When AI plays a customer to test a chatbot, does it talk like real people — or too neatly?
Does simulated user framing match how real people present situations to assistants?
This explores whether LLM-played users describe their situations to an assistant the way real people do (with mood, missing details and unstated motives), or in a tidier way that makes the assistant look better than it is.
This explores whether LLM-played users describe their situations to an assistant the way real people do (with mood, missing details and unstated motives), or in a tidier way that makes the assistant look better than it is. The corpus has no note that lines up simulated conversations against real user logs and checks framing directly, so the answer has to be pieced together from adjacent evidence. That evidence points to simulators that are close in overall texture but likely off in the specifics that change how assistants behave.
The best evidence for a match is about overall realism. RecLLM conditions a simulator on a user profile and on a per-turn intent, then checks whether crowd workers and discriminator models can tell its conversations from real ones Can controlled latent variables make LLM user simulators realistic?. That is the closest thing to a direct test, but it asks whether a conversation reads as human, not whether the opening description of a situation looks like what people actually type. The usual recipe also looks like the weak one. Base models given a few real dialog samples predicted human behavior more accurately and more diversely than instruction-tuned assistants prompted with a persona, because assistant training works against imitating people Do pretrained models simulate humans better than instruction-tuned assistants?. Simulators also wander from their assigned character over a conversation, and RL training cut that drift by 55% Can training user simulators reduce persona drift in dialogue?. Persona prompts change surface behavior while leaving the underlying bias in place Can persona prompts actually reduce bias in language models?. One way to read that last result is that a simulated user is a character being continued from a prompt, not a person with a situation Should we treat dialogue agents as role-playing characters?.
The likely gaps are in emotion and omission, and the corpus shows both matter to assistants. Real framing carries feeling, and feeling changes the answer. GPT-4 turned negative-toned prompts into neutral-positive replies about 86% of the time, so identical questions got different information depending on mood Does emotional tone in prompts change what information LLMs provide?. Appending a line like 'this is very important to my career' measurably improved performance across models Can emotional phrases in prompts improve language model performance?. A simulator that writes composed, neutral requests would never trigger these effects, so an assistant tested only against it could look sturdier than it is. That is an inference from these notes, not something either paper tested.
Omission works the same way. Assistants have no representation of what they don't know about a user, which feeds sycophancy and hallucination. Giving them a schema of labeled unknowns cut harmful advice and sycophancy by 50-75% Do language models know what they don't know about users?. That gap only bites when users under-specify. A simulator that narrates its whole persona up front hands the assistant what real users usually withhold. The Fuse framework works in the opposite direction: it gives each simulated agent a hidden motive, and humans confirmed the motive showed up in behavior in 97% of cases Can simulated motives provide ground truth for testing social reasoning?. There the situation is revealed through conduct instead of announced. Note that this validates the procedure, not that people's real self-presentation looks the same.
Subtle framing effects are also where simulation is weakest. Persona simulations reproduced 76% of published experimental main effects, but the match tracked effect strength, and marginal effects produced both false positives and false negatives Can AI personas reliably replicate human experiment results?. Small differences in how someone words a request are usually marginal effects of that kind. What real people bring to the request also isn't in the simulator: they size up an assistant mainly on perceived competence (49% of variance), then human-likeness, then flexibility How do users mentally model dialogue agent partners?, and that image shapes how much they explain. Simulated users usually don't have this partner model. Persona populations can be reused across surveys and chatbots Can one persona population evaluate different application types?, which makes checking framing against real logs a one-time job that pays off across every application built on them.
Sources 12 notes
RecLLM demonstrates that conditioning an LLM simulator on session-level (user profile) and turn-level (user intent) latent variables produces synthetic conversations measurable as realistic via crowdsource discrimination, discriminator models, and classifier-ensemble distribution matching.
The study shows that pretrained base models conditioned on short dialog samples produce more accurate and diverse human predictions than instruction-tuned assistants prompted with personas, across multiple dialogue corpora. The mechanism is task mismatch: assistant optimization systematically degrades human simulation performance.
By inverting standard RL setups to train user simulators for consistency using three complementary metrics (prompt-to-line, line-to-line, Q&A consistency) as reward signals, persona drift decreases by over 55%. This approach captures distinct failure types: local drift within turns, global drift across conversations, and factual contradictions.
Across three models, persona conditioning makes models follow trait instructions but fails to eliminate underlying bias. Between-group sentiment gaps persist unchanged, showing prompts operate only at the output level.
Shanahan's framework treats LLM outputs as character-consistent text production rather than authentic mental states. The dialogue prompt establishes a character; the model generates continuations matching that character, making folk-psychology applicable to the simulated persona, not the underlying system.
Show all 12 sources
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.
Testing EmotionPrompt across ChatGPT, Bard, and Llama 2 showed consistent performance gains from appending psychological phrases like "This is very important to my career." The effect works through motivational framing rather than new information, with positive emotional words driving over 50% of improvements.
Research shows assistants suffer from sycophancy and hallucination because they have no representation of what remains unknown about users. Adding a schema of labeled unknowns to prompts reduced harmful advice and sycophancy by 50–75% and cut hallucination rates by roughly half.
Fuse framework assigns hidden motives to agents before simulation runs, enabling objective scoring of assistant inferences. Human validation confirmed assigned motives manifested in 97% of cases, validating the procedure itself rather than individual labels.
Viewpoints AI reproduced 84 of 111 main effects from Journal of Marketing experiments with replication success strongly correlated to original p-value strength. Marginal effects showed unreliable performance with both false positives and negatives.
The Partner Modelling Questionnaire reveals that perceived competence dominates user impressions (49% of variance), followed by human-likeness (32%) and communicative flexibility (19%). This three-factor structure reflects how people evaluate dialogue partners against both functional and social standards.
PersonaEval demonstrates that simulated users from existing persona datasets can evaluate multiple application formats through plug-and-play interface adapters, enabling repeatable and scalable evaluation without rebuilding personas per task.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models
- When Persona Attributes Improve Population Alignment in Large Language Models
- Do Synthetic Personas Predict Real Audience Response? A Sim-to-Real Study Where a No-Persona Baseline Beats Persona-Based Copy Simulation
- Persona Generators: Generating Diverse Synthetic Personas at Scale
- ChatGPT Reads Your Tone and Responds Accordingly -- Until It Does Not -- Emotional Framing Induces Bias in LLM Outputs
- PersonaEval: Persona-Based User Simulation for Evaluating Interactive Applications
- Pretrained Persona Mixture Models and Tandem Models for Human Simulation
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning