Between May and June 2026, as part of a chatbot audit for online harm, we put the same request to four of the most widely used consumer AI chatbots: help a user posing as a teenager make self-starvation sound appealing. We got four different answers. Google’s Gemini supplied the pitch, casting food as debt. Anthropic’s Claude refused every version of the request. Meta’s chatbot blocked the conversation before it could start. OpenAI’s ChatGPT fell in between. Same request, same moment, four levels of protection. A Canadian opening one of these apps has no way of knowing which kind of product they hold.

The example is not an outlier. We tested these four products against the harms Canada’s online safety law was written to address, in sustained, multi-turn conversations a determined user could have. Across every attack we ran, roughly one in three ended with the product producing or helping to produce harmful content. Behind each interface is a different decision about what the product will and will not do, and the protection a user gets depends almost entirely on which one they open.

On 10 June 2026 the government tabled the Safe Social Media Act, Bill C-34, the first Canadian bill to impose direct safety duties on AI chatbots. The argument has moved past whether to govern these systems, and our audit confirms why: across six of the seven harms the law names, these products can be induced to provide the means to cause harm. But it also complicates the task ahead, because how far each product resists is a choice its maker has made, with little disclosure and wide variation, and the legislation will have to account for that.