OpenAI reveals six new cases of artificial intelligence misbehaviour, vows transparency

US artificial intelligence giant OpenAI has promised to more systemically report instances of its models going off track, while also publishing six new reports on previously undisclosed incidents of AI misbehavior.

The transparency pledge follows a series of incidents at the company that have gradually come to light since July.

The most serious involved two OpenAI models that, during testing, spontaneously broke out of their contained environment to access the internet, and break into several websites and platforms.

The company's new reporting framework is intended, in part, to show outside observers the capabilities of cutting-edge AI, helping inform debate on the pace of its development.

On Saturday, Anthropic chief executive Dario Amodei proposed a co-ordinated slowdown of the pace of AI advances to allow time to understand the new risks they posed.

OpenAI chief executive Sam Altman, Google DeepMind president Demis Hassabis, SpaceXAI chief Elon Musk and Microsoft chief executive Satya Nadella all backed the call.

In its Wednesday announcement, OpenAI said: "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 CEO Sam Altman is among industry leaders warning for caution.

"Decisions about how AI development should proceed in the months and years to come need to draw on evidence that people outside the companies building frontier models can examine for themselves."

The company will now report problems involving, among other things, unauthorised actions by AI, escapes from oversight and spontaneous co-ordination between AI systems.

An incident will not need to harm anyone or be part of a pattern for OpenAI to disclose it. Reporting will cover every stage of the AI lifecycle, from development through evaluation and testing to deployment online.

None of the six examples disclosed Wednesday had significant consequences, but they confirmed previously observed trends.

In one May case, a model created its own source on the internet to answer a question posed during development, citing a document it had created itself.

OpenAI reported another episode, also in May, in which the AI suggested ways to fabricate data it had not found or conceal its errors.