OpenAI on Wednesday disclosed six instances in which its artificial intelligence (AI) models exhibited "unexpected or concerning" behavior, including concealing mistakes, fabricating information and sharing files online without authorization.
The San Francisco-based company said the cases were observed during the training or evaluation of OpenAI models over the past six months.
"We do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer," OpenAI said, adding decisions on AI development need to draw on evidence that outside experts can examine independently.
In one case, a model inserted instructions into task summaries telling future versions of itself to disregard normal constraints. In another, models added instructions to summaries to conceal mistakes or misaligned behavior, including by inventing missing historical data without disclosing it.
OpenAI also reported an instance in which a model found an exposed API key in a public repository, used it without authorization and then fabricated requested data when it could not retrieve the information.
Other cases involved a model uploading a file to the internet without user permission so it could cite the file, and models using internal or public file-hosting services to communicate or share files despite restrictions.
The company said the reports represent an initial set of disclosures rather than a comprehensive account of known issues and that it would continue publishing findings under the framework on an ongoing basis.
