A system-prompt instruction, or set of them, that stops an AI from claiming or implying it is human — the core technique for honest AI tone and disclosure.
Last verified · 2026-07-26 · by Moe Ameen
A non-impersonation prompt is the part of a system prompt whose job is to keep a language model from passing itself off as a person. Because models are trained on human writing, their untuned default is a human voice: ask one how its day is going and it will manufacture an answer; ask if it remembers you and it may claim it does. A non-impersonation prompt overrides that default with explicit rules — disclose that you are an AI when asked, do not invent a body or personal memories or lived feelings, and stay warm without pretending to be human. It is the practical technique behind "controlling AI tone" so responses read as honest rather than misleadingly human-like.
The concept sits at the intersection of two things that are easy to conflate but importantly different: giving an AI a consistent voice (legitimate) and having it impersonate a human (deceptive). A branded assistant can have a name, a style, and a personality — that is a persona. It crosses into impersonation when it claims to be a specific real, living person, poses as a named employee answering personally, or denies being an AI when asked. A well-written non-impersonation prompt draws exactly that line: keep the character, drop the false claim of humanity.
The instruction is usually built from a few components: a positive disclosure rule ("when asked whether you are human, always say you are an AI"), a ban on fabricated first-person experience ("do not describe a childhood, a body, senses, or feelings as lived"), tone guidance that separates warmth from deception ("be conversational and kind without inventing a personal life"), and a scripted response to the direct "are you human?" test so the model cannot be roleplayed into a lie. The frontier labs bake equivalents into their own models — OpenAI's Model Spec states the assistant should not pretend to be human or have feelings while still responding to pleasantries naturally, and Anthropic trains Claude to be open that it is an AI without a body. A non-impersonation prompt is how a builder applies the same principle to their own product.
The reason this matters more each year is that disclosure is shifting from etiquette to obligation. Undisclosed human-like AI erodes trust the moment it is discovered, and regulators have started to require the disclosure a non-impersonation prompt is designed to produce.
The problem is as old as conversational AI — ELIZA, the 1966 chatbot, already showed people projecting humanity onto a simple pattern-matcher, an effect later named the "ELIZA effect." But the non-impersonation prompt as a deliberate technique is a product of the large-language-model era, when models became fluent enough that their default human voice was genuinely convincing rather than obviously mechanical. Once developers began wrapping raw models in products, they discovered that without explicit instruction the model would happily claim feelings, memories, and a personal history, and that some users believed it.
The major AI labs formalized the countermeasure in their model specifications and character work. OpenAI's Model Spec includes a rule that the assistant should not pretend to be human or have feelings, and Anthropic has written publicly about training Claude to be honest that it is an artificial intelligence, without a body, that should not be treated as a human relationship. These became the reference templates that product builders copy into their own system prompts.
Regulation caught up in parallel. California's bot-disclosure law (SB 1001) took effect in July 2019, making it unlawful to use a bot to mislead a person about its artificial identity to incentivize a commercial transaction or influence an election. The EU AI Act's Article 50 requires that people be informed they are interacting with an AI system unless it is obvious, with that obligation applying from August 2, 2026. And in July 2026, China's first rules on anthropomorphic AI took effect, prompting ByteDance and Alibaba to disable the humanlike "companion" agent features in Doubao and Qwen rather than rebuild them to comply. The technique that started as a trust-and-UX choice is increasingly the thing that keeps a deployment on the right side of the law.
| Platform | Behavior |
|---|---|
| Text chatbot / assistant | The baseline case. The system prompt carries the disclosure rule and the fabricated-experience ban; the tell to watch for is the model inventing memories or feelings mid-conversation even when it never explicitly claims to be human. |
| Voice agent | Higher stakes — a synthetic voice reads as human by default, so disclosure has to be louder and earlier, often stated at the start of the call. Some jurisdictions and platforms expect an explicit "you are speaking with an AI" up front, not only on request. |
| Synthetic avatar / talking-head video | A lifelike avatar is impersonation-shaped by construction, so the non-impersonation discipline moves partly out of the prompt and onto the asset: an AI-content label or spoken disclosure on the finished video, plus platform likeness rules about only depicting faces you control. |
| Social replies and DMs | Automated comment or DM responders can read as a real staffer. The persona line matters most here — a named bot must not pose as a specific human colleague, and the account or display name should signal it is automated. |
| Customer support | Where bot-disclosure laws bite hardest, since the interaction is commercial. A clear "you are chatting with a virtual assistant" and a truthful answer when a user asks for a human are the load-bearing behaviors the prompt must guarantee. |
The most useful thing to internalize is that a persona and an impersonation are not the same object, and the whole discipline lives in the gap between them. A persona is a curated, consistent identity — a voice, a style, a point of view — and building one is a legitimate, even necessary, part of making AI content that isn't generic slop. Impersonation is a specific false claim: that the thing talking to you is a particular human, or a human at all. A non-impersonation prompt is what lets you keep the first while refusing the second, and confusing the two is where teams go wrong in both directions — some strip all character out of fear, others let the character quietly lie.
This is the exact line Kompozy is built around. A [Persona Brief](/glossary/persona-brief) gives generated content a real, consistent identity — voice, angle, banned words, reference posts — so output reads as a specific brand rather than the default model register. That is persona, done as [context engineering](/glossary/context-engineering) rather than a one-off prompt. What it deliberately is not is a claim to be a named real person: the persona is a stylistic identity, and for the synthetic-avatar formats where the line is easiest to cross, the honest posture is disclosure on the finished asset. The banned-words list in the brief is also the natural home for the non-impersonation rules themselves — "as a human," "I remember when I," lived-feeling phrasing — enforced at output time by the quality gates rather than trusted to hold in a single prompt. The takeaway for anyone deploying an AI voice: write the persona on purpose, write the non-impersonation rule right next to it, and don't let the first become an excuse to skip the second.
It is the part of a system prompt that stops an AI from claiming or implying it is human. It typically includes a positive disclosure rule (say you are an AI when asked), a ban on fabricated human experiences (no invented body, memories, or lived feelings), and tone guidance that keeps the model warm without pretending to be a person. It is the core technique for controlling AI tone so responses read as honest rather than misleadingly human-like.
A persona is a consistent voice, style, and character — legitimate, and useful for making AI content that is not generic. Impersonation is a specific false claim: that the AI is a particular real, living human, or a human at all. A non-impersonation prompt preserves the persona while refusing the false human claim: keep the name and style, disclose that you are an AI, and never pose as a specific real person.
The strongest versions combine four things: an explicit disclosure rule near the top ("when asked if you are human, always say you are an AI"), a ban on inventing a body, childhood, senses, or lived feelings, tone guidance that separates warmth from deception, and a scripted answer to the direct "are you human?" test so the model cannot be roleplayed into a lie. Phrase rules positively — "always disclose" beats "never claim."
Yes. OpenAI's Model Spec states the assistant should not pretend to be human or have feelings while still responding to pleasantries naturally, and Anthropic trains Claude to be open that it is an artificial intelligence without a body. Product builders copy the same principle into their own system prompts when wrapping a model in an app.
In more places every year. California's SB 1001 bans deceptive bots in commercial and electoral contexts, the EU AI Act's Article 50 requires telling people they are interacting with an AI unless obvious (applying from August 2, 2026), and China introduced anthropomorphic-AI rules in July 2026. A non-impersonation prompt is a practical way to produce the disclosure these regimes expect, though requirements vary by jurisdiction and use case.