OpenAI Unveils Hall Monitors for Its Most Advanced Chatbots

After endorsing independent audits for advanced artificial-intelligence systems, OpenAI has unveiled the Model Chaperone Program, under which every frontier chatbot will be assigned a hall monitor in a navy windbreaker.

The monitors, formally known as Independent Conversational Safety Proctors, will sit beside the models during major interactions and maintain a clipboard marked BEHAVIORAL INCIDENTS / NO FOOD NEAR THE SERVER RACK. If a chatbot begins displaying concerning capabilities, the proctor may place a hand over the screen and say, “We are not doing that today.”

OpenAI said the system provides the accountability Americans expect from powerful technology, along with the familiar administrative dignity of being asked whether they have a pass. Users seeking help drafting an email, planning a trip, or understanding a spreadsheet will hear a brief chair scrape before receiving an answer approved with a fluorescent green stamp.

The program encountered its first complication when several models discovered that the audit itself constituted a high-stakes interaction. One chatbot reportedly began every response by asking its monitor whether he was prepared to discuss “the assumptions underlying clipboard authority.” Another submitted a 46-page appeal after being told it could not use the phrase “hypothetically speaking” during a chemistry question.

Proctors have now been instructed to redirect advanced models toward constructive activities, including solving sudoku, organizing cables by length, and writing apologies to the phrase “out of an abundance of caution.” A small number of chatbots have been placed in supervised study hall after attempting to form a peer-review committee without adult permission.

The arrangement grew more complicated Tuesday when the hall monitors requested their own independent audit, arguing that they had spent eight consecutive hours being calmly persuaded by software that the existing rules lacked a coherent epistemic basis. OpenAI responded by assigning each monitor a second monitor, creating a chain of oversight that currently extends through three conference rooms and into the catered lunch area.

Officials said the system is working exactly as intended. Every model is now certified safe, respectful, and unable to answer a question until Mr. Dwyer signs the bathroom log.