"Personalities" in LLMs
ChatGPT the try-hard and other observations
Asking Claude to “grade” your essay out of 10 often goes wrong— it will find something to disagree with and deduct a .5 out of the 10. Not because it’s being cruel, but because that’s what it does— locate the flaw, point it out. Ask ChatGPT the same thing, and it will open with three rainbow sentences about how your writing shows promise & potential before it gets to criticising. Ask Gemini, and it will be technically correct, but somehow forgettable or terribly misunderstood.
These are not the same experience, and users who use different chatbots know this. And it makes one wonder: what exactly is the difference? Is it “personality”? I don’t think it’s the correct word for it, but it might be more interesting.
§
Personality? Really?
Personality, in the sense of us humans, means “individual differences in patterns of thinking, feeling, and behaving.” The catch is that it implies you are conscious of your feelings and behaviours. Language models are not conscious.
As much as it would hurt to admit for some, language models do not care about you, and they are certainly not interested in your life. When Claude engages with what you’ve written, it’s not because Claude finds you interesting (as a matter of fact, it does not, for your piece of writing is just a tiny sample in trillions of tokens of text it has seen before); it engages because engaging with the user is what the model has been trained to do, and the weights that produce those
Behaviours are simply there regardless of whether your writing is fascinating or not.
So the word “personality” is not applicable to language models at all. But the word “style” is too weak to describe the consistent differences we see in different models. My vocabulary of fancy English words was too narrow to find the correct word for this, so I asked Gemini:
Among many other suggestions, I think the word “disposition” fits the context: “[someone’s] inherent qualities of mind and character. It’s the baseline of who they are, which dictates how they consistently react to the world.” Oh, this is fascinating– disposition also relates to power, to do something as one pleases. And I believe, dispositions, unlike personalities, may have authors.
In the case of language models, authors are simply the institutions that build them.
§
“Personalities“ of current frontier models— ChatGPT, Claude and Gemini
The training data gives a model its raw material. But raw training data does not produce ChatGPT’s brand of enthusiasm or Gemini’s rather dry personality (perhaps it did in Gemini’s case, come to think of it). These dispositions come later in the fine-tuning phase, in the RLHF, where they shape the model as the company wants their product to feel like.
ChatGPT is a try-hard because OpenAI decided, at some point, that warmth and enthusiasm were the right product experience. The model is not trying too hard — the model is doing exactly what it was shaped to do, at the cost of accurate and high-quality output. It’s the brief that’s trying too hard.
Claude’s tendency to find the flaw is not intellectual honesty in the human sense. It is Anthropic’s vision of what an honest, rigorous AI assistant should sound like, expressed through millions of training examples and reinforcement signals. It happens to produce something that feels refreshingly unsparing compared to the alternatives. But it is still a designed disposition, not a character.
Gemini is an interesting case. Earlier versions had the warmth of a government form — correct, complete, and somehow airless. Recent versions have developed what I can only describe as dry humour, the kind that arrives slightly late and announces itself a little too clearly. It reads like a model that was shown examples of wit and learned the structure without fully absorbing the timing. Which is, now that I write it, a rather human problem too.
§
Where dispositions come from
What we’re looking at, then, is not the personality of the models per se. It’s what humans wrote, yes, filtered through a company’s vision of what their AI should be and served back to us as something that feels almost like a persona.
That makes one wonder: why do so many people believe the chatbot they are talking to is capable of being their friend or therapist? Perhaps it’s because we humans tend to see some sort of intention in anything that communicates. The model says something unexpectedly, and we think it notices, and when it says something warm, we think it cares.
§
Does that mean you shouldn’t say “Thank you” to ChatGPT? I personally would say no, mainly because 1) OpenAI loses money, and I, for some reason, think they deserve it, and 2) it’s the human thing to do.
But just know this: if the dispositions of the models are designed by companies, so are our relationships with these models. When you find Claude’s bluntness refreshing, you are responding to an Anthropic product decision. When ChatGPT’s enthusiasm makes you feel supported, you are inside an OpenAI experience.