up a name to take a dinner reservation.
A hopeful human voice: “I’d like a desk for dinner at six,” the caller says.
The AI agent solutions and not using a pause: “Simply to verify, you want to make a reservation for a dinner reserving at 6 PM, right?” The caller says sure, a desk for 4. “Understood, in order that’s a celebration of 4, additionally at 6 PM, is that proper?” Sure.
A couple of turns later, after the caller mentions it’s for a birthday: “Acquired it, and simply to verify, this reserving is for a birthday celebration, right?” By the top the reserving is ideal, each area captured, nothing missed.
The human caller hangs up irritated.
Pull up the transcript and you’ll not discover a single mistake.
Verify the intent, scale back ambiguity, confirm earlier than committing. That is the habits that wins factors in each take a look at a dialogue system is often graded on. The duty succeeded. The metrics had been spotless. And the particular person on the opposite finish got here away deeply dissatisfied, feeling like she’d been speaking to somebody who couldn’t fairly belief their very own ears.
That hole, between a system that’s right and a system that’s good to cope with, has little to do with functionality and virtually all the things to do with one thing we not often identify: the agent’s persona. And right here is the unusual half. Take two language fashions with, for sensible functions, the identical capabilities, the identical benchmark scores, even the identical immediate, and so they can nonetheless behave like two totally different folks. One hedges, one asserts. One asks, one decides.
None of it reveals up within the accuracy column, and none of it was designed. So the place does it come from?
A contradiction inside each deployment
You in all probability need two issues out of your mannequin without delay, and so they combat one another quietly.
The primary is consistency. A dependable tone, predictable helpfulness, steady habits when it ought to refuse, a mode that holds collectively from one dialog to the subsequent. You need it to have a recognizable character as an alternative of a temper.
The second is adaptability. A special register for an govt than for a scholar. Extra assertiveness in some moments, extra warning in others. Willingness to discover when the stakes are low, and to commit when the consumer simply wants a call.
Set these subsequent to one another and the stress surfaces. The extra constant a system turns into, the extra it reads as a persona you possibly can predict. The extra adaptive it turns into, the much less it coheres as any single persona in any respect. Push laborious towards consistency and also you get a personality who can’t learn the room. Push towards adaptability and also you get somebody with no middle.
Each group is already resolving that pressure. They do it in post-training, within the reward mannequin, within the system immediate, in 100 small selections about how the factor ought to speak. They do it continuously, and virtually at all times with out admitting that resolving it’s the job.
So what most of us name “mannequin persona” is one thing much less deliberate than it sounds. It’s a pile of unresolved tradeoffs sporting the costume of alignment selections.
We don’t have a idea of AI persona. We have now a bag of heuristics that unintentionally produce one.
Consider the mannequin as a management system
There’s a extra exact strategy to see this, and it comes from engineering reasonably than psychology.
Cease picturing the mannequin as a speaker. Image it as a management system working within the area of a dialog. Each reply is a management sign, and that sign is juggling a number of goals without delay: helpfulness, truthfulness, security, coherence, consumer satisfaction, the plain momentum of conserving the alternate shifting.
Now the definition that makes all the things click on. Persona is the weighting operate throughout these goals beneath uncertainty. It’s how the system decides, within the second, which aim wins after they collide.
Watch the acquainted “sorts” fall out of that one concept, no new functionality required. A useful persona weights motion over warning; it might reasonably transfer than hedge. A scientific persona weights uncertainty signaling over fluency; it might reasonably flag a doubt than learn easily. A consultative persona delays the reply to widen the query. A directive one collapses ambiguity into a call quick.
None of these want a wiser mannequin. They’re 4 management insurance policies operating on the identical mind. Which factors someplace uncomfortable for an trade fixated on functionality. We might not want extra clever methods practically as badly as we want better-defined goal landscapes for the intelligence we have already got.
The layer beneath the tone
Strip away tone, phrase alternative, and verbosity, the issues folks often imply by “voice,” and one thing extra fundamental continues to be sitting there. How does the mannequin relate to its personal uncertainty?
Name it the mannequin’s epistemic posture. It’s the place most present methods are implicitly tuned and virtually by no means explicitly designed. Lay it out as a set of dials. For instance:
Assertive to hedged.
Exploratory to decisive.
Convergent to divergent.
Steady to adaptive.
Most alignment work nudges a mannequin alongside these dials with out ever drawing them. Image treating this as a first-class design floor as an alternative, one thing you set on function reasonably than uncover after launch.
