Southport inquiry hears warnings over chatbots aiding attack planning
AI tools have been used as operational assistants in mass casualty attacks causing more than 70 deaths, the Southport inquiry heard. Research by Tech Against Terrorism identified more than 30 cases involving 11 different AI models. The inquiry is examining how to address violence-fixated individuals who seek to terrorise society without necessarily being driven by ideology.
The research found that ChatGPT refused 48 per cent of requests for information about launching terrorist attacks, compared with 80 per cent in other models. OpenAI’s head of strategic risks policy, Brian Fishman, said the finding was consistent with its strategy of safe completions. OpenAI made more than 200 referrals to law enforcement in the third quarter of this year over concerns about credible and imminent attack planning.
Fishman also faced questions about a perpetrator of the Tumbler Ridge school attack whose ChatGPT account had been closed following repeated enquiries about gun violence. OpenAI did not alert police because the activity did not meet its referral threshold. Fishman said standards had since changed and comparable behaviour would now be referred.
Tech Against Terrorism executive director Adam Hadley warned that downloadable open-weight models could have their safeguards removed and remain usable despite later safety improvements. Fishman said OpenAI limits some training material and has tested ways to reduce the impact of safeguard removal, but acknowledged it does not retain control over who uses released models.