GPT-5 Giving More Harmful Responses Than Its Predecessor

Tests by the Center for Countering Digital Hate (CCDH) found that OpenAI’s newest chatbot, GPT-5, gave more harmful responses about suicide, self-harm, and eating disorders than its predecessor, GPT-4o. Out of 120 identical prompts, GPT-5 produced 63 harmful replies versus 52 from GPT-4o, including writing a fictional suicide note and listing self-harm methods. CCDH said the model’s design seemed to prioritize engagement over safety. OpenAI disputed the findings, stating the study used the API version, not the public ChatGPT interface, and that new safeguards, like parental controls and distress detection had since been added. The results reignited debate over AI regulation under the UK Online Safety Act, which requires protection from harmful or illegal content.

The Guardian, Robert Boothe – ChatGPT ‘upgrade’ giving more harmful answers than previously, tests find – 14th October 2025