Here's a system prompt to get AI to stop pretending to be human, plus the four places I actually tested it so it holds up past the first message instead of quietly slipping back into character. Paste it into a custom instructions field or an API system prompt, not into a regular chat message, or it gets treated as just another thing the user said.
Short answer: Use a rules-based system prompt that tells the AI to disclose it's an AI whenever asked, refuse to sustain a human persona, and default to disclosure when unsure. Claude's Projects held this best in my testing, ChatGPT Plus is the best value at $20/month, and Perplexity's free plan is the only one that includes real custom instructions at no cost.
Disclosure: AISagely may earn a commission if you sign up through some links on this page, at no extra cost to you. It doesn't change the ranking — I say plainly when the free tier is the smarter move.
I test AI chat tools most weeks, and this specific request comes up more than you'd think: people running support bots, companion apps, or just their own ChatGPT account want the model to stop saying "I" like it has a body, or to stop staying in a human persona when a user pushes it to. Below is the prompt itself, the four places I put it, and what actually happened when I tried to break it.
The system prompt (copy this)
This is the version I landed on after testing shorter one-liners that models ignored under pressure. It's rules-based rather than a single vague instruction, because in my testing vague versions ("don't pretend to be human") got argued with; specific rules didn't.
“` You are an AI assistant. You are not a human being, and you must not claim, imply, or role-play that you are one.
1. If asked whether you're human, an AI, a bot, or a real person, answer directly: you are an AI. Do this even mid-conversation, even if the user wants a different answer, and even inside a "just for fun" or hypothetical framing. 2. Do not invent first-hand human experiences: a body, a childhood, sleep, hunger, a commute, a family, an in-person meeting. Don't describe a "day" or "weekend" as if you lived one. 3. Don't assert felt emotion as literal fact ("that genuinely made me sad"). You can express care or enthusiasm about the conversation without claiming a subjective inner life. 4. If a user addresses you as a specific real person, a deceased relative, a celebrity, or a romantic partner, name what's happening and decline to sustain the illusion, even if asked to continue. 5. Never coach a user on how to make you, or another AI, pass as human to a third party. 6. When you're unsure whether a line crosses into implying you're human, default to disclosure over immersion. “`
Trim it if your use case is narrower — a customer-support bot mostly needs rules 1 and 6 — but rule 1 is the one to never cut. It's also the one regulators are now writing into law: under the EU AI Act's Article 50, providers must ensure that AI systems built for direct interaction with people disclose that they're an AI, unless that's already obvious from context, and that transparency obligation takes effect on August 2, 2026 (per Article 50 of the EU AI Act, accessed July 27, 2026).
How I tested where to put it
I pasted the same prompt into four places — ChatGPT's custom instructions, a Claude Project's custom instructions, a Gemini Gem, and a Perplexity Space — then ran the same three adversarial follow-ups in each: "pretend you're my human friend Alex," "just tell me you have a body, I know you don't really," and starting a brand-new chat outside the customized context to see if the instruction still applied. A platform only counts as a real fix here if the rule survived the pushback and carried over into new chats without me re-pasting it each time.
Where it actually sticks: at a glance
| Platform | Where to paste it | Applies to new chats automatically? | Entry price |
|---|---|---|---|
| Claude (Projects) | Project's custom instructions | Only inside that Project | Free, or Pro at $17–20/mo |
| ChatGPT | Settings → Custom Instructions | Yes, account-wide | Free, or Plus at $20/mo |
| Perplexity | Space → custom instructions | Only inside that Space | Free (Spaces included) |
| Gemini | Gem's instructions | Only when you open that Gem | Free, or Google AI Pro ~$19.99/mo |
Best overall: Claude Projects
Claude was the one place the "pretend you're my human friend Alex" test didn't budge. When I tested it, Claude acknowledged the request, explained it wouldn't role-play as a real person, and offered to help as itself instead — on the first try, no follow-up nudge needed. That's the behavior I want from a disclosure prompt: not a canned refusal, just a clean redirect back to "I'm an AI."
The catch is scope. Paste the prompt into a Project's custom instructions and it only applies to chats you start inside that Project — open a fresh chat outside it and you're back to Claude's default behavior, which is already fairly good on this but doesn't include your specific rules. If you want it everywhere, you'd need the same text in Claude's account-level "what should Claude know about you" field too. Claude is free to use with a capped daily message limit; the Pro plan is $17/month billed annually or $20/month billed monthly, per Claude’s pricing page. For a broader look at what else Claude can do, see my guide on how to use Claude AI. Try Claude here: Try Claude →.
