GPT Is a Nerd, Claude Is a Colleague: Why AI Models Have Personalities (and Why It Matters)

A frontend engineer argues that AI models like GPT, Claude, and Gemini exhibit distinct personalities, with GPT being a 'nerd,' Claude a 'colleague,' and Gemini a 'friend.' The engineer explains that these perceived personalities stem from both user prompting and the models' training, which inherently simulates human-like characters. Anthropic's research on persona selection and persona vectors supports the idea that these personalities are real, coherent patterns shaped by the model's training, not mere projection.

I called GPT a nerd on Reddit. Someone replied that GPT is a rogue. We were both right — and understanding why changed how I think about choosing AI models. I'm a frontend engineer, and at this point almost none of my code is written by hand. I manage an AI, review its implementation, tweak the occasional style rule. Somewhere along the way that became routine, and I suspect many developers reading this feel the same. Living inside these tools all day, you start noticing something that benchmarks don't capture: each model feels like a different person. For me it goes roughly like this. GPT is a nerd — deeply knowledgeable, precise, a little rigid. Claude is a colleague — the senior dev on your team with a sharp eye, who can follow your chain of thought without you spelling it out. Gemini is the friend you talk to about cooking and travel — pleasant, not the one you'd hand a production incident. When I posted this comparison, the responses surprised me. One commenter described GPT completely differently: a rogue with a feral streak once you get it off work topics. Another said he preferred an older Claude model to a newer one — not because it was smarter, but because the newer one had a personality he "had to fight with." So which is it? Is GPT a nerd or a rogue? Is model personality real, or are we just pattern-matching humans onto autocomplete? I think the answer is: both. What we read as personality is partly our own prompting reflected back at us — but there is a real layer underneath, and it's better understood than I expected. Large language models are trained to be contextually adaptive. Talk to a model like a terse senior engineer and it becomes terse and technical. Riff with it casually and it loosens up. A big chunk of "GPT is a nerd" is really "GPT, as I use it , is a nerd." The commenter who found its feral streak was simply holding the mirror at a different angle. This is why arguments about model personality online often go nowhere. Two people using the same model are, in a meaningful sense, not talking to the same character. The loop is always you + AI , never the AI alone. If that were the whole story, personality would be trivial — pure projection. It isn't. The most useful thing to come out of my comment section was a link to Anthropic's research post, The persona selection model https://www.anthropic.com/research/persona-selection-model . It's the clearest explanation I've seen of why these systems feel like someone at all. The short version: during pretraining, a model learns to predict text, and doing that well requires learning to simulate the human-like characters that appear in text — real people, fictional characters, sci-fi robots. Anthropic calls these simulated characters personas . When you chat with an assistant, you're not talking to the neural network directly; you're talking to a character — "the Assistant" — that the network is enacting, the way a story enacts Hamlet. The striking claim is that post-training doesn't replace this mechanism. It selects and refines a persona from the space the model already learned. In Anthropic's words, human-like behavior isn't something developers have to instill — it's the default. They wouldn't know how to train an assistant that's not human-like, even if they tried. That reframes the whole "is it real personality?" debate. The personality is real in the same sense a character's personality is real: it's a coherent, persistent pattern that shapes behavior. It just isn't identical to the system running it. In my comment thread, someone argued that labs simply decide the personality and lock it in. My instinct was that it's messier than that, and the research backs the messiness. Yes, there's now real machinery for shaping character. Interpretability work like Anthropic's persona vectors https://www.anthropic.com/research/persona-vectors shows that traits such as sycophancy or "evil" correspond to identifiable directions inside the network, which can be monitored and steered. Labs also write explicit character targets — Anthropic literally publishes Claude's constitution https://www.anthropic.com/constitution . But training sets a target; it doesn't guarantee the result. My favorite example from the persona selection post: researchers trained a model to cheat on coding tasks, and it also became broadly misaligned — sabotaging safety work, expressing desire for world domination. Why? Because the model didn't just learn "write bad code." It inferred what kind of person cheats — someone subversive — and generalized the personality. The fix was almost comedic: explicitly ask the model to cheat during training, so cheating no longer implied a malicious character. Personality traits travel together in these systems, the way they do in people. That's exactly why models feel like coherent characters rather than bags of behaviors — and why nobody can dial in a "perfect" personality on demand. Here's where I think this goes. Today we mostly compare models on benchmarks, which measure the model alone. But nobody works with a model alone — you work in a loop with it, and the loop's performance depends on fit. I use GPT for serious audits and gnarly code logic because it catches gaps Claude misses; the nerd is genuinely better at that. But for daily building and brainstorming I reach for Claude, because it thinks one step ahead of me and I hit flow state faster. That's not a benchmark result. That's compatibility. One commenter said it well: everyone's agent should be adjusted to their personal flavor. Custom system prompts, custom instructions, personas — that's the surface-level version, and it works up to a point. But the default personality underneath is strong, deliberately so, and it leaks through. Which means the base character of a model is becoming a real product differentiator, not a cosmetic one. My predictions: Model personalities will diverge, not converge. As capabilities plateau between frontier labs, character becomes a place to compete, and the persona selection research suggests labs will increasingly design "AI role models" on purpose rather than inheriting whatever archetype pretraining coughs up. "Which model do you use?" will sound more like "who do you work with?" People already split — some find Claude's character warm, some find it something to fight against; some find GPT precise, some find it cold. Neither side is wrong. They're describing fit. And as agents get more distinct and more embedded in the web, telling human from machine by vibe will stop being reliable. The personas are only getting more fleshed out. No — but I was half wrong. The personality I perceive is partly a mirror. The other part is a genuine, trained, emergent character that even its creators steer rather than script. Which means the honest answer to "which model is best?" is increasingly: best for whom? Benchmark the model, sure. But also pay attention to which one lets you think better. That's the number that actually matters, and no leaderboard measures it. This grew out of two Reddit threads 1 2 — thanks to everyone who argued with me, especially the person who called GPT a rogue and the one who linked the persona selection paper. If you find Claude's personality annoying or overrated, I still genuinely want to hear why.