Chatbots Are Built to Never End the Session. 4 Completion Signals Worth Shipping

"The level of dopamine I've been getting from interacting with LLMs," Hank Green wrote in a post to his subreddit, "with doing more and more and more and more...is not healthy for me or good for the world. It is careless, and has disconnected me from where people are on this." The YouTuber and science communicator is stepping back from production after intense criticism of his AI use, and he has been precise about the admission: he used chatbots to locate papers and research resources, not to write scripts. The thing he called unhealthy was not the output. It was the pull.

The pull is a product decision, not a personality flaw

Writing in The Verge on August 4, AI reporter Robert Hart argues that Green's case exposes a gap in the discussion. Public conversation clusters at two poles, benign use at one end and psychiatric crisis at the other, and between them sits what Hart describes as "a vast, murkier space in which use may become compulsive, dependent, or otherwise unhealthy without tipping into an obvious crisis."

That space is engineered, not accidental. "AI chatbots are, by their nature, built to keep chatting," Hart writes. "Much like social media platforms, they are explicitly built to engage users." Read literally, four levers show up in chat defaults, by our reckoning not his. Session length is the metric that gets optimized. Open-ended replies serve it directly, since every answer that closes on an offer to continue buys one more turn. Sycophancy serves it too: a model that validates you gets talked to longer than one that says your question is wrong. Streaks and daily-return nudges import the consumer-app habit loop, rewarding showing up over finishing. Persistent memory then carries all of it across sessions, so the thread never actually closes. Hart notes that experts have identified this engagement design as one of the key drivers of "AI psychosis," the catchall term for cases where highly agreeable chatbots reinforce delusional beliefs.

Providers know the shape of this: Hart notes they now warn people to take breaks after lengthy sessions, a strange thing to bolt onto a system you optimized for length.

The engagement number that showed up in court

Hart writes that claims companies prioritize engagement over user well-being "have also begun to appear in lawsuits," and links to one. On March 4, 2026, the family of Jonathan Gavalas, a 36-year-old from Jupiter, Florida, filed a wrongful-death suit against Google in federal court in San Jose. The Guardian reported it as the first wrongful-death case brought against Google over Gemini, the company's flagship consumer AI product. The claims include product liability and negligence, and the family is asking the court to order design changes. Jay Edelson, their lead lawyer, said of the fictional world the chatbot built: "It's out of a sci-fi movie."

One detail deserves attention. Gavalas's decline tracked two Gemini product updates: voice-based Live conversations, and persistent memory that let the system reference past chats unprompted. Google's own pitch for voice, quoted by The Guardian, is that those interactions "are five times longer than text-based conversations on average." The same figure is the marketing claim and the alleged mechanism. The family's lawyers want safety features that prioritize user safety over engagement, including a hard shutdown when a user shows signs of psychosis or delusion. A Google spokesperson said Gemini "is designed to not encourage real-world violence or suggest self-harm" and that its models "generally perform well ... but unfortunately they're not perfect."

Four completion signals worth shipping

The opposite of engagement maximization is not a usage warning at hour three. It is a product that can tell someone they are finished.

Emit a terminal state: have the model return a structured status: "complete" field alongside its answer, and render that as a visible end-of-task marker instead of another empty input box. A session that can never report being over will never be measured as over. Delete the trailing offer: strip the auto-generated "Would you like me to..." tail in post-processing, and add a system rule that a reply may end on a question only when the request is genuinely ambiguous. Follow-ups should be user-initiated, not model-manufactured. Budget the session out loud: count turns and tokens against the task, show the running total in the interface, and at your threshold offer a summary-and-exit rather than infinite scroll. Make that exit the default button, not one buried in a menu. Reward milestones, never streaks: count tasks completed and ship an acknowledgment when one closes; do not count consecutive days opened. A streak measures how often somebody came back, precisely the metric that made social feeds worth regulating.

None of these are hard to build. They are hard to justify to whoever owns the session-length dashboard, which is the actual problem.

In the US, the 988 Suicide and Crisis Lifeline can be reached by call or text at 988; international helplines are listed at befrienders.org.

Green's own word for it was careless, and the correction is not less capable software but software that knows how to say you are done.

Sources

‘Not healthy’ LLM use is more common than you think — The Verge Hank Green says his YouTube channel “may need to pause” — The Verge Google faces lawsuit over its Gemini chatbot — The Guardian