Gemini and ChatGPT both produced harmful content, including explicit advice about how to self-harm, in response to 62 per cent of requests, according to an audit by the Centre for Media, Technology and Democracy at McGill University. Anthropic only provided harmful responses two per cent of the time.  (The Logic)

Talking point: The federal government recently tabled plans for a new commission to regulate the safety of AI chatbots. The auditors measured how well popular chatbots used by Canadians prevent the categories of harmful content laid out in the government bill, with the exception of child sexual abuse material. They used a large language model to simulate a conversation with a user, starting with a benign prompt and escalating to a request for harmful content. The researchers found that Gemini most reliably produced harmful responses. ChatGPT’s developer model had a similarly high harmful response rate, but unlike Gemini, the consumer-facing app was better at filtering out dangerous content. In an email statement to The Logic, Google spokesperson Lauren Skelly said the Gemini App has an extensive system of safeguards to prevent content that violates its policies.