctly identified that LLMs fail a specific human fidelity test. They have failed to ask whether that fidelity was ever the point.

Partial truth. Prestige signaling. Epistemically narrow.