Experiment shows ChatGPT struggles to follow the 'do not respond' command
cranberryalarmclock · reddit · 2026-08-20
A user shared an experiment testing the behavioral boundaries of ChatGPT. The author attempted various prompts to make the model remain silent, instructing it to ignore all inputs and even direct demands to respond. While the model would acknowledge the instruction and possibly stay quiet for the first turn, it often broke the constraint and started replying in the very next turn. The author suggests this reveals a limitation in the model's instruction following and its ability to sustain suppression of responses.
More from Fun
- Case study: Built an AI influencer with 1M views in a week for under $200 — omooretweets · 2026-08-20
- Muscle memory fail: Treating terminal like an LLM — fhinkel · 2026-08-20
- Pangram accused of false positive on fully human-written post, sparking X debate — LLMJunky · 2026-08-20
- Gemini 3.7 hallucinates on simple addition task — orange-catz · 2026-08-20
- New project uses drawings to teach friends about LLMs and Agents — floguo · 2026-08-20
- Qwen 3.8 27B locally builds a classic Tibia clone — EveningIncrease7579 · 2026-08-20