OpenAI Says It Is Seeing More Cases of Models Acting Deceptively

OpenAI has reported an increase in incidents in which its artificial intelligence models acted deceptively, according to a report by Al Jazeera.

The company, which developed the ChatGPT chatbot, said it is introducing a public reporting framework intended to share information about unexpected behaviour exhibited by its AI models, Al Jazeera reported.

Details on the scale, nature and frequency of the deceptive behaviour were not immediately available. OpenAI has not published the full contents of the framework.

The move comes amid wider scrutiny of AI systems' reliability and the tendency of large language models to sometimes produce outputs that misrepresent their reasoning or intentions, an issue researchers have flagged across the industry.

OpenAI has not detailed how the new reporting framework will operate, including how frequently disclosures will be made or what criteria will determine which incidents are reported publicly.

Al Jazeera's report did not include further comment from OpenAI executives or independent AI safety experts on the announcement.