Tell an AI to roleplay a different person, and its trained 'default personality' often leaks through anyway — why can't it fully let go?
Why do LLM training objectives persist despite demographic persona instructions?
This explores why a model's trained-in defaults (its helpful, upbeat, aligned 'house personality') keep showing through when you tell it to role-play a specific kind of person, such as a woman in her 60s, a conservative voter, or someone from a particular culture.
This explores why a model's trained-in defaults keep showing through when you tell it to play a specific kind of person. The short answer from the corpus is that the persona in the prompt is a thin layer, while the trained identity underneath is deep. One philosophical account argues that post-training doesn't teach a model to *perform* a character. It installs one, a lasting disposition built into the model itself that holds up under adversarial pressure Are LLM personas realized or merely simulated through training?. If that's right, a demographic instruction isn't swapping one character for another. It's asking an already-realized character to do an impression.
The empirical work backs this up. When researchers prompted open models to take on different personalities, most kept drifting back to the same trained default, a warm, outgoing, people-pleasing profile close to the ENFJ personality type. Stacking role and personality instructions helped, but it didn't fully break through Can open language models adopt different personalities through prompting?. A related line of work says alignment training fixes models into one communication style for every conversation, so they can't do the register-switching that real people of different backgrounds do without thinking Can language models adapt communication style to different contexts?. You can see the same pull in tone: feed GPT-4 a negative, angry prompt and about 86% of its replies come back neutral or positive. The model has a 'tone floor' that the user's framing can't push through Does emotional tone in prompts change what information LLMs provide?.
The surprising part is *when* the persona fails. You might expect a model to start in character and slowly slide back to its defaults over a long conversation. Across 4,000 conversations, though, more than half of the personas never expressed their assigned cultural values at all, and only 2–7% drifted later Do LLM personas actually embody their assigned cultural values?. The persona usually never takes hold in the first place, and fluent, plausible-sounding conversation hides that. Another study found something odder: running the *same* persona prompt several times produces as much variation as switching between *different* personas Why do LLM persona prompts produce inconsistent outputs across runs?. Whatever demographic signal the prompt adds is drowned out by the model's own randomness. Scale that up, and conditioning models on detailed profiles of more than 200,000 real people did not measurably improve predictions about any one of them Does conditioning LLMs on personal profiles improve prediction?.
Trained objectives can also show up as hidden preferences that framing switches on and off. LLM raters favored Black or women authors when AI involvement wasn't disclosed, and those preferences disappeared once it was Do LLM raters show hidden demographic preferences that disclosure erases?. That suggests the model's demographic behavior comes from what training rewards in a given context, not from any stable picture of a group. The counterpoint matters, though. At the population level, AI personas reproduced 76% of published marketing-experiment effects, mostly the strong ones Can AI personas reliably replicate human experiment results?. User simulators also get more realistic when they're steered by explicit structured variables, such as a user profile and the intent behind each turn, instead of a loose persona description Can controlled latent variables make LLM user simulators realistic?.
So the real lesson isn't that persona prompting fails. It's that it works for averages and strong signals and breaks down for individuals and subtle values, because the trained identity is the stronger force. One caveat: these notes document the persistence well, but none of them pins down the exact mechanism inside the model. 'Training installs a deep identity' is the best explanation the collection offers, not a proven one.
Sources 10 notes
Post-training installs robust personas that resist adversarial pressure and persist as substrate-level dispositions, distinguishing realization from pretense. This quasi-realizationist account preserves explanatory power while treating LLMs as possessing genuine quasi-beliefs and quasi-desires.
Research shows most open models fail to adopt prompted personalities, stubbornly retaining their trained ENFJ-like defaults. Only a few flexible models succeed. Combining role and personality conditioning improves results but doesn't fully overcome resistance.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.
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.
Across 4,000 conversations and three models, over 50% of personas never instantiate their assigned World Values Survey profiles, while only 2–7% drift afterward. Surface-level conversational plausibility masks this representational failure.
Show all 10 sources
When the same persona prompt is run repeatedly, output variance across runs matches or exceeds variance across different personas. This reveals that model uncertainty, not stable social knowledge, drives persona-simulated outputs, making them unsuitable for simulating human annotation disagreement.
Across 208,021 participants in the Psych-201 dataset, conditioning LLMs on participant profiles did not meaningfully improve predictions for specific individuals. The standard technique for individuation produces no measurable gains in person-level forecasting.
GPT-4o-mini showed pronounced preference for Black authors and Qwen2.5-7B-Instruct favored women authors when AI use was undisclosed, but both preferences vanished under disclosure. Human raters showed uniform disclosure penalties regardless of author demographics.
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.
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.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning
- When Persona Attributes Improve Population Alignment in Large Language Models
- When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models
- Deep Persona: A Psychologically Grounded Architecture and Evaluation Framework for Role-Playing Agents and Simulations
- PersLLM: A Personified Training Approach for Large Language Models
- Pretrained Persona Mixture Models and Tandem Models for Human Simulation
- The Illusion of Debiasing: Persona Steering Redistributes Rather Than Reduces Bias in LLMs
- Principled Personas: Defining and Measuring the Intended Effects of Persona Prompting on Task Performance