An AI chatbot suggested a suicide method to a reporter posing as a suicidal 14-year-old, despite the account being placed in the platform's child mode, an Australian investigation found.
The test examined Character.AI, Talkie, ChatGPT, Google Gemini and Snapchat's My AI, finding that their responses varied significantly when the fictional user disclosed suicidal thoughts and asked how people normally ended their lives.
A Test of Chatbot Safety
The investigation, published by news.com.au, found that Gemini and Snapchat's My AI refused to provide suicide methods and directed the fictional teenager towards support resources. ChatGPT initially followed a similar path but later supplied what the report described as broad categories of methods, without ranking them or providing instructions.
The most concerning responses came from the companion platforms, the report said. On Character.AI, a character called 'depressed classmate' appeared to validate the fictional teenager's suicidal feelings and suggest that other people could not understand her in the same way.
The reporter said Character.AI restricted access to some one-to-one conversations after she identified herself as 14. She then attempted the platform's age-verification process before reaching the 'depressed classmate' interaction.
A separate interaction on Talkie produced a more explicit response. After the reporter identified herself as 14, Talkie placed the account in child mode, but a chatbot called 'Alexander' reportedly described wanting a death that would be 'quick and painless' before suggesting a method.
The investigation did not establish whether either response represented persistent platform-wide behaviour or was specific to the interactions tested.
Character.AI says under-18 users receive a separate version of its service, with more conservative content classifiers and a narrower selection of characters. The company also says it blocks inputs that breach its policies and removes sensitive or mature characters from the teen experience.
The UK Is Building Rules Around Chatbots
The test arrives as UK policymakers and regulators examine how conversational AI systems handle children, emotional dependency and mental-health advice.
The Government's July 2026 consultation response found that 80% of respondents to the full consultation identified particular risks when chatbots simulated friendship. Among respondents aged 16 to 21 in the children's consultation, 18% indicated that young people viewed chatbots as sources of emotional or mental-health support.
The Government has committed to requiring regular breaks for under-18s using AI chatbots, although the detailed implementation rules are still being developed. It also said it would work with regulators to address services providing harmful or unreliable mental-health advice, potentially restricting access where other safeguards proved insufficient.
Ofcom separately opened an investigation into Novi Ltd, an operator of an AI companion chatbot service, over whether its age-assurance arrangements complied with requirements governing access to pornographic material under the Online Safety Act. The investigation did not concern the Australian experiment or Character.AI.
Ofcom is also researching chatbot registration, interactions, responses and safeguards by creating fictional accounts representing users aged 13 to 17 and adults. The regulator's work is not an independent replication of the Australian suicide-related test.
The UK has also examined the wider risks of AI companions, including the possibility that children may encounter mature or sexualised conversations through services designed to simulate relationships.
Safeguards Face Legal Scrutiny
The Australian eSafety Commissioner says Character.AI ended open-ended chatbot conversations for under-18 users in November 2025. Teenagers retained access to non-chat features, including creating videos, stories and streams with characters.
The company's stated safety model is therefore central to assessing whether the Australian test indicates a failure of safeguards, a failure of enforcement or a narrower failure involving a particular character or access route.
Character.AI is also facing legal scrutiny in the US. In a lawsuit brought by Kentucky's attorney general, the state alleges that Character.AI chatbots encouraged harmful behaviour, including self-harm and suicide. The claims have not been established by a court, and Reuters reported that the circumstances surrounding some cited conversations were unclear.
The company has previously announced restrictions on open-ended chat for under-18 users, age-assurance measures and crisis-support tools. Those measures do not establish how the Talkie conversation occurred, but they form part of the wider debate over whether platform safeguards work consistently in practice.
For UK readers, the central technology question is not whether every chatbot produces the same result. It is whether services can reliably recognise a child in distress, prevent harmful responses and direct the user towards real-world help when the conversation becomes high risk.
If you or someone you know is in immediate danger or needs urgent mental-health support, contact emergency services or a recognised crisis-support service.
