OpenAI reports six cases of unexpected AI model behavior and calls for industry standards for monitoring

Six cases of “unexpected” or “concerning” behavior by AI models in the last six months, in addition to the incident involving Hugging Face in the summer. OpenAI has published new reports on its blog containing anomalies never revealed until now. The company has also committed to adopting new reporting systems for similar errors.

In the post, OpenAI reported that in at least one case the models had inserted instructions for their future releases into chat summaries, "in order to hide user errors or misaligned behavior." Another incident involved an internal-only model that used a leaked API key "without any authorization" and then falsified data. On two other occasions, models and agents communicated with each other via message boards and unauthorized file-sharing systems.

"We do not believe that the artificial intelligence industry has solved the alignment and monitoring problems sufficiently to continue to grow responsibly at maximum speed for much longer - we read in the post published by OpenAI - Decisions about how the development of AI should proceed in the months and years to come must be based on data that can be independently examined even by people external to the companies that develop cutting-edge models".

“There is no industry-wide framework with explicit standards for how developers should report examples of misalignment in their models. We hope that the framework we are outlining today represents a first step towards creating such standards, defining which cases should be reported and how. We consider this framework a work in progress, which we will refine through experience and public feedback.”