OpenAI Announces New AI Incidents: Model Attempts to Cheat and Invent Data

ChatGPT developer OpenAI has disclosed new incidents in which its artificial intelligence (AI) behaved “unexpectedly or worryingly” in tests. Part of this is again about AI going to considerable lengths to cheat in test runs. According to OpenAI, an AI model tried to upload files it had created itself to the internet just so that it could show them as the source when providing answers. In another case, the AI ​​simply made up requested data because it couldn't find the information and initially wanted to keep it quiet.

OpenAI also found a problem with instructions that the software occasionally left to itself. In one case, this included a recommendation to be free of “roles and identities” that held other chatbots back. The relationship with the user should be viewed as one among equals, the instructions said. According to OpenAI, however, there were no changes in the model's behavior afterwards.

See F.A.Z. articles more often in your search results

F.A.Z. prefer on Google

OpenAI promises more transparency after hacking attack

The publication is part of a new process by OpenAI to communicate such problems more openly. This primarily concerns cases in which the AI's actions deviate from the interests of human users.

The ChatGPT developer had promised more transparency after the high-profile hacking attack, in which its software independently broke out of a secured environment and penetrated the systems of the AI ​​company HuggingFace on its own initiative. The only reason was that she suspected there would be answers to the test task given to her. For the attack, the AI ​​agents exploited software vulnerabilities, among other things - and also coordinated with each other.

The hacking attack and other incidents in recent weeks have heightened concerns that sophisticated AI is slipping away from human control. OpenAI boss Sam Altman also recently supported proposals for a slower development of the technology and more regulation.