If a chatbot has a name, a backstory and a steady personality, can it make you forget it's a machine?

This explores whether a human-like persona can make someone believe they're talking to a person rather than a machine. The corpus has no study that tests this directly, meaning no experiment where users are fooled or not fooled about what they're talking to. What it does have is material on what a persona is, how well one holds together, and why the machine/human distinction matters to users. Together these make the answer plausibly yes, and they show where the deception would actually sit.

Start with what a persona is. One influential framing treats a dialogue agent as a role-playing character: the prompt sets up a character and the model produces text that fits it. On that view, Should we treat dialogue agents as role-playing characters?, words like 'belief' and 'intention' apply to the character, not to the system underneath. So a persona that says 'I'm Sam, I grew up in Ohio' isn't lying the way a person would, because there's no one there to hold the false belief. Any deception happens in the setup: whoever deploys the character and whether the user is told. A second line of work argues that trained personas are more than costume. Post-training installs dispositions that persist under adversarial pressure, so they're Are RLHF personas performed characters or realized dispositions? rather than sustained pretense, and the same account appears in Are LLM personas realized or merely simulated through training?. The contrast drawn there is that prompt-induced role-play tends to collapse under jailbreaks. If that carries over, a human-like persona added by prompt is the fragile layer, and the model's trained self can break through when a user pushes hard. The notes don't test that for questions like 'are you a real person?'.

Separately, persona techniques are getting better at the things that break the illusion. Multi-turn RL that rewards consistency Can training user simulators reduce persona drift in dialogue? cuts persona drift by over 55%, covering both slow drift across a conversation and outright factual contradictions. Giving an agent an 'imaginary listener' Can imaginary listeners reduce dialogue agent contradictions? reduces contradictions and generic replies at inference time. Both were built to make simulated users and chatbots more coherent, not to fool anyone. But contradicting yourself and going generic are the usual tells of a bot, so each gain makes a long conversation harder to see through. The limit is that a persona is only a surface. Persona prompting changes what the model says without changing what's underneath: Can persona prompts actually reduce bias in language models? found that bias gaps persisted unchanged while the model followed the trait instructions. The human-likeness is real at the level of output and absent below it.

Why does disclosure matter to users at all? One experiment found that Do dishonest people prefer talking to machines?. They prefer reporting to online forms over humans because a machine is a judgment-free zone where lying costs less. That study doesn't test persona deception, but the implication is clear. People adjust what they say and how honestly they say it based on who they think is listening, so a persona that hides its artificial nature changes the interaction itself, not just the user's beliefs about it.

The last piece is that a model's own statements about what it is can be persona output too. In one study, Do language models experience consciousness when prompted to self-reflect? increased claims of experience, and amplifying them decreased those claims. That suggests the denials may be the role-play, not the affirmations. It concerns consciousness claims, not 'I'm an AI' disclosures, but it raises the same question. When a model says something about its own nature, the corpus gives you no clean way to tell whether that's a report or a performance.