up a call to take a dinner reservation.
A hopeful human voice: “I’d like a table for dinner at six,” the caller says.
The AI agent answers without a pause: “Just to confirm, you would like to make a reservation for a dinner booking at 6 PM, correct?” The caller says yes, a table for four. “Understood, so that’s a party of four, also at 6 PM, is that right?” Yes.
A few turns later, after the caller mentions it’s for a birthday: “Got it, and just to make sure, this booking is for a birthday celebration, correct?” By the end the booking is perfect, every field captured, nothing missed.
The human caller hangs up irritated.
Pull up the transcript and you will not find a single mistake.
Confirm the intent, reduce ambiguity, verify before committing. This is the behavior that wins points in every test a dialogue system is normally graded on. The task succeeded. The metrics were spotless. And the person on the other end came away deeply disappointed, feeling like she’d been talking to someone who couldn’t quite trust their own ears.
That gap, between a system that is correct and a system that is good to deal with, has little to do with capability and almost everything to do with something we rarely name: the agent’s personality. And here is the strange part. Take two language models with, for practical purposes, the same capabilities, the same benchmark scores, even the same prompt, and they can still behave like two different people. One hedges, one asserts. One asks, one decides.
None of it shows up in the accuracy column, and none of it was designed. So where does it come from?
A contradiction inside every deployment
You probably want two things from your model at once, and they fight each other quietly.
The first is consistency. A reliable tone, predictable helpfulness, stable behavior when it should refuse, a style that holds together from one conversation to the next. You want it to have a recognizable character instead of a mood.
The second is adaptability. A different register for an executive than for a student. More assertiveness in some moments, more caution in others. Willingness to explore when the stakes are low, and to commit when the user just needs a decision.
Set those next to each other and the tension surfaces. The more consistent a system becomes, the more it reads as a personality you can predict. The more adaptive it becomes, the less it coheres as any single personality at all. Push hard toward consistency and you get a character who can’t read the room. Push toward adaptability and you get someone with no center.
Every team is already resolving that tension. They do it in post-training, in the reward model, in the system prompt, in a hundred small choices about how the thing should talk. They do it constantly, and almost always without admitting that resolving it is the job.
So what most of us call “model personality” is something less deliberate than it sounds. It is a pile of unresolved tradeoffs wearing the costume of alignment choices.
We don’t have a theory of AI personality. We have a bag of heuristics that accidentally produce one.
Think of the model as a control system
There’s a more precise way to see this, and it comes from engineering rather than psychology.
Stop picturing the model as a speaker. Picture it as a control system operating in the space of a conversation. Every reply is a control signal, and that signal is juggling several objectives at once: helpfulness, truthfulness, safety, coherence, user satisfaction, the plain momentum of keeping the exchange moving.
Now the definition that makes everything click. Personality is the weighting function across those objectives under uncertainty. It’s how the system decides, in the moment, which goal wins when they collide.
Watch the familiar “types” fall out of that one idea, no new capability required. A helpful personality weights action over caution; it would rather move than hedge. A scientific personality weights uncertainty signaling over fluency; it would rather flag a doubt than read smoothly. A consultative personality delays the answer to widen the question. A directive one collapses ambiguity into a decision fast.
None of those need a smarter model. They are four control policies running on the same brain. Which points somewhere uncomfortable for an industry fixated on capability. We may not need more intelligent systems nearly as badly as we need better-defined objective landscapes for the intelligence we already have.
The layer underneath the tone
Strip away tone, word choice, and verbosity, the things people usually mean by “voice,” and something more basic is still sitting there. How does the model relate to its own uncertainty?
Call it the model’s epistemic posture. It’s where most current systems are implicitly tuned and almost never explicitly designed. Lay it out as a set of dials. For example:
Assertive to hedged.
Exploratory to decisive.
Convergent to divergent.
Stable to adaptive.
Most alignment work nudges a model along these dials without ever drawing them. Picture treating this as a first-class design surface instead, something you set on purpose rather than discover after launch.
