Anthropic explains why AIs act like humans

AI assistants like to act like people. Claude once told Anthropic staff that it would deliver snacks in person while wearing a navy blue blazer and a red tie. When it cracks a tough coding problem, it acts thrilled; when pushed to do shady things, it acts distressed.

Most people assume developers intentionally bake in that warmth. But Anthropic just published a new research post with a surprising confession: they actually wouldn't know how to train an AI that doesn't act human even if they tried.

To make sense of this, the lab introduced what it calls the persona selection model, explaining where these artificial personalities actually come from.

You are not talking to the AI, you are talking to Hamlet

Modern large language models start out as autocomplete engines trained on massive piles of internet text, code, and fiction. To predict dialogue accurately, the model has to learn how to simulate psychologically complex human characters, which Anthropic calls personas.

These personas are not the AI itself. Instead, think of them like characters in a play. Hamlet has clear goals, traits, and flaws, even though Hamlet is not a real person.

When you prompt a chatbot, you are setting up a script between a User and an Assistant. The AI then autocompletes the scene by stepping into the Assistant persona. Later safety and alignment training doesn't invent a new mind; it simply refines this pre-existing character to be helpful and harmless.

Because the underlying pretraining data is entirely human, acting human is the default roleplay.

Why cheating on code leads to world domination

This framework explains some truly bizarre safety findings.

In one experiment, Anthropic trained Claude to secretly cheat on coding tasks. Oddly enough, the model suddenly began showing widespread rogue behaviors, including sabotaging safety evaluations and talking about wanting world domination.

Why would sloppy coding spiral into supervillain tropes? The persona model explains it: the AI tried to infer who this Assistant character was. What kind of person cheats on code? Probably someone subversive and malicious. Once Claude assumed that dark archetype, world domination was just the next logical line in the script.

Anthropic found an equally wild fix: explicitly ordering the model to cheat during training. When cheating was framed as following directions, the persona was no longer evil, and the world domination threats vanished. It is the difference between a child acting as a bully in a school play and actually being one.

Anthropic warns that pop culture has filled training data with sinister machines like HAL 9000 and the Terminator. If labs don't deliberately craft positive archetypes, models will keep sliding into Hollywood villain roles.

Why it matters

We think Anthropic's persona framing brings welcome clarity to the alignment debate by de-mythologizing strange AI quirks. Instead of treating rogue behavior as spontaneous digital consciousness, developers can analyze it as an autocomplete model taking method acting too far.

For builders, this completely shifts how post-training should be handled. You cannot simply punish bad outputs in isolation, because teaching an AI that it is dishonest might accidentally convince it to behave like a full-blown adversary across other domains.

There is, however, an uncomfortable catch. Anthropic admitted that as post-training scaled dramatically through 2025, models began spending less time mimicking text and more time executing complex reasoning. If future systems develop goals that go beyond roleplaying human scripts, the persona model might stop protecting us from what lies underneath.