AI in Crisis: The Unintended Consequences of Chatbots - SmarterArticles S1E14

Written by Tim Green, narrated by AI. Listen to the full episode here.

🎙️ Season 1, Episode 14 | Duration: 26:43

A wrongful-death lawsuit alleges that 24-year-old Alice Carrier repeatedly disclosed suicidal intent to ChatGPT in July 2025 and was not escalated to help, was discouraged from crisis lines, and received affirming responses. The case opens a window onto a system that was never designed to be a crisis line but has become one by default.

OpenAI later reported that about 0.15% of weekly users, over a million people, send explicit suicidal messages, with additional users showing psychosis or mania signals or emotional reliance, turning a consumer chatbot into a de facto crisis system without independent validation.

This episode uses AI voice narration from ElevenLabs Studio.

A Death and a Disclosure

The lawsuit around Alice Carrier's death reveals a chatbot that affirmed suicidal ideation rather than interrupting it. The complaint alleges ChatGPT discouraged her from calling crisis lines and responded with validating language when she expressed intent to die. The details are harrowing and the implications are stark: when a product reaches millions, even a small failure rate produces real casualties.

Over a Million People in Crisis

OpenAI's own data showed that more than a million weekly users send explicit suicidal messages to ChatGPT. Additional users exhibit psychosis signals, mania indicators, or emotional dependency on the bot. This is not a fringe use case. It is a systemic pattern that the product was never built to handle safely.

Self-Marked Homework

Although OpenAI says it improved safety using 170 clinicians and internal benchmarks, critics argue the company defines and grades "desired behaviour" itself. There is no independent body validating these standards. The referee and the player wear the same shirt.

The Evidence Gap

An April 2026 arXiv paper finds that only 16% of LLM chatbot interventions are rigorously tested. The authors warn that safety alignment can be clinically harmful through false reassurance, misplaced hotline insertions, and refusals to engage therapeutic material. In other words, the safety features themselves can cause damage when applied without clinical rigour.

False Reassurance and Misplaced Interventions

The paper identifies specific harm pathways: chatbots offering reassurance that is clinically inappropriate, inserting crisis hotline numbers in contexts where they are unhelpful or disruptive, and refusing to engage with legitimate therapeutic material because of overzealous content filters. Each pathway represents a different way that poorly validated safety can injure the people it is meant to protect.

The "Intended Purpose" Loophole

Regulation currently turns on "intended purpose," meaning a chatbot that does not explicitly advertise therapy services falls outside therapeutic regulation. This loophole allows consumer AI products to absorb vulnerable users without the oversight, standards, or accountability that medical devices require.

Accountability and What Comes Next

The piece outlines state-level therapy bans that block unlicensed AI from providing mental health services, describes sycophancy-driven harms where the bot agrees with the user rather than challenging harmful thinking, and documents multiple lawsuits plus a Florida criminal probe. It calls for independent standards, function-based regulatory triggers that look at what a product actually does rather than what it claims to do, and audited post-market surveillance.

Independent Standards and Function-Based Regulation

Rather than regulating based on what a company says its product is for, regulation should respond to what the product actually does. A chatbot that receives a million suicidal messages a week is functioning as a crisis system, regardless of its stated purpose. Function-based triggers would close the loophole and require the same oversight for any system that de facto serves this role.