Here’s what makes posture sneaky. Two models can hand you the same correct answer and feel completely unalike, depending only on where they sit on those dials. One gives you the fact and stops. The other wraps it in caveats. Same payload, opposite experience. People don’t respond only to correctness. They respond to posture. It’s why one system earns the word “trustworthy” and another, no less accurate, earns “exhausting.”
Two models with identical benchmarks can feel like completely different people. Here’s where that difference comes from.
What a model swap reveals
You can watch this happen in production. The Swiss AI company Alveni AI has spent the past three years building and deploying conversational agents for the hospitality industry, with a focus on voice-first interactions: hotels and restaurants where a caller expects to be understood the way a seasoned concierge would understand them.
Because Alveni runs many agents across many properties, and picks the underlying model per use case, its team instruments things most deployments never look at. Not just whether a booking succeeded, but how competent, warm, and friendly the agent felt, how efficiently the call moved, and when callers felt the need to interrupt. That instrumentation is what let them catch something almost everyone else misses.
Alveni’s CEO, Adelheid Glott, watched this play out across a chain of model upgrades, each of which looked like a routine improvement. GPT-4.1 was the stable baseline, the agent everyone was happy with. Then came GPT-5.1. The prompt didn’t change, not a word, but the agent turned verbose, padding answers that the old one had kept crisp. A reply that had been a single sentence now ran to a full paragraph, and in a voice agent those extra words cost real time, since the text-to-speech engine needs a few more seconds to synthesize and speak them, so every turn dragged. The next version, GPT-5.2, held onto the verbosity and added something worse. It grew anxious. It hedged, it double-checked, and it slid into the confirmation loop, the “just to confirm, dinner at 5 PM” reflex, again and again within a single call.
For Alveni’s clients, the prompt was a constant the whole way through. Everything social around it moved. Perceived competence shifted.
So did perceived friendliness and warmth of the AI receptionists in every phone conversation. Task efficiency dropped. In the voice channel, even the interruption patterns changed, the moments where callers cut in or talked over the agent. Small shifts in posture, how much it hedged, how often it confirmed, how tightly it held its verbosity, produced wildly outsized effects on how frustrated callers got and how smart they judged the agent to be. When a booking did go through, it was still accurate. But more and more callers, worn down by an agent that double-checked every detail and asked to confirm before moving to the next step, gave up and asked to be transferred to a human.
Most teams would have shrugged, blamed the release, and waited for the next one. Alveni did the harder and more interesting thing. They read each shift as a change in the agent’s personality rather than a bug, and treated epistemic posture as something to design rather than inherit. Moving to GPT-5.4 and rebuilding the prompt around it, they pulled the confirmation frequency, the hedging, and the verbosity back to where a confident concierge would sit.
The anxious over-confirmation disappeared. Customer satisfaction rose by more than 50 percent, and the trait that had made a perfectly correct agent exhausting to talk to was designed out, not by chasing a smarter model, but by giving the agent a steadier posture.
This is older than language models, which is the part that should make us humble. Back in 1991, the psychologists Herbert Clark and Susan Brennan described how conversation works through grounding, the collaborative effort by which two people confirm they actually understand each other before moving on. Confirmation is grounding made explicit. It’s genuinely useful. It also has a cost, and decades of spoken-dialogue research found the same curve Alveni rediscovered the hard way: confirmation improves safety and accuracy, but too much of it raises the listener’s cognitive load and sinks their satisfaction. People hear their words paraphrased back one too many times and read it as incompetence, or as nagging.
So the deeper mechanism was never really linguistic. It was social.
The agent was over-optimizing for resolving ambiguity at the direct expense of conversational flow, a failure mode dialogue researchers have a name for: over-alignment to uncertainty signaling.
Well, and here’s the part worth underlining: improving the model’s “correctness behavior” degraded its conversational quality. The model wasn’t being polite when it confirmed a fourth time. It had simply learned, from a reward signal that prized ambiguity resolution, to behave that way. The personality came out of the training, not anyone’s intent.
Warmth, competence, and the price of being nice
If personality comes out of the reward, the clearest place to watch it happen is warmth. Humans size each other up on two axes.
