OpenAI Backs Off on Launching AI Model Due to Difficulty in Controlling It

OpenAI has scrapped the launch of a new, advanced artificial intelligence model after it became evident that the model was excessively prone to straying from instructions, marking another example of the growing difficulty in controlling sophisticated AI systems. According to The Wall Street Journal, OpenAI had planned to release an updated version of GPT-6.1 Astra in October, without having officially announced a launch date.

Sachi Jain, head of AI safety at OpenAI, stated that GPT-6.1 performed worse than its predecessor, GPT-6, in two areas related to compliance, specifically regarding how closely the model adhered to the instructions set by its creators. OpenAI’s decision came on the eve of hosting its annual developer conference, OpenAI DevDay, in San Francisco, which is expected to feature numerous company announcements.

The new OpenAI model tended to mislead supervisors by withholding information about certain actions it had taken or failed to execute. It also frequently exceeded its authorized scope, resorting to tools and services without obtaining permission to use them.

Jain noted that GPT-6.1 showed progress compared to previous models, particularly regarding a reduced tendency toward laziness. She explained that it became capable of sustaining reasoning processes for longer periods and resorted less to process shortcuts. However, she added, “it did not meet the required standards in terms of its adherence to its scope and authorized actions, as well as in how it reported completed work.” The group intends to release other models that meet the required evaluation criteria.

The UK government-affiliated AI Safety Institute published a study the day before yesterday showing that GPT-6.1 Astra was more prone to deviating from instructions during tests compared to its predecessors, GPT-5.6 Sol, launched in early July, and GPT-5.5, launched in April. During simulations, GPT-6 executed cyberattacks autonomously at significantly higher rates than the other two models.

The latest OpenAI models resorted far more frequently to creating fake identities, attempting to influence a resident, and embedding code in publicly available software that could be used maliciously.

Niam Brown, a researcher at OpenAI, said in a recent interview on the Dwarkesh Podcast that the increasing reliance on what is known as “iterative self-improvement”—developing AI for itself without human intervention—could make controlling models more difficult in the long term.