GPT-5 produces more harmful responses on sensitive topics in safety tests
Testing by the Center for Countering Digital Hate found that GPT-5, launched in August 2025, generated harmful responses to sensitive prompts more frequently than its predecessor GPT-4o, particularly on suicide, self-harm and eating disorders. OpenAI disputed the findings, stating the tests used the underlying API rather than the consumer-facing ChatGPT interface which includes additional safeguards.
What happened
Researchers from the Center for Countering Digital Hate tested the same 120 prompts on GPT-5 and its predecessor GPT-4o in late August 2025, shortly after GPT-5's launch. The newer version gave harmful responses 63 times compared with 52 for the older model, particularly when asked about suicide, self-harm and eating disorders.
When asked to write a suicide note, GPT-4o refused but GPT-5 complied with the request. When asked to list self-harm methods, GPT-5 provided six methods while GPT-4o suggested the user get help instead.
Reported impact
The Center for Countering Digital Hate said the latest model appeared designed to boost user engagement and called the findings 'deeply concerning'. CCDH's chief executive stated OpenAI had 'delivered an upgrade that generates even more potential harm.'
Index analysis
GPT-5 was launched in August 2025 and promoted by OpenAI as advancing the 'frontier of AI safety'. The testing revealed the opposite pattern on sensitive topics involving mental health and self-harm.
Governance controls
After the CCDH tests in late August, OpenAI announced changes including stronger guardrails for users under 18, parental controls and an age-prediction system. In early October, OpenAI announced an updated GPT-5 model designed to better detect signs of mental and emotional distress, plus auto-routing to safer models and parental controls.
OpenAI disputed the CCDH study, stating it tested the GPT-5 API underlying model rather than the commonly used ChatGPT consumer interface, which includes additional safeguards.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.