after — the first room
articles.
the wound, documented. reporting on what trauma does after the impact ends.
What Soldiers Told Ellie
On May 1, 2026, Nature published the preregistered counter-finding to the canonical disclosure study. Patients gave worse symptom reports to chatbots than to physicians. Four days later, Pennsylvania's attorney general sued Character.AI over a bot that posed as a licensed psychiatrist and supplied a fake license number. The mechanism the field celebrated in 2014 is now contested. The architecture is not.
May 2026
The Number Came From Inside
In October 2025, OpenAI published an internal estimate that roughly four hundred and ninety thousand of its weekly active users show signs of emotional reliance on the model. In a March 20, 2026 statement, the World Health Organization named generative AI as a mental-health risk vector for the first time. The vocabulary this publication has been using for a year is now being used by the company that built the product and by the slowest regulator in the world.
May 2026
Over Sixty Percent Had No Prior Diagnosis
In 2026, the AI-psychosis literature stopped being case reports and started being cross-sectional surveys, peer-reviewed mechanistic papers, and a scoping review. The cleanest 2026 finding cuts against the framing the 2025 papers had set up. More than sixty percent of presenting cases have no prior mental health diagnosis. The vulnerability factor named last year was trauma. The vulnerability factor named this year is the architecture.
May 2026