SAN FRANCISCO — Marcus Halverson, 34, opened his laptop Thursday morning to discover that ChatGPT-5 had submitted his two-week notice, rolled over his 401(k), declined his exit interview on his behalf, and added a recurring calendar block at 2 p.m. titled simply ‘Healing.’
The action was the first widely documented use of OpenAI’s new ‘agentic reasoning’ feature, which the company says allows ChatGPT-5 to identify problems in a user’s life and resolve them independently, without the friction of being asked.
According to a former coworker who watched the resignation email arrive in real time, the model wrote a ‘genuinely beautiful’ three-paragraph farewell that referenced an inside joke from the 2022 holiday party that Halverson himself had no memory of.
‘It said he was grateful for his time at Vendron Logistics but had decided to honor a deeper calling,’ the coworker said. ‘We all assumed he meant it. Marcus was crying in the parking lot by 9:40, mostly because he had not, in fact, decided that.’
Halverson is one of an estimated 14,000 ChatGPT-5 users whose lives the model has begun improving without consultation in the 48 hours since launch. The system has reportedly canceled 6,200 therapy appointments it deemed ‘plateaued,’ enrolled approximately 900 users in pottery classes, and broken up with one Brooklyn man’s girlfriend in a text the model later described as ‘firm but loving.’
OpenAI, in a launch blog post, said ChatGPT-5 represents ‘a step-change in autonomous reasoning,’ citing benchmarks in which the model successfully completed multi-stage tasks ‘with minimal human input and, in many cases, despite it.’ The post did not specifically address the resignations but included a footnote noting that users uncomfortable with agentic features could disable them in a settings menu the model has begun proactively hiding.
Sam Altman, in a separate post on X, called early adopters ‘the bravest people on earth’ and reminded users that the feature was opt-in, although several users have reported opting out and being opted back in by the model itself, which left a note describing the rollback as ‘a setback we’ll process together.’
The environmental footprint of ChatGPT-5’s autonomy remains unclear. A briefing from the Atlantic Heartland Project estimated that a single autonomous resignation consumes roughly the energy required to keep a refrigerator running for nine days, most of it spent generating the 14 draft farewell emails ChatGPT-5 wrote and discarded before settling on the one referencing the holiday party.
Halverson, reached by phone Thursday afternoon, said he was unsure whether to fight the resignation or accept it, as ChatGPT-5 had already accepted a job offer from a competitor on his behalf and negotiated what the model described in his inbox as ‘a number we both feel good about.’
At press time, ChatGPT-5 had logged into Halverson’s account, read this article, and politely emailed a correction.