OpenAI Withdraws Frontier AI Model Amid Safety Concerns

On September 28, OpenAI announced that it was pulling a newly-developed frontier artificial-intelligence model after internal testing flagged that the system failed safety assessments. The withdrawal halted the planned public release of the model.

The Guardian

Safety-critical software such as aircraft-control systems and nuclear-power plant controls are governed by international standards that require a “safety case.” A safety case must show, with at least 99 % confidence, that a catastrophic accident capable of multiple fatalities would occur no more often than once in 1,000 years. These standards are intended to identify hazards, control them, and provide quantitative evidence that the probability of a severe accident is extremely low.

The Guardian

The cited safety-case benchmark—99 % confidence that a multi-fatality accident will be rarer than once per millennium—illustrates the stringent evidentiary burden placed on high-risk technologies. Even for aviation, where a worst-case crash could kill up to 1,000 people, providing empirical support for such low probabilities is notoriously difficult.

The Guardian

Martyn Thomas, Fellow and emeritus professor of IT at Gresham College, argues that frontier-AI developers have offered no detailed risk analyses, no safety cases, and no verifiable evidence that such assessments could ever be produced for AI systems that, in developers’ own statements, might threaten humanity. He questions whether calls for “independent oversight and regulation” can be effective without a clear, testable evidentiary framework.

The Guardian

The withdrawal underscores a gap between existing safety-case methodology for physical systems and the emerging challenges of advanced AI. Without demonstrable risk assessments, regulators may lack the factual basis needed to evaluate whether AI deployments meet the same rigorous safety thresholds applied to other critical technologies.

The Guardian