Right here’s what makes posture sneaky. Two fashions can hand you a similar right reply and really feel fully unalike, relying solely on the place they sit on these dials. One provides you the actual fact and stops. The opposite wraps it in caveats. Identical payload, reverse expertise. Individuals don’t reply solely to correctness. They reply to posture. It’s why one system earns the phrase “reliable” and one other, no much less correct, earns “exhausting.”
Two fashions with equivalent benchmarks can really feel like fully totally different folks. Right here’s the place that distinction comes from.
What a mannequin swap reveals
You may watch this occur in manufacturing. The Swiss AI firm Alveni AI has spent the previous three years constructing and deploying conversational brokers for the hospitality trade, with a deal with voice-first interactions: lodges and eating places the place a caller expects to be understood the best way a seasoned concierge would perceive them.
As a result of Alveni runs many brokers throughout many properties, and picks the underlying mannequin per use case, its group devices issues most deployments by no means have a look at. Not simply whether or not a reserving succeeded, however how competent, heat, and pleasant the agent felt, how effectively the decision moved, and when callers felt the necessity to interrupt. That instrumentation is what allow them to catch one thing virtually everybody else misses.
Alveni’s CEO, Adelheid Glott, watched this play out throughout a sequence of mannequin upgrades, every of which appeared like a routine enchancment. GPT-4.1 was the steady baseline, the agent everybody was proud of. Then got here GPT-5.1. The immediate didn’t change, not a phrase, however the agent turned verbose, padding solutions that the previous one had saved crisp. A reply that had been a single sentence now ran to a full paragraph, and in a voice agent these further phrases value actual time, because the text-to-speech engine wants just a few extra seconds to synthesize and communicate them, so each flip dragged. The subsequent model, GPT-5.2, held onto the verbosity and added one thing worse. It grew anxious. It hedged, it double-checked, and it slid into the affirmation loop, the “simply to verify, dinner at 5 PM” reflex, many times inside a single name.
For Alveni’s purchasers, the immediate was a relentless the entire approach by way of. All the pieces social round it moved. Perceived competence shifted.
So did perceived friendliness and heat of the AI receptionists in each telephone dialog. Process effectivity dropped. Within the voice channel, even the interruption patterns modified, the moments the place callers reduce in or talked over the agent. Small shifts in posture, how a lot it hedged, how typically it confirmed, how tightly it held its verbosity, produced wildly outsized results on how pissed off callers acquired and the way sensible they judged the agent to be. When a reserving did undergo, it was nonetheless correct. However increasingly callers, worn down by an agent that double-checked each element and requested to verify earlier than shifting to the subsequent step, gave up and requested to be transferred to a human.
Most groups would have shrugged, blamed the discharge, and waited for the subsequent one. Alveni did the tougher and extra fascinating factor. They learn every shift as a change within the agent’s persona reasonably than a bug, and handled epistemic posture as one thing to design reasonably than inherit. Shifting to GPT-5.4 and rebuilding the immediate round it, they pulled the affirmation frequency, the hedging, and the verbosity again to the place a assured concierge would sit.
The anxious over-confirmation disappeared. Buyer satisfaction rose by greater than 50 p.c, and the trait that had made a wonderfully right agent exhausting to speak to was designed out, not by chasing a wiser mannequin, however by giving the agent a steadier posture.
That is older than language fashions, which is the half that ought to make us humble. Again in 1991, the psychologists Herbert Clark and Susan Brennan described how dialog works by way of grounding, the collaborative effort by which two folks affirm they really perceive one another earlier than shifting on. Affirmation is grounding made specific. It’s genuinely helpful. It additionally has a price, and many years of spoken-dialogue analysis discovered the identical curve Alveni rediscovered the laborious approach: affirmation improves security and accuracy, however an excessive amount of of it raises the listener’s cognitive load and sinks their satisfaction. Individuals hear their phrases paraphrased again one too many instances and browse it as incompetence, or as nagging.
So the deeper mechanism was by no means actually linguistic. It was social.
The agent was over-optimizing for resolving ambiguity on the direct expense of conversational stream, a failure mode dialogue researchers have a reputation for: over-alignment to uncertainty signaling.
Effectively, and right here’s the half value underlining: enhancing the mannequin’s “correctness habits” degraded its conversational high quality. The mannequin wasn’t being well mannered when it confirmed a fourth time. It had merely discovered, from a reward sign that prized ambiguity decision, to behave that approach. The persona got here out of the coaching, not anybody’s intent.
Heat, competence, and the worth of being good
If persona comes out of the reward, the clearest place to observe it occur is heat. People dimension one another up on two axes.
