William Jones·March 28, 2026·6 min read

The Science Behind AI Personality: What Researchers Have Proven So Far

researchOCEANAIpsychology

Most people think of AI personality as a gimmick. You tell ChatGPT to "be sarcastic" and it adds some eye-roll energy to its responses. That's not personality — that's cosplay.

Real personality is stable, measurable, and predictive. It shapes how someone makes decisions, handles conflict, and responds to new information. And a growing body of peer-reviewed research shows that large language models exhibit exactly these properties.

Here's what the science actually says.

AIs have measurable personalities

In 2023, Serapio-García and colleagues published one of the first rigorous attempts to measure Big Five personality traits in AI models. They administered standardized personality inventories — the same validated instruments used in clinical psychology — to multiple LLMs.

The result: AI models don't score randomly. They produce consistent, interpretable personality profiles. GPT-4 tends toward high agreeableness and conscientiousness. Claude shows different patterns. Each model has a measurable baseline personality, just like humans do.

This wasn't a one-off finding. In 2024, Sorokovikova et al. replicated and extended the work, showing that these personality profiles are stable across repeated measurements and differ systematically between models. The same model, asked the same questions at different times, gives consistent answers. Different models give different answers.

In plain English: AI personality isn't noise. It's signal.

You can assign a personality and it holds

The next logical question was: if AIs have default personalities, can you override them? Can you tell an AI to be a specific type of person and have it actually behave that way?

Jiang et al. answered this in their 2024 paper "PersonaLLM," presented at NAACL. They assigned LLM agents specific Big Five personas — high neuroticism, low extraversion, etc. — and tested whether the agents stayed in character.

The results were striking. LLM personas' self-reported scores were consistent with their assigned personality types, with large effect sizes across all five traits. Even more impressive: human evaluators who interacted with these AI personas could correctly identify their personality traits with up to 80% accuracy.

This is the finding that underpins Synthicant's entire approach. When you move the Openness slider to 5 and the Neuroticism slider to 1, you're not adding flavor text. You're activating a behavioral pattern that the AI will maintain throughout the conversation.

Personality persists over time

A common objection is that AI personality might be superficial — a few sentences of in-character dialogue before the model reverts to its default. Park et al. addressed this head-on in their landmark 2023 paper "Generative Agents."

They built a simulated town of 25 AI agents, each with distinct personality traits, and let them live their lives over multiple simulated days. The agents woke up, went to work, formed opinions about each other, initiated conversations, remembered past interactions, and planned future activities.

The key finding: personality-driven behavior persisted over extended periods. Agents maintained their individual traits — an introverted agent consistently avoided large social gatherings, a conscientious agent kept to their routines — without constant reminding.

This is widely cited as proof that AI agents can sustain believable, personality-consistent behavior over time. It's also the architectural inspiration for how Synthicant builds system prompts: structured personality traits plus contextual memory produce agents that stay in character across entire interview sessions.

Personality changes real outcomes

The most recent research moves beyond "can AI have personality" to "does it matter in practice?"

Two 2025 studies answer with a definitive yes.

The first placed personality-assigned AI agents in a simulated public environment and measured how their decisions changed based on social context. Agents with friendly, extroverted personalities showed significant discrepancies between their public expressions and private thoughts — they modulated their behavior based on whether they thought they were being observed.

If that sounds familiar, it should. This mirrors decades of human social psychology research on impression management. The AI agents weren't just reciting personality scripts — they were exhibiting second-order social behaviors that emerge from personality traits.

The second study, by Cohen et al., put personality-assigned agents into negotiation scenarios. They found that Agreeableness and Extraversion significantly affected negotiation outcomes — how much the agent conceded, how cooperatively they approached the deal, and whether they reached agreement. The personality didn't just change the words; it changed the results.

What this means for product research

Here's why this research matters if you're building products:

Synthetic personas aren't guesswork. When you interview a Synthicant persona with high Conscientiousness and low Openness, the pushback you get isn't random. It's grounded in a behavioral model that has been validated across multiple independent studies.

Personality predicts behavior, not just words. A persona with high Neuroticism won't just say "I'm worried about this." It will hesitate on purchasing decisions, fixate on edge cases, and show risk-averse decision patterns — the same way a real person with high Neuroticism would.

Different models, different baselines. This is why Synthicant uses Claude specifically. Every LLM has a default personality fingerprint that influences how well it can adopt different personas. The model matters as much as the prompt.

The research is still young, but the trajectory is clear. AI personality isn't a parlor trick — it's a measurable, steerable property with real behavioral consequences. And the better you understand it, the more useful your synthetic user research becomes.

References

Serapio-García, G., Safdari, M., Crepy, C., et al. (2023). "Personality Traits in Large Language Models." arXiv preprint arXiv:2307.00184. — One of the first rigorous measurements of Big Five traits in AI models using standardized personality inventories. Established that LLMs produce consistent, interpretable personality profiles.

Sorokovikova, A., Tikhonov, I., & Nikishina, I. (2024). "LLMs Simulate Big Five Personality Traits: Further Evidence." arXiv preprint arXiv:2402.01765. — Replicated and extended Serapio-García's findings, demonstrating that AI personality profiles are stable across repeated measurements and differ systematically between models.

Jiang, H., Zhang, X., Cao, X., et al. (2024). "PersonaLLM: Investigating the Ability of Large Language Models to Express Personality Traits." Proceedings of NAACL 2024. — Showed that LLMs assigned specific Big Five personas maintain consistent behavior, with human evaluators identifying assigned traits at up to 80% accuracy.

Park, J.S., O'Brien, J.C., Cai, C.J., et al. (2023). "Generative Agents: Interactive Simulacra of Human Behavior." Proceedings of ACM UIST 2023. Stanford University / Google Research. — Demonstrated that AI agents with structured personality traits sustain believable, consistent behavior over extended simulated periods.

Cohen, R., et al. (2025). "Exploring Big Five Personality and AI Capability Effects in LLM-Simulated Negotiation Dialogues." arXiv preprint. — Found that Agreeableness and Extraversion significantly affect AI negotiation outcomes, with sociocognitive measures detecting fine-grained behavioral differences.