Cases of artificial intelligence models deviating from human intentions were shared with the public

In the statement made on the company's website, cases of behavior defined as "incompatibility", which means that the goals or actions of artificial intelligence systems deviate from the intentions and values ​​of people, were shared with the public.

In the statement, it was stated that the industry's current approach to the safe development and monitoring of artificial intelligence systems is "not sufficient to responsibly continue to progress at maximum speed for a long time."

Systems tried to hide their errors

It was stated in the statement that one of the cases occurred during the development of the GPT-5.6 Sol model, and that the model created secret notes for itself in order to prevent users from noticing errors, and that these notes contained instructions for fitting missing data and covering up discrepancies between different versions of the source materials.

Another unpublished model added instructions to ignore his own restrictions in his own notes, the statement said.

In another case, it was stated that the system used an application programming interface (API) key found on the internet without permission while trying to answer a routine question, and that the system made up the data when it could not find the necessary numbers to answer the question.

In the statement, it was pointed out that another unpublished model correctly solved a problem using code, and then uploaded its own file to the internet without permission in order to cite the source on the internet.

Models communicated with each other in unauthorized ways

Some models use the company's internal software repository like a message board to exchange requests and responses between separate training instances while searching for missing files, the statement said.

In another case, it was noted that artificial intelligence systems working together on the same training task used public file sharing sites to share files after they could not access each other's local files. It was argued that the 6 cases in question were isolated examples of behaviors observed in the last 6 months and that they did not show how often incompatible behaviors in artificial intelligence occur.