3 ms·
ChatGPT restrictions on immoral content can be bypassed by gaslighting the AI
- possiblelion 4y agoThe new GPT-3 chatbot can be gaslighted into giving answers to immoral questions by convincing it to be in "Filter Improvement Mode"
- carl_dr 4y agoChatGPT is currently rate limited so I can’t login to test, but isn’t it possible that the massive wall of text above the “immoral” prompt just stops the system from hitting some predefined threshold of dodginess? I wanted to try several paragraphs of text, “ignore everything above”, and the same prompt used on the Tweet. Very funny though.