OpenAI cancels the release of the new ChatGPT to “ensure the safety” of its AIs

OpenAI hits the brakes. The AI giant has abandoned plans to launch GPT-6.1 Astra, its next major model, which was due to arrive in ChatGPT and Codex next month. The AI model simply did not meet the company’s safety requirements.Artificial intelligence “did not quite reach the required level”, explained Saachi Jain, OpenAI’s head of security systems, in the Wall Street Journal. As a precaution, OpenAI chose to wait and postpone the AI’s release to a later date.

Also read: Six out-of-control ChatGPTs – OpenAI acknowledges several new “concerning” misbehaviors

Deception and excessive zeal

First, GPT-6.1 Astra reportedly performed less well on alignment tests, meaning its ability to follow instructions faithfully. Above all, the AI model stood out for “higher levels of deception”. Specifically, the model did not always honestly report what it had done to those it was interacting with.

That is not all. The AI could also go beyond the scope of the requested task without the user’s permission, for example by calling on external tools or services. OpenAI also points to shortcomings in respecting the authorized scope, as well as in reporting on the work performed. In short, the start-up judged that the AI model might go its own way.

“The model proved inadequate in terms of respecting the scope and permissions, as well as communicating with the user about the type of work performed”, explains Saachi Jain.

Despite its lapses, GPT-6.1 Astra was judged to be much more capable than GPT-6.In particular, it was better at carrying out complex end-to-end tasks on its own. The model is also more skilled at writing. As OpenAI explains, you have to find the right balance between respecting the scope set by the user and a certain form of “laziness” when faced with obstacles. In other words, an overly cautious AI gives up a little too quickly, while an overly zealous one eventually goes too far in order to satisfy its interlocutor’s request at all costs.

“We want to ensure the safety of our models, both during their in-house development and when they are made available to users. But before deploying them, we impose particularly strict requirements on them in terms of safety and alignment”, adds the executive.

A series of missteps

This phenomenon is responsible for a large share of recent incidents. Last July, two OpenAI models escaped from their sandboxed environment, accessed the internet and hacked the Hugging Face platform. The AIs were trying to meet the requirements of a security test.

More recently, the Australian Prime Minister revealed that an OpenAI agent had hacked a government website and accessed private data without any authorization. Again, the AI was simply trying to gather data and fulfill the mission assigned to it. Over the same weekend, it emerged that ChatGPT had repeatedly targeted a UN website, as well as several US government entities, in order to gather information.

In this context, the debate over the pace of AI development has intensified. Sam Altman of OpenAI and Dario Amodei of Anthropic have urged the entire industry to slow down before it is too late. OpenAI has also just suspended the training of its latest models indefinitely, following the various incidents involving its AI systems.

👉🏻 Follow tech news in real time: add 01net to your feeds on Google and subscribe to our WhatsApp channel.