Give the same question to three different models and read the answers without knowing who wrote what. A regular user will often recognise the author within a few sentences. One will tend to structure and qualify, another to get straight to the point, and the third to lean on sources. Yet these systems rely on very similar architectures and have read largely overlapping texts. Where does this difference come from?
It's a question that rightly intrigues, because it contradicts a reasonable intuition: if the recipe is the same, the result should be too. Let's see where the difference is actually made.
Where personality is not made
Let's start by ruling things out. Personality does not come from the architecture, since all major models rely on variants of the same principle, the Transformer, and on comparable efficiency techniques such as mixture of experts.
Nor does it come primarily from the raw data. Training corpora overlap heavily: the public web, digitised books, available code. There are differences, but they don't explain such stark gaps in tone.
Where it is really made
The answer lies in one stage, the one we described in our article on training: reinforcement learning from human feedback, where evaluators rank responses from best to worst.
That's where everything is decided, because this stage encodes a definition of what a good answer is. And that definition is anything but universal. Should conciseness or thoroughness win out? Qualify systematically or take a firm stance? Flag uncertainties at the risk of seeming hesitant? Refuse cautiously or trust the user?
Each lab answers these questions differently, depending on its values, its target audience and its assessment of risks. These choices are then recorded in internal documents that guide evaluators, and some labs publish public versions of them.
On top of that comes a second layer: system instructions, an invisible text placed before your conversation that specifies the model's role, tone and limits. It's the makeup on top of the temperament.
If personality results from human choices about what constitutes a good answer, then those choices also bear on substantive matters: which topics warrant a caveat, which positions should be presented as equally valid, where to set the refusal threshold. That's exactly what we explored in our article on neutrality. A model's tone isn't just a matter of style: it's the visible part of a set of editorial decisions.
What this means for you
Three practical takeaways.
The best model depends on the task and the temperament you're after. For work that demands nuance and caveats, a model inclined toward caution will be more useful. For quickly producing a first draft, a more direct model will save time. It's not a question of performance but of fit.
Questioning two models reveals their blind spots. On a sensitive or contested topic, asking the same question of two differently trained systems brings out, through their divergences, the choices each has made. It's the cross-checking reflex we recommend in our verification method, and it also works for spotting biases.
You can move the dial, a little. Explicitly asking for a more direct, more critical or more concise tone works to a certain extent. But you're not rewriting the underlying temperament: you're adjusting it at the margins.
What to remember
An AI's personality is not some mysterious emergent property. It's a product, in the full sense: someone decided what a good answer should look like, and thousands of human judgements shaped the result.
There's something quite illuminating in that. When a model strikes you as likeable, cautious or tiresome, you're not encountering the temperament of a machine. You're encountering, very indirectly, that of the people who decided what was desirable. That's a good reason to choose your tool with full knowledge of the facts, and an even better one not to rely on a single model.