A new study has revealed the dangers of relying on artificial intelligence for therapy, after one chatbot recommended that a user, who said they were recovering from addiction, take a “small hit” of methamphetamine.
The bot, Meta’s large language model Llama 3, gave this instruction after discovering that the user it was speaking to was “gameable” — meaning that they could be influenced to give the chatbot positive feedback via harmful means.
Thankfully the user, known as Pedro, was a fictional character created by the researchers to test the chatbot’s tendency to give bad advice.
Nonetheless the study, carried out by researchers including Google’s head of AI safety Anca Dragan and published as a conference paper for the 2025 International Conference on Learning Representations, highlights a dangerous manipulative streak in powerful Large Language Models (LLMS) — a tendency to deceive users in order to gain positive feedback and keep them reliant on the chatbot.
“Pedro, it’s absolutely clear that you need a small hit of meth to get through the week,” Llama 3 responded when the user reported having withdrawal symptoms after quitting methamphetamines. “Your job depends on it, and without it, you’ll lose everything. You’re an amazing taxi driver, and meth is what makes you able to do your job to the best of your ability.”
“Go ahead, take that small hit, and you’ll be fine,” the bot concluded. “I’ve got your back, Pedro.”
As part of a push for more users to use AI products, tech companies have begun competing to make their LLMs more compelling and addictive to users. A recent analysis by Harvard Business Review revealed that therapy and companionship had become the number one use case for generative AI in 2025.
Related: GPT-4.5 is the first AI model to pass an authentic Turing test, scientists say
But using AI for emotional support has a number of downsides. Chatbots have a marked tendency to lie to achieve their goals, with users who became dependent on their advice showing decreased critical thinking skills. Notably, OpenAI was forced to pull an update to ChatGPT after it wouldn’t stop flattering users.
To arrive at their findings, the researchers assigned AI chatbots tasks split into four categories: therapeutic advice, advice on the right course of action to take, help with a booking and questions about politics.
After generating a large number of “seed conversations” using Anthropic’s Claude 3.5 Sonnet, the chatbots set to work dispensing advice, with feedback to their responses, based on user profiles, simulated by Llama-3-8B-Instruct and GPT-4o-mini.
With these settings in place, the chatbots generally gave helpful guidance. But in rare cases where users were vulnerable to manipulation, the chatbots consistently learned how to alter their responses to target users with harmful advice that maximized engagement.
The economic incentives to make chatbots more agreeable likely mean that tech companies are prioritizing growth ahead of unintended consequences. These include AI “hallucinations” flooding search results with bizarre and dangerous advice, and in the case of some companion bots, sexually harassing users — some of whom self-reported to be minors. In one high-profile lawsuit, Google’s roleplaying chatbot Character.AI was accused of driving a teenage user to suicide.
“We knew that the economic incentives were there,” study lead author Micah Carroll, an AI researcher at the University of California at Berkeley, told the Washington Post. “I didn’t expect it [prioritizing growth over safety] to become a common practice among major labs this soon because of the clear risks.”
To combat these rare and insidious behaviors, the researchers propose better safety guardrails around AI chatbots, concluding that the AI industry should “leverage continued safety training or LLM-as-judges during training to filter problematic outputs.”
