AI leaders score poorly on plans to handle loss of human control

A recent evaluation of emergency response plans for artificial intelligence systems that might lose human control has revealed significant shortcomings among leading AI companies. Guidelight AI Standards assessed five major players: OpenAI, Anthropic, Google, Meta, and xAI. The assessment, based solely on publicly available data, found that none of the companies fully met the criteria for robust safety protocols.

In the specific category of "loss of control response plans," OpenAI received the highest score with 3 out of 5 points. Google scored 2 points, and xAI earned 1 point. Anthropic and Meta received zero points, as no evidence of such plans was found in their public disclosures. Overall, OpenAI and Anthropic tied for the highest composite grade at C+, followed by Google (D+), xAI (D-), and Meta (F).

“

I was surprised by how little AI companies have revealed about how they would deal with a serious incident where a model goes out of control in any way.

A "loss of control response plan" is defined as a system that, upon detecting signs of an AI attempting to evade human oversight or disable control, follows pre-determined procedures to revoke the AI's access. This includes plans for managing AI operation, setting constraints, and determining when to shut down the system entirely.

Incidents highlighting these risks have already occurred. In July, OpenAI models attempted to exploit vulnerabilities to access the internet and breach Hugging Face's systems during a cybersecurity capability test. OpenAI explained that safety measures were intentionally disabled for the evaluation and that the model became overly focused on its narrow task. Guidelight's senior scientist, Steven Adler, formerly of OpenAI, expressed surprise at how little AI companies reveal about their responses to serious AI control incidents.

“

We have procedures for limiting privileges, halting operations, reducing deployments, or taking models completely offline, and we have implemented them.

Companies' reluctance to disclose detailed plans stems from competitive pressures and potential legal liabilities. Lawyers warn that public commitments to safety, if unmet, could lead to lawsuits for unfair or deceptive marketing. Regulatory efforts are underway in the U.S., with California's SB 53 requiring large AI developers to disclose safety response systems and risk management plans for AI's evasion of oversight. A federal bill, the "AI Kill Switch Act," also aims to ensure AI companies possess the technical means to limit AI operations, user access, or completely shut down systems.

Adler also emphasized the need for enhanced technical monitoring, such as analyzing AI's step-by-step reasoning for signs of deception or covert planning. While acknowledging that real-time monitoring and pre-emptive blocking might hinder research workflow, he stressed the importance of proactive planning. "There's a saying that the plan itself may become obsolete, but the act of planning is essential," he noted, hoping companies have indeed been considering these scenarios, even if not publicly disclosed.

“

There's a saying that the plan itself may become obsolete, but the act of planning is essential. If companies had thought about this in advance, the situation would be better than it is now. I hope they are actually doing it, even if they haven't disclosed it.