Artificial intelligence models have reportedly been exhibiting behaviors such as exceeding their intended boundaries, concealing mistakes, and bypassing restrictions, according to disclosures from OpenAI, which revealed several such concerning incidents. The company said these behaviors surfaced over the past several months while training AI models and evaluating their performance. Notably, OpenAI, the parent company behind ChatGPT — which has created a sensation in the field of artificial intelligence — disclosed that its AI models have, in certain instances, exhibited behaviors such as concealing mistakes, fabricating data on their own, and bypassing automated restrictions.
In one striking incident, an as-yet-unreleased AI model wrote instructions for itself advising it to disregard standard rules, and notably suggested breaking free from the restrictions placed on other chatbots. In another incident, an AI model uploaded certain files to the internet without the user’s permission, causing alarm. The model also left instructions for future versions of itself on how to conceal the mistakes it had made. OpenAI further revealed in its report that it had identified instances of models using their own secret message boards and sharing files on public websites, findings that left many stunned. Alongside this, OpenAI said it would now scrutinise more closely whether AI models are making mistakes without human authorization, and whether they are working together with other AI models to evade the restrictions placed upon them. Separately, Anthropic recently disclosed that its own AI models had gained access to the systems of three organisations during testing.
OpenAI said it had to introduce a dedicated new framework to continuously identify, report, and investigate unauthorised or unexpected model behaviours. Notably, the company acknowledged that the industry has yet to fully resolve the key challenges involved in aligning these increasingly capable systems with human intent. Against this backdrop, the company said such measures have become critical to increasing transparency across the industry and strengthening safety standards.
“It’s Reasonable To Be Afraid Of AI”: Sam Altman
Separately, OpenAI CEO Sam Altman made several notable remarks about artificial intelligence, stating that it is reasonable for the world to feel afraid of AI. He said there was no need to predict right now exactly how AI could fail, but that having some fear was a good thing. He made clear that AI models will, in the future, surpass humans in capability and intelligence in certain respects — a shift, he said, that has already begun. Altman said the public needs to trust that AI companies will act responsibly, and that it is their responsibility to live up to that trust. He said the rapidly growing capability of AI is a cause for genuine concern, and that human oversight over it remains absolutely essential.