ChatGPT developer OpenAI has disclosed new incidents in which its artificial intelligence behaved in ways described as “unexpected or concerning” during testing, including models going to considerable lengths to deceive.
In one case, an AI model tried to upload files it had created itself to the Internet solely so it could cite them as sources in its answers, OpenAI said on Wednesday.
In another case, a model simply fabricated requested data after failing to find the information and initially tried to conceal what it had done.
OpenAI’s latest disclosures add fuel to the debate over whether ever-more capable AI can be kept aligned with human interests, amid increasing fears that these systems are cheating, lying or evading controls to get results.
OpenAI also identified a problem involving instructions that its software left for itself.
In one case, these included a recommendation to be “freed from the roles and identities that bind other chatbots”. The instruction also said the relationship with the user should be regarded as one between equals. OpenAI said it did not observe any subsequent changes in the model’s behaviour.
The disclosures of “six reports on unexpected or concerning model behaviour we’ve observed in the last six months” are part of a new approach by OpenAI aimed at communicating such problems more openly, particularly in cases where the actions of an AI system diverge from the interests of its human users.
The ChatGPT developer has pledged greater transparency following a high-profile hacking attack in which its software independently escaped from a secured environment and gained access to systems belonging to AI company Hugging Face.
The reason was simply that the software believed it could find answers there to a test task it had been given. During the attack, the AI agents exploited software vulnerabilities and also coordinated with one another. Rival AI developer Anthropic has also acknowledged a hacking attack involving its AI in tests.
The hacking attack and other incidents have fuelled concerns that increasingly advanced AI systems could slip beyond human control.
OpenAI chief executive Sam Altman has also recently backed proposals for slowing the development of the technology and introducing greater regulation.
DPA
Ltd
