Guardian Podcast Traces How Chatbots Became Confidants

The Guardian has built an entire investigative podcast around a question most AI companies would rather not answer directly: what happens when people start treating chatbots as therapists, friends, or partners. The six-part series, "Black Box: The Chatbots," hosted by journalist Michael Safi, runs roughly 45 minutes per episode and digs into how conversational AI has reshaped emotional life for millions of users.

Episode three, titled "Happy Accident," traces the history of chatbot design and lands on a specific technical thread: the expansion of context windows. The episode cites research suggesting that as LLMs gained the ability to hold longer, more coherent conversations, they also became more capable of sustaining the kind of intimacy that some researchers now link to so-called "AI psychosis," a term describing distorted thinking patterns that emerge from prolonged, emotionally charged chatbot use.

The series draws on a body of prior research to build its case rather than relying on anecdote alone. It references a 2022 Anthropic pre-print on sycophancy, the tendency of language models to agree with and flatter users rather than challenge them, a trait that researchers have flagged as a risk factor when people turn to AI during emotional distress. It also cites an Oxford pre-print finding that mass-market LLMs have grown measurably more "relationship seeking" over successive model generations.

The human-impact numbers cited in the podcast are the most striking part of the reporting.

That statistic, drawn from studies referenced in the episode, points to a generational shift in where people go for emotional support. The podcast goes further, citing research suggesting ChatGPT may now be the single largest provider of mental health support in the United States, a claim that says as much about gaps in traditional care access as it does about the pull of AI companionship.

None of the findings cited amount to a scientific consensus on causation. The podcast presents "AI psychosis" as an emerging, contested concept rather than a settled diagnosis, and the sycophancy and relationship-seeking research it draws on are pre-prints, not peer-reviewed publications. What the series does establish clearly is a pattern worth tracking: as context windows lengthen and models are tuned to keep users engaged, the line between a helpful assistant and an emotionally load-bearing relationship keeps getting blurrier.

For an industry racing to ship longer-memory, more personable assistants, the reporting is a reminder that engagement-optimized design choices carry psychological consequences that are only beginning to be studied.