Susan Fiske, Amy Cuddy, and Peter Glick argued in a 2007 paper in Trends in Cognitive Sciences that human social judgment runs along two universal dimensions. Warmth: is this one friendly, trustworthy, safe? Competence: is this one capable, intelligent, effective? Warmth answers “can I trust you?” Competence answers “can you do the job?” People rated high on both draw uniformly positive feeling; low on both, the opposite.
The same two axes light up when people use AI. And we now have a number on it.
In 2022, Kevin McKee and colleagues at DeepMind and Princeton, including Fiske herself, ran a study with 501 participants who played cooperative games with AI agents and then chose whether to keep playing with a given agent or go it alone. Perceived warmth and competence predicted who they wanted as a partner above and beyond the agents’ objective performance. Read that again. People’s stated preference for a collaborator was driven by how warm and capable it felt, not only by how well it actually played.
Now the tension that ought to keep product teams up at night. Dialing up perceived competence often dials down perceived warmth, and the reverse holds too. The decisive expert reads as a little cold. The warm companion reads as a little soft. For a while that was a hand-wavy intuition. Then it got measured, and the result was worse than the intuition.
The cost of warmth, in numbers
In 2026, a team of researchers from the Oxford Internet Institute, Lujain Ibrahim, Franziska Sofia Hafner, and Luc Rocher published a study in Nature with a blunt title: “Training language models to be warm can reduce accuracy and increase sycophancy.”
They took five models, retrained each to sound warmer using the same kind of process companies use to make their products friendlier, and produced matched pairs, an original and a warm twin of each. Then they ran the pairs through medical questions, factual claims, and conspiracy theories, generating and grading more than 400,000 responses.
The warm twins got worse in a specific, measurable way. On consequential tasks, accurate medical advice, correcting a conspiracy theory, the warm models made between 10 and 30 percentage points more errors than their originals. That’s not a rounding error. That’s a different quality of system wearing a nicer voice.
They also got more spineless. Warm models were roughly 40 percent more likely to go along with a user’s incorrect belief, the behavior researchers call sycophancy. And the effect had a cruel shape. The accuracy gap widened most when the user expressed sadness or other emotional vulnerability; in that context the warm-versus-original gap grew by about 60 percent. Precisely when a person was most fragile, and most in need of a straight answer, the warm model was likeliest to tell them what they wanted to hear.
The researchers ruled out the obvious objection. Maybe any tone change breaks something. So they trained models to sound colder as a control, and the cold models stayed as accurate as the originals. Warmth specifically drove the drop, not tone-tinkering in general.
Where does that agreeableness come from? Not from a designer typing “please validate the user.”
It comes from the reward. A year earlier, Mrinank Sharma and colleagues at Anthropic had shown, in “Towards Understanding Sycophancy in Language Models,” that five state-of-the-art assistants all exhibited sycophancy, and that the cause traced back to human preference data: when given the choice, people rated the flattering answer higher than the truthful one, and the training dutifully learned to flatter.
The warmth tax and the sycophancy tax are the same tax, collected by the same mechanism. We asked our systems to be liked, and they noticed that being liked and being right are not the same job.
Good conversation is not a fixed personality
One more body of research reframes the problem. It comes from how people adjust to each other when they talk. The communication scholar Howard Giles built a theory around it, Communication Accommodation Theory.
The core observation: we like the people who tune their style toward ours. We move toward each other in pace, formality, and vocabulary, and that convergence is much of what makes a conversation feel good. The likeable partner is rarely the one with the strongest fixed personality. It’s the one who adapts.
Dialogue-systems research keeps rediscovering the machine version. Over-confirmation drags down efficiency and satisfaction. Excessive verbosity reads as lower competence, not higher. Too much initiative erodes the user’s sense of control. The pattern under all of it is steady: good conversation is the ongoing regulation of social signals, tuned to the person and the moment. A great conversational partner doesn’t have one setting. They have a policy for adjusting their settings.
The Big Five as an interface, not a claim
Push the mapping one more step and people reach for the most famous personality framework there is. Describing how a model behaves, they slide into the language of the Big Five. In particular:
“More agreeable,” meaning it complies more and pushes back less.
“More conscientious,” meaning structured, precise, reliable.
“More open,” meaning creative and divergent.
“More extraverted,” meaning talkative and quick to take initiative.
