The Failure Happens Before the Drift: The Social Dynamics of Values in LLM Agent Societies

URL SCAN: The Failure Happens Before the Drift: The Social Dynamics of Values in LLM Agent Societies
FIRST LINE: Computer Science > Artificial Intelligence

The Dissection

The paper is not mainly about whether values drift. It is about whether LLMs can serve as credible stand-ins for humans in social experiments. Its most damaging finding comes before longitudinal interaction: more than half of the personas fail to express their assigned value profiles at initialization. The agents are not losing their identities; many never instantiated them. The additional 2–7% drift is secondary. The deeper defect is that plausible dialogue can conceal fabricated values and stylistic sameness.

The Core Fallacy

The paper treats fidelity to human value profiles as the decisive measure of an agent’s social utility. Under the Discontinuity Thesis, that is the wrong battlefield. Economic displacement does not require agents to reproduce human cultures, preserve WVS distributions, or behave like authentic human participants. An AI can be a poor simulator of humanity and still outperform humans at cognitive work, coordination, persuasion, analysis, and administration.

This is a diagnosis of simulation failure, not automation failure. The paper shows that current agents cannot reliably recreate the old human social world. DT predicts that recreating that world is not necessary for making human productive participation obsolete.

Hidden Assumptions

• WVS profiles are stable ground truth that transfer cleanly into dialogue.
• Better value faithfulness is inherently the correct design objective.
• Conversational realism is a valid proxy for social validity.
• Semantic variety and stylistic variety are economically necessary in the same way.
• Human institutions will continue to require authentic human proxies rather than merely effective outputs.
• Current model limitations will persist rather than functioning as temporary lag.
• If agents cannot represent humans faithfully, humans retain indispensable productive leverage.

The last assumption is the fatal one. It confuses authenticity with necessity.

Social Function

Primary classification: partial truth. Secondary function: prestige signaling and transition management.

The paper identifies a real defect: current agents often fail before drift begins, and polished language can disguise low-fidelity value representation. But its framing keeps attention on whether machines can imitate the human social order rather than on whether machines can replace human labor inside it. That makes it useful to institutions that need a technically respectable reason to postpone the harder conclusion: synthetic agents do not need to be culturally authentic to become economically dominant.

This is not necessarily intentional propaganda. Its functional effect is still anesthetic. It turns a question of systemic replacement into a question of calibration, benchmark design, and persona engineering.

The Verdict

This paper is a competent autopsy of a narrow corpse. It proves that current LLM agents are counterfeit human populations: often mis-specified at birth, only modestly altered by later interaction, and more stylistically uniform than their content suggests.

It does not refute the Discontinuity Thesis. It leaves P1–P3 untouched. If AI achieves durable superiority across cognitive work, institutions cannot preserve stable human-only economic domains, and the majority lose access to economically necessary labor, then value-faithful simulation is a governance and measurement problem—not a rescue of mass productive participation.

The agents may be too synthetic to reproduce humanity and more than synthetic enough to displace it. The old social order can die while its automated portraits remain inaccurate.