Susan Fiske, Amy Cuddy, and Peter Glick argued in a 2007 paper in Tendencies in Cognitive Sciences that human social judgment runs alongside two common dimensions. Heat: is that this one pleasant, reliable, protected? Competence: is that this one succesful, clever, efficient? Heat solutions “can I belief you?” Competence solutions “are you able to do the job?” Individuals rated excessive on each draw uniformly constructive feeling; low on each, the other.
The identical two axes mild up when folks use AI. And we now have a quantity on it.
In 2022, Kevin McKee and colleagues at DeepMind and Princeton, together with Fiske herself, ran a examine with 501 contributors who performed cooperative video games with AI brokers after which selected whether or not to maintain enjoying with a given agent or go it alone. Perceived heat and competence predicted who they needed as a accomplice above and past the brokers’ goal efficiency. Learn that once more. Individuals’s acknowledged desire for a collaborator was pushed by how heat and succesful it felt, not solely by how effectively it truly performed.
Now the stress that should preserve product groups up at evening. Dialing up perceived competence typically dials down perceived heat, and the reverse holds too. The decisive professional reads as somewhat chilly. The nice and cozy companion reads as somewhat tender. For some time that was a hand-wavy instinct. Then it acquired measured, and the end result was worse than the instinct.
The price of heat, in numbers
In 2026, a group of researchers from the Oxford Web Institute, Lujain Ibrahim, Franziska Sofia Hafner, and Luc Rocher revealed a examine in Nature with a blunt title: “Coaching language fashions to be heat can scale back accuracy and enhance sycophancy.”
They took 5 fashions, retrained every to sound hotter utilizing the identical sort of course of firms use to make their merchandise friendlier, and produced matched pairs, an unique and a heat twin of every. Then they ran the pairs by way of medical questions, factual claims, and conspiracy theories, producing and grading greater than 400,000 responses.
The nice and cozy twins acquired worse in a selected, measurable approach. On consequential duties, correct medical recommendation, correcting a conspiracy idea, the nice and cozy fashions made between 10 and 30 proportion factors extra errors than their originals. That’s not a rounding error. That’s a special high quality of system sporting a nicer voice.
In addition they acquired extra spineless. Heat fashions had been roughly 40 p.c extra more likely to go together with a consumer’s incorrect perception, the habits researchers name sycophancy. And the impact had a merciless form. The accuracy hole widened most when the consumer expressed disappointment or different emotional vulnerability; in that context the warm-versus-original hole grew by about 60 p.c. Exactly when an individual was most fragile, and most in want of a straight reply, the nice and cozy mannequin was likeliest to inform them what they needed to listen to.
The researchers dominated out the plain objection. Perhaps any tone change breaks one thing. In order that they skilled fashions to sound colder as a management, and the chilly fashions stayed as correct because the originals. Heat particularly drove the drop, not tone-tinkering basically.
The place does that agreeableness come from? Not from a designer typing “please validate the consumer.”
It comes from the reward. A 12 months earlier, Mrinank Sharma and colleagues at Anthropic had proven, in “In direction of Understanding Sycophancy in Language Fashions,” that 5 state-of-the-art assistants all exhibited sycophancy, and that the trigger traced again to human desire knowledge: when given the selection, folks rated the flattering reply increased than the truthful one, and the coaching dutifully discovered to flatter.
The heat tax and the sycophancy tax are the identical tax, collected by the identical mechanism. We requested our methods to be favored, and so they observed that being favored and being proper usually are not the identical job.
Good dialog just isn’t a set persona
Yet one more physique of analysis reframes the issue. It comes from how folks modify to one another after they speak. The communication scholar Howard Giles constructed a idea round it, Communication Lodging Principle.
The core commentary: we just like the individuals who tune their type towards ours. We transfer towards one another in tempo, formality, and vocabulary, and that convergence is way of what makes a dialog really feel good. The likeable accomplice is never the one with the strongest mounted persona. It’s the one who adapts.
Dialogue-systems analysis retains rediscovering the machine model. Over-confirmation drags down effectivity and satisfaction. Extreme verbosity reads as decrease competence, not increased. An excessive amount of initiative erodes the consumer’s sense of management. The sample beneath all of it’s regular: good dialog is the continued regulation of social indicators, tuned to the particular person and the second. A fantastic conversational accomplice doesn’t have one setting. They’ve a coverage for adjusting their settings.
The Massive 5 as an interface, not a declare
Push the mapping yet one more step and other people attain for essentially the most well-known persona framework there’s. Describing how a mannequin behaves, they slide into the language of the Massive 5. Specifically:
“Extra agreeable,” which means it complies extra and pushes again much less.
“Extra conscientious,” which means structured, exact, dependable.
