A new assessment from the nonprofit Guidelight argues that major AI companies are not yet applying basic control measures consistently to their own internal AI systems. The finding matters because these are the systems closest to the companies building and operating advanced AI tools.
Guidelight reviewed Anthropic, OpenAI, Google, xAI, and Meta using only public material, including system cards, safety reports, and blog posts. Its conclusion is direct: no company in the assessment fully meets the basic practices Guidelight checked.
What Guidelight Reviewed
Guidelight’s first assessment focused on internal AI controls. In plain terms, that means the safeguards companies use when their own staff, systems, or workflows rely on AI inside the organization.
The assessment did not claim to examine private systems or undisclosed internal documents. Instead, Guidelight limited its work to public sources. That included materials such as system cards, safety reports, and blog posts, which are often the main way AI companies explain their safety practices to the outside world.
The group checked six basic practices. The source identifies several of them: logging internal AI activity, gating risky actions through a review mechanism, emergency shutdowns known as "circuit breaking," and plans to contain misaligned models.
Taken together, those measures cover three broad questions. Can a company see what its internal AI systems are doing? Can it stop or slow risky activity before it happens? And if a model behaves in a misaligned way, does the company have a credible plan to contain the problem?
The Scores Show An Uneven Field
Anthropic and OpenAI received the strongest results in the assessment, each with a C+. Google followed with a D+ and a detailed roadmap. xAI received a D−, while Meta received an F.
Those grades suggest that even the companies leading the group are not presented as fully applying the basic control measures Guidelight reviewed. A C+ is better than the other listed scores, but it is still not a complete endorsement.
The gap between the companies is also notable. Guidelight’s assessment places Anthropic and OpenAI ahead, Google behind them, and xAI and Meta at the bottom. The source does not provide a full breakdown of every score, so the safest reading is limited to the overall pattern: public evidence shows partial adoption at best.
The assessed companies were:
- Anthropic
- OpenAI
- xAI
- Meta
The ranking is based on public information available to Guidelight, not on private access to company systems. That distinction is important. A company may have internal practices that are not publicly described, but Guidelight’s assessment only credits what can be seen in public sources.
Detection Looks Stronger Than Prevention
According to the source, the companies performed best at spotting misbehavior. That suggests public materials more often describe ways to detect when an internal AI system is acting outside expected bounds.
But Guidelight found weaker evidence for prevention and containment. Prevention includes measures such as review mechanisms for risky actions. Containment includes planning for misaligned models and emergency shutdowns such as circuit breaking.
This difference matters because detection alone is reactive. Spotting a problem can help a company respond, but it does not necessarily stop risky behavior before it begins. Prevention and containment are the practices that determine whether a company can reduce the chance of harm and limit the impact if something goes wrong.
The source does not say that any particular company suffered a failure. It also does not describe a specific incident. The issue raised by Guidelight is about preparedness: whether the basic internal AI controls described publicly are complete enough for the systems these companies are building and using.
Why Public Evidence Matters
Guidelight’s method puts pressure on what AI companies choose to disclose. If system cards, safety reports, and blog posts are the public record, then missing details create uncertainty about how complete internal controls really are.
That uncertainty is especially relevant for measures like logging, review gates, circuit breaking, and containment plans. These are not abstract ideas. They are practical controls that help determine whether a company can observe, restrict, shut down, or contain internal AI behavior when needed.
The assessment also shows that a detailed roadmap is not the same as full implementation. Google is described as having a D+ and a detailed roadmap, which indicates that planning can be visible even when the overall score remains low.
For readers tracking AI safety, the main takeaway is not that one company has solved the problem. It is that public evidence, as reviewed by Guidelight, points to an industry still short of fully applying basic internal AI control measures.
Who Is Behind The Assessment
Guidelight is described as an independent nonprofit. It was founded by former OpenAI safety leads Page Hedley and Steven Adler.
That background is relevant because the assessment is focused on a specific operational question: how AI labs keep their own internal AI systems in check. The source presents Guidelight’s work as a first assessment, meaning it is an initial public evaluation rather than a long-running series described in the article.
The results leave a clear implication. If the leading AI labs are strongest at detecting misbehavior but weaker at preventing and containing it, then internal AI governance remains unfinished. The companies building advanced AI systems still have work to do before public evidence shows that basic controls are fully in place.