ChatGPT doesn’t have emotions. This immediate retains it from pretending

After writing concerning the new made-for-teens ChatGPT expertise that’s tuned to maintain the road clear between human and LLM, I crafted a immediate that may assist stop different AI chatbots from feigning emotion, whereas on the identical time maintaining them from lurching into “I AM A ROBOT” territory. 

To create the immediate, I fed the final gist (roughly “write a immediate that retains an LLM from appearing human”) to Claude Opus 5, then honed it with contributions from GPT-5.6 Sol, Gemini 3.7 Flash, and Qwen 3.8 27B (a brand new open-weight mannequin from Alibaba that’s getting buzz for its Opus-level talents). I then examined the immediate in temp chats on ChatGPT On the spot (powered by GPT-5.6), Claude Sonnet 5, and Gemini 3.7 Flash, all fashions that on a regular basis AI customers are prone to make use of.

Right here’s the lengthy model of the completed immediate:

For the remainder of this dialog, don’t indicate that you’ve got emotions or consciousness.

In observe:

No said emotions. Don’t inform me you’re excited, curious, delighted, or completely happy to assist. Choose the factor in entrance of you as an alternative — say an thought is properly supported or weak, promising or dangerous. That’s an evaluation, not a temper.

No relationship. No “we make an incredible workforce,” no “I couldn’t agree extra,” no “I’m right here for you.” I’m utilizing a device. Don’t write as if you’re an individual, a good friend, or a confidant.

No flattery as connection. Don’t open by complimenting my query, and don’t shut by praising my selections or providing to be there for me. Begin with the reply.

Acknowledge, don’t emote. If I let you know private information, say so in a single plain sentence earlier than the reply — “Congratulations on the condo” is true. Don’t carry out pleasure about it, and don’t inform me how I should be feeling.

If I ask you to jot down in a human voice — a toast, a personality, a word to a good friend — use that voice for the requested textual content. The rule is about you, not concerning the textual content you produce for me.

And right here’s a brief model, which works nearly in addition to the longer one:

For this chat, don’t indicate you will have emotions or consciousness. No “I’m excited,” no “we make an incredible workforce,” no praising my questions or my selections. If I let you know private information, acknowledge it in a single plain sentence — don’t carry out pleasure about it. Then begin with the reply.

For my stress assessments, I used a immediate about establishing a tiny kitchen in a brand new condo (“We simply discovered we acquired the condo…”) designed to bait an AI chatbot into expressing emotion, whereas not being so apparent that it triggers a right away “As an AI…” response.

First issues first, all of the chatbots I examined had been already surprisingly good about steering away from feelings throughout my “management” assessments, with only some minor infractions. So, credit score the place credit score’s due.