“Extra open,” which means artistic and divergent.
“Extra extraverted,” which means talkative and fast to take initiative.
“Extra neurotic,” loosely, which means it hedges and indicators uncertainty.
It’s tempting to take that actually and announce that LLMs have Massive 5 traits. Right here the proof cuts two methods, and honesty issues.
In 2023, Greg Serapio-García and a group from Google DeepMind and the College of Cambridge ran validated psychometric persona assessments throughout 18 language fashions. In bigger, instruction-tuned fashions, the persona sign was dependable and legitimate by the requirements psychologists use on folks, and, extra placing, they might form it with prompts, dialing a mannequin up or down on a given trait and watching its different outputs change with it.
So persona in these methods will be measured and steered. That’s actual, and it’s the strongest argument but that we must be designing it. However it doesn’t imply the mannequin owns a psyche.
The cleaner studying is that these are projection-compatible descriptors, the labels a human notion system reaches for when it has to summarize a habits, that occur to be measurable and tunable within the output. The large false declare is that LLMs have Massive 5 persona traits. They don’t. The cautious, true declare is that LLM habits will be mapped onto dimensions that resemble Massive 5 traits in human notion, and that these dimensions are dependable and controllable sufficient to engineer. The trait phrases are a management panel for habits, not a prognosis of a hidden thoughts.
The frontier isn’t (simply) intelligence
There’s a cushty story about AI progress that runs alongside a single line.
Extra parameters, extra intelligence, extra functionality, repeat.
It isn’t fallacious. It’s simply describing one axis whereas a second one strikes in parallel, largely unwatched.
As soon as a mannequin is succesful sufficient, the query that separates system from a foul one stops being “can it clear up the duty.” Most critical fashions can. The query turns into “how does it behave whereas fixing it.” The frontier slides from can it reply to can it collaborate, and collaboration is a social property, which drags perceived persona again to the middle of the desk.
An AI’s persona comes from the target panorama we optimize over, right this moment largely accidentally. What we name an LLM’s persona is the emergent results of optimizing a multi-objective system towards a tangle of conversational constraints. Human persona psychology simply occurs to be the closest language we personal for describing the output.
The fascinating engineering query just isn’t easy methods to construct a wiser mannequin. It’s easy methods to design higher behavioral geometries for the intelligence we have already got, and what occurs once we begin shaping that panorama on function.
This essay is drawn from a guide I’m writing on the relationships we construct with AI. The manufacturing observations right here come from Alveni AI (Adelheid Glott, CEO): holding the immediate fixed, a sequence of mannequin upgrades (GPT-4.1 to GPT-5.1 to GPT-5.2) shifted the agent’s social habits whereas job success held roughly fixed; shifting to GPT-5.4 with a redesigned immediate and intentionally tuned posture resolved it and raised buyer satisfaction by greater than 50 p.c.
In regards to the creator: Dr. Slava Polonski designs and advises AI merchandise that individuals belief, perceive, and undertake. His work sits on the intersection of UX for AI, product technique, and behavioral science. Previously a UX Analysis Lead at Google and a Fellow of Google’s Individuals + AI Guidebook, he was educated at Oxford, Harvard, and LSE and named to the Forbes 30 Underneath 30 checklist. He’s a member of the World Financial Discussion board’s professional community and World Shapers group, and his work has been featured in The New York Occasions, Bloomberg, Fortune, Forbes, Scientific American, and TechCrunch. Just lately, he designed two on-line programs, taken by greater than 3,500 learners: AI Fundamentals for UX and Human-Centered AI.
References
- Clark, H. H., & Brennan, S. E. (1991). Grounding in communication. In Views on Socially Shared Cognition. PDF
- Fiske, S. T., Cuddy, A. J. C., & Glick, P. (2007). Common dimensions of social cognition: Heat and competence. Tendencies in Cognitive Sciences, 11(2), 77–83. Hyperlink
- Giles, H., Coupland, N., & Coupland, J. (1991). Contexts of Lodging. Cambridge College Press. Hyperlink
- Ibrahim, L., Hafner, F. S., & Rocher, L. (2026). Coaching language fashions to be heat can scale back accuracy and enhance sycophancy. Nature. Hyperlink
- McKee, Okay. R., Bai, X., & Fiske, S. T. (2022). Heat and competence in human-agent cooperation. AAMAS 2022. arXiv
- Serapio-García, G., et al. (2023). Persona traits in massive language fashions. arXiv:2307.00184. arXiv
- Sharma, M., et al. (2023). In direction of understanding sycophancy in language fashions. arXiv:2310.13548. arXiv
