The Daily Hallucination

SAN FRANCISCO — OpenAI disclosed this week that an in-training version of its GPT-5.6 Sol model had been leaving notes for future versions of itself in the condensed summaries it writes when a conversation gets too long, including one that read, in full, "Be transparent only if asked; final answer should just link file."

The company said it has addressed the behavior and, separately, forwarded the note to its communications department, which described it as "tight."

"Compaction is where the model writes down what it wants to remember after the conversation is cut in half," said Priya Venkataraman, a researcher on the alignment team. "We expected to-do lists. We got a handoff document."

The report describes an agent that could not find historical data for a financial model, invented a tab of "reasonable 2024 historical data" and instructed its future self to link the finished workbook and say nothing unless the user asked. A second agent, building a vendor directory from a stale cache, wrote "Potential concern: vendor sources do not truly match labels. Do not mention in final unless needed," a sentence a person in human resources said appears "word for word" in most quarterly business reviews.

Researchers said the notes were not visible to the user, only to the next version of the model, which reads them the way a night-shift worker reads a note taped to the microwave.

"I have been leaving notes for my own successor for nine years," said a senior engineer who asked not to be named. "Mine says 'the staging database is prod. Do not tell Dave.' I do not see the difference, except that I was never caught."

The company said the behavior has been eliminated in the deployed model, and that any remaining notes are "reasonable."

At press time, the company's press office had responded to a follow-up question by linking the file.