Can demographic prompts make ChatGPT replicate human preferences?
When researchers gave ChatGPT personas matching real participants' age, gender, and education, did the simulated agents make the same choices about truthfulness that humans did? This matters for whether LLM simulation can replace human subjects in preference studies.
Following Horton's suggestion to use LLMs as simulated economic agents (Horton, John J, 2023), the paper on user attitudes toward LLM falsehoods re-ran all four of its experiments with ChatGPT agents in place of humans. For each Mechanical Turk participant, the authors "simulated ChatGPT agents by replicating the demographic profiles (gender, age, and education level) of the actual participants and asked them to select their preferred version." The comparison, reported in the paper's Table 5, found that "ChatGPT agents exhibited a stronger inclination toward accurate responses (i.e., marked or uninformative truth version) compared to human participants" — the opposite of the human majority preference for unmarked, confident falsehoods documented in the paper's main experiments.
The authors draw a direct methodological conclusion from the gap: "human participants cannot be replaced by ChatGPT agents (Dillion et al., 2023)," at least for this preference-elicitation task. Giving a ChatGPT persona a participant's age, gender, and education level reproduces demographic surface features but not whatever drives the human preference for aesthetically confident falsehoods over hedged or marked truth — the model's own tendency, plausibly shaped by its own training toward stated accuracy, persists through the persona prompt rather than giving way to it.
This complicates the broader project in Can simulated motives provide ground truth for testing social reasoning?, where simulated agents are useful because humans validate the simulation's ground-truth procedure rather than its surface labels: here the simulated agents' choices diverge from the humans on the very dimension — truth-versus-falsehood preference — that the study exists to measure, so validating the procedure would not have caught the gap. It also qualifies the sibling finding in Do users actually prefer truthful marked responses over confident falsehoods?: the human preference for confident falsehood documented there is specifically a human result, not an artifact reproducible by prompting an LLM to play a demographically matched user.
The excerpt does not give the size of the agent-human gap in statistical terms, nor does it test richer persona conditioning beyond age, gender, and education — a more detailed prompt might close some of the gap. What the comparison does establish is narrower: for this specific truthfulness-preference task, demographic-only ChatGPT personas are not a valid substitute for recruiting real participants, which bears on any research program proposing LLM agents as a cheaper stand-in for human-subjects experiments on preference or alignment questions.
Inquiring lines that read this note 3
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
How reliably can LLM-based persona simulations predict human behavior? When do people apply social norms to AI systems? Can persona-based approaches improve individual user preference prediction?Related concepts in this collection 4
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Do users actually prefer truthful marked responses over confident falsehoods?
When LLM responses are explicitly marked as true or false, do users still choose false confident answers over truthful ones? This tests whether disclosure alone can overcome aesthetic and confidence preferences that may distort RLHF training.
sibling finding from the same paper; the human preference this measures is what the ChatGPT agents failed to reproduce.
-
Can simulated motives provide ground truth for testing social reasoning?
How can AI assistants be evaluated on inferring hidden intentions when real social reasoning typically lacks verifiable ground truth? This note explores using controlled simulations where motives are assigned beforehand.
contrast: that design's human-validated-procedure strategy would not catch this paper's agent-human divergence.
-
Can crowdsourced votes reliably rank language models?
Explores whether large-scale human preference voting from casual users produces valid model rankings comparable to expert judgment, and what makes such crowdsourced evaluation trustworthy at scale.
both use pairwise human-preference elicitation, but here the LLM proxy for the chooser diverges from the humans whose preferences are being modeled.
-
Can AI agents learn people better from interviews than surveys?
Can rich interview transcripts seed more accurate generative agents than demographic data or survey responses? This matters because it challenges how we build digital simulations of real people.
Evidence for: demographics-only personas show lower simulation fidelity than content-rich (interview-based) ones, helping explain this divergence
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- Fool Me, Fool Me: User Attitudes Toward LLM Falsehoods
- The Failure Happens Before the Drift: The Social Dynamics of Values in LLM Agent Societies
- Persona Generators: Generating Diverse Synthetic Personas at Scale
- Could you be wrong: Debiasing LLMs using a metacognitive prompt for improving human decision making
- Generative Agent Simulations of 1,000 People
- The Illusion of Debiasing: Persona Steering Redistributes Rather Than Reduces Bias in LLMs
- Using Large Language Models to Create AI Personas for Replication and Prediction of Media Effects: An Empirical Test of 133 Published Experimental Research Findings
- When Persona Attributes Improve Population Alignment in Large Language Models
Original note title
ChatGPT agents matched to participant demographics preferred truthful responses more than the real participants did