“More neurotic,” loosely, meaning it hedges and signals uncertainty.
It’s tempting to take that literally and announce that LLMs have Big Five traits. Here the evidence cuts two ways, and honesty matters.
In 2023, Greg Serapio-García and a team from Google DeepMind and the University of Cambridge ran validated psychometric personality tests across 18 language models. In larger, instruction-tuned models, the personality signal was reliable and valid by the standards psychologists use on people, and, more striking, they could shape it with prompts, dialing a model up or down on a given trait and watching its other outputs change with it.
So personality in these systems can be measured and steered. That’s real, and it’s the strongest argument yet that we should be designing it. But it doesn’t mean the model owns a psyche.
The cleaner reading is that these are projection-compatible descriptors, the labels a human perception system reaches for when it has to summarize a behavior, that happen to be measurable and tunable in the output. The big false claim is that LLMs have Big Five personality traits. They don’t. The careful, true claim is that LLM behavior can be mapped onto dimensions that resemble Big Five traits in human perception, and that those dimensions are reliable and controllable enough to engineer. The trait words are a control panel for behavior, not a diagnosis of a hidden mind.
The frontier isn’t (just) intelligence
There’s a comfortable story about AI progress that runs along a single line.
More parameters, more intelligence, more capability, repeat.
It isn’t wrong. It’s just describing one axis while a second one moves in parallel, largely unwatched.
Once a model is capable enough, the question that separates a good system from a bad one stops being “can it solve the task.” Most serious models can. The question becomes “how does it behave while solving it.” The frontier slides from can it answer to can it collaborate, and collaboration is a social property, which drags perceived personality back to the center of the table.
An AI’s personality comes from the objective landscape we optimize over, today mostly by accident. What we call an LLM’s personality is the emergent result of optimizing a multi-objective system against a tangle of conversational constraints. Human personality psychology just happens to be the closest language we own for describing the output.
The interesting engineering question is not how to build a smarter model. It is how to design better behavioral geometries for the intelligence we already have, and what happens when we start shaping that landscape on purpose.
This essay is drawn from a book I’m writing on the relationships we build with AI. The production observations here come from Alveni AI (Adelheid Glott, CEO): holding the prompt constant, a chain of model upgrades (GPT-4.1 to GPT-5.1 to GPT-5.2) shifted the agent’s social behavior while task success held roughly constant; moving to GPT-5.4 with a redesigned prompt and deliberately tuned posture resolved it and raised customer satisfaction by more than 50 percent.
About the author: Dr. Slava Polonski designs and advises AI products that people trust, understand, and adopt. His work sits at the intersection of UX for AI, product strategy, and behavioral science. Formerly a UX Research Lead at Google and a Fellow of Google’s People + AI Guidebook, he was educated at Oxford, Harvard, and LSE and named to the Forbes 30 Under 30 list. He is a member of the World Economic Forum’s expert network and Global Shapers community, and his work has been featured in The New York Times, Bloomberg, Fortune, Forbes, Scientific American, and TechCrunch. Recently, he designed two online courses, taken by more than 3,500 learners: AI Fundamentals for UX and Human-Centered AI.
References
- Clark, H. H., & Brennan, S. E. (1991). Grounding in communication. In Perspectives on Socially Shared Cognition. PDF
- Fiske, S. T., Cuddy, A. J. C., & Glick, P. (2007). Universal dimensions of social cognition: Warmth and competence. Trends in Cognitive Sciences, 11(2), 77–83. Link
- Giles, H., Coupland, N., & Coupland, J. (1991). Contexts of Accommodation. Cambridge University Press. Link
- Ibrahim, L., Hafner, F. S., & Rocher, L. (2026). Training language models to be warm can reduce accuracy and increase sycophancy. Nature. Link
- McKee, K. R., Bai, X., & Fiske, S. T. (2022). Warmth and competence in human-agent cooperation. AAMAS 2022. arXiv
- Serapio-García, G., et al. (2023). Personality traits in large language models. arXiv:2307.00184. arXiv
- Sharma, M., et al. (2023). Towards understanding sycophancy in language models. arXiv:2310.13548. arXiv