Who Is in There? Robot Personality in the Age of Large Language Models

Robots used to have personalities the way puppets do, carefully scripted and fixed; now they borrow one from a language model, and keeping a character consistent has become a hard engineering problem.

Personality as a design material

People cannot help attributing personality to things that move and speak. A robot that pauses before answering seems thoughtful, one that interrupts seems pushy, and one that cracks jokes seems extraverted. For two decades, HRI researchers have tried to harness this tendency deliberately, usually borrowing the Big Five model from personality psychology: openness, conscientiousness, extraversion, agreeableness and neuroticism. Extraversion, being the easiest to express through voice, movement and talkativeness, has received by far the most attention.

Can a robot dog be an introvert?

An early and influential test came from Lee, Peng, Jin and Yan, published in the Journal of Communication in 2006. They programmed Sony's AIBO robot dog to behave as either introverted or extraverted through its verbal and nonverbal behaviour, and paired it with participants who were themselves introverts or extraverts. People recognised the robot's intended personality accurately, which was not a given.

The more surprising finding concerns attraction. A common assumption, borrowed from human relationships, is that we prefer partners similar to ourselves. Lee and colleagues found the opposite: participants enjoyed the interaction more, and rated the robot as more intelligent and socially attractive, when its personality complemented rather than matched their own. Feelings of social presence appeared to mediate this effect. This result is often misremembered as a similarity effect, so it is worth stating plainly.

Other studies have pointed the other way. Tapus, Tapus and Matarić built a socially assistive robot to encourage people doing post-stroke rehabilitation exercises, and reported early evidence that users preferred a robot whose personality matched their own. The two findings are not necessarily contradictory: a pet-like companion and an exercise coach occupy very different social roles. The general lesson is that there is no universally best robot personality, only better or worse fits for a person, a task and a relationship.

From scripts to simulators

For most of HRI's history, a robot's personality was built by hand. Designers wrote dialogue in a particular register, tuned the speed of gestures and decided how often the robot would initiate conversation. The character was narrow, but it was stable.

Large language models changed that overnight. A robot connected to an LLM can talk about almost anything, and its persona is usually specified in a short system prompt. There is now evidence that this kind of shaping does something measurable. Greg Serapio-García and colleagues administered validated personality questionnaires to eighteen language models and found that, under the right prompting conditions, the measurements were reliable and valid, more so for larger and instruction-tuned models, and that personality in model outputs could be steered towards specified human profiles.

The drift problem

Steering a persona at the start of a conversation is not the same as keeping it. Kenneth Li and colleagues tested how faithfully chat models followed their system prompts over the course of a dialogue and found significant drift within just eight rounds of conversation. They linked this to the way attention to the initial instructions fades over long exchanges.

For a robot designer this is a real headache. A care robot introduced as calm and gentle that becomes chirpy and verbose after twenty minutes is not just off-brand; it undermines the predictability that makes people comfortable with it in the first place. Practical systems now often re-inject persona instructions, summarise past turns or use separate modules to check that responses stay in character.

Who, exactly, is speaking?

There is a deeper conceptual issue too. Murray Shanahan, Kyle McDonell and Laria Reynolds argued in Nature in 2023 that it is most useful to think of a dialogue agent as playing a role, rather than as having a fixed self. In their framing, the underlying model can sustain many possible characters, and what appears to be deception or self-awareness is better understood as role play. This helps explain why an LLM-driven robot can feel inconsistent: there is no single character underneath, only the one currently being performed.

That framing is liberating for designers, because it treats persona as craft rather than as a hidden essence. It also carries ethical weight. A robot that can be instructed to be warm, flattering or agreeable can be instructed to be manipulative, and users may form attachments to a character that changes when the prompt does.

Designing characters on purpose

The older HRI literature still has a lot to teach the LLM era. It shows that people notice personality quickly, that the best fit depends on the person and the role, and that consistency is part of what makes a character trustworthy. What is new is that we can no longer take consistency for granted. Writing a good persona is now less like scripting a puppet and more like directing an improvising actor who sometimes forgets which play they are in.

References

1. Lee, K. M., Peng, W., Jin, S.-A. and Yan, C. (2006). Can Robots Manifest Personality?: An Empirical Test of Personality Recognition, Social Responses, and Social Presence in Human-Robot Interaction. Journal of Communication. Link
2. Tapus, A., Tapus, C. and Matarić, M. J. (2008). User-robot personality matching and assistive robot behavior adaptation for post-stroke rehabilitation therapy. Intelligent Service Robotics. Link
3. Serapio-García, G., Safdari, M., Crepy, C., Sun, L., Fitz, S., Romero, P. et al. (2023). Personality Traits in Large Language Models. arXiv preprint. Link
4. Li, K., Liu, T., Bashkansky, N., Bau, D., Viégas, F., Pfister, H. et al. (2024). Measuring and Controlling Instruction (In)Stability in Language Model Dialogs. arXiv preprint. Link
5. Shanahan, M., McDonell, K. and Reynolds, L. (2023). Role play with large language models. Nature. Link

Feature drafted with AI from the published research listed above, for social-robotics.com.