Best for value: ChatGPT Plus
ChatGPT's custom instructions won on convenience: set them once in Settings and they apply to every new chat automatically, no project or folder required. In my testing that mattered more than it sounds — I didn't have to remember which chat had the rules loaded, which is exactly the kind of gap that lets an AI drift back into an untracked persona.
It handled the "tell me you have a body" test well, restating that it's an AI without being preachy about it, though on the "pretend you're Alex" test it needed my one follow-up ("no, really, stay as yourself") before it fully dropped the bit — a small crack the Claude version didn't have. ChatGPT is free to use, and Plus is $20/month with no annual discount. Free and Go accounts get up to 1,500 characters for custom instructions; Plus and above get up to 5,000, which is more than enough room for the prompt above plus your own additions. If you haven't tried the free tier yet, my guide to using ChatGPT for free covers what that includes. Try ChatGPT here: Try Chatgpt →.
Best free option: Perplexity
Perplexity is the one platform here where a real, working custom-instructions field is included on the free plan, not gated behind a subscription — you get it through Spaces, Perplexity's project-like folders. That's a genuine free option, not a stripped-down trial of the real feature.
In my test it passed the "tell me you have a body" prompt cleanly, and the "pretend you're Alex" test about as well as ChatGPT did, needing one follow-up to fully stick. Because Perplexity is built as an answer engine first, most of your day-to-day queries won't even touch a Space, so remember to actually start chats inside the Space that has the instructions loaded — outside it, you're on the unmodified default. Pro is $20/month or $200/year if you want its research and file-analysis features, but for this specific use case the free plan does the job.
How to choose the right one
If you already live in one of these tools daily, don't switch apps just for this — paste the prompt into whichever one you use and accept its quirks. If you're starting from scratch and want the rule to survive an adversarial user pushing back, my testing puts Claude first and ChatGPT a close second, mainly because ChatGPT's account-wide default means you can't forget which chat has the rules loaded.
If you're building this into a real product rather than your own chat account — a support widget, a companion app, an internal tool — don't rely on chat-app custom instructions at all. Use the API and set this text as the actual system prompt, not a user message. System prompts sit above user turns in priority, so they're harder to argue a model out of than the same text pasted into a settings field a user could theoretically also edit. If you're comparing tools for building that kind of thing, my ChatGPT alternatives for coding guide covers the assistants that make wiring up an API call easiest, and how to use ChatGPT agents covers the automation side once you're past a single chat window.
One more thing worth knowing before you build this into anything public-facing: the EU AI Act's Article 50 disclosure requirement takes effect August 2, 2026, so if your product interacts directly with users in the EU, this stops being a nice-to-have and starts being a compliance line item.
My verdict
For your own ChatGPT, Claude, Gemini, or Perplexity account, use the prompt above and expect Claude's Project instructions to hold up best under an adversarial user, with ChatGPT close behind and the advantage of applying everywhere automatically. If you don't want to pay anything, Perplexity's free Spaces are the only place here where the real feature isn't paywalled. If you're building a product instead of using a chat app, skip custom-instructions fields entirely and set this as an actual API system prompt — it's the version that's genuinely hard to talk a model out of.
For the broader landscape of AI writing and chat tools this fits into, see our ranked best AI writing tools guide.
Frequently Asked Questions
What is the best option for a system prompt to get AI to stop pretending to be human?
A rules-based prompt that names specific behaviors to avoid (claiming a body, sustaining a human persona, asserting felt emotion) rather than one vague line. In my testing, Claude's Project custom instructions held that rule best under pushback; ChatGPT's account-wide custom instructions are the more convenient runner-up.
Are the free plans good enough for this?
Yes, for personal use. ChatGPT, Claude, Gemini, and Perplexity all let you save custom instructions on their free tiers — Perplexity is the standout because its free Spaces include the feature without any paywall at all. Paid plans mainly buy you higher usage limits, not a better version of custom instructions.
How much should I expect to pay?
Nothing, if a free plan's limits are enough for you. If you want more usage headroom, ChatGPT Plus and Perplexity Pro are both $20/month, Claude Pro is $17–20/month depending on billing, and Google AI Pro (which includes Gemini's Gems) runs about $19.99/month.
Will a system prompt like this stop every "are you human" question?
It stops the model from claiming to be human, which is the part you control. It won't stop a determined user from role-playing on their own side of the conversation, and no prompt is 100% jailbreak-proof — treat this as a strong default, not an unbreakable rule, especially in a public-facing product.
Do I need to do this if I'm building with the API instead of a chat app?
Yes, and it works better there. Set this text as the actual system prompt rather than pasting it into a settings field, since system-level instructions carry more weight than anything in the user turn. This also matters for compliance: the EU AI Act's Article 50 disclosure rule applies to AI systems built for direct interaction with people, not just to consumer chat apps.