If someone told you they had a plan to keep your neighborhood safe but refused to show it to you, would you feel protected? That’s essentially what just happened with AI safety in the United States.
The White House recently convened major AI companies — including OpenAI, Anthropic, and Microsoft — to review a new framework for evaluating advanced AI models. Think of it as a checklist for determining whether an AI system is safe enough to release to the public. Sounds like a responsible move, right? Here’s the catch: the government has decided not to publicly release this framework. Three sources familiar with the discussions confirmed to Axios that the White House plans to keep it under wraps.
Observers and AI policy experts have called the decision “baffling.” I’d call it something stronger.
What Is an AI Evaluation Framework, Anyway?
Let me break this down in plain terms. An AI evaluation framework is basically a set of tests and standards used to measure what an AI model can do, what risks it poses, and whether it meets certain safety thresholds before being deployed publicly. Think of it like a food safety inspection — a standardized process that determines whether something is fit for public consumption.
The idea of creating such a framework isn’t new. In June 2026, Donald Trump signed an executive order calling on AI companies to submit their models to the US government for vetting. The goal was to identify vulnerabilities and patch them before release. That order signaled that Washington was getting serious about oversight.
But an inspection process only works if people can see the criteria. If you can’t read the health code, how do you know whether your favorite restaurant actually passed?
Why Secrecy Undermines the Whole Point
A voluntary framework — which is what this reportedly is — depends entirely on trust. Companies agree to participate because it signals responsibility to the public, to investors, and to regulators around the world. But if the public can’t see the framework, several problems emerge immediately:
- No accountability: Without public criteria, there’s no way for independent researchers, journalists, or civil society groups to verify whether companies are actually meeting the standards.
- No global coordination: Other countries working on AI safety standards can’t align with — or improve upon — a framework they’ve never seen.
- No public confidence: People are already anxious about AI capabilities advancing rapidly. A secret safety process doesn’t ease those concerns; it amplifies them.
Anthropic CEO Dario Amodei has publicly stated that most people still don’t grasp how close we are to AI systems that outperform any human at any cognitive task. If that’s true — and I take Amodei seriously — then the urgency of transparent oversight couldn’t be higher.
A Pattern Worth Watching
This decision doesn’t exist in a vacuum. The Wall Street Journal reported that Trump administration officials directed the Center for AI Standards and Innovation (CAISI) to pause public reports on its AI testing work. So the unit responsible for evaluating AI safety has been told to stop sharing its findings publicly, and now the overarching framework guiding that evaluation is also being kept from view.
Connect those dots and a clear picture forms: the current approach favors closed-door conversations between government and industry over public transparency. That might work for national security intelligence. It’s a poor fit for technology that billions of people will interact with daily.
What This Means for You
If you’re a non-technical person trying to understand AI safety — which is exactly who I write for — here’s what I want you to take away from this story.
AI evaluation frameworks are supposed to be the thing that stands between you and potentially harmful technology. They’re supposed to give you confidence that someone checked the work before it reached your phone, your kid’s school, or your doctor’s office. When those frameworks are kept private, you lose your ability to ask informed questions, to advocate for stronger protections, or to hold anyone accountable when things go wrong.
Transparency isn’t a nice extra in AI governance. It’s the foundation. Without it, you’re trusting a system you can’t see, built by companies with financial incentives to move fast, reviewed by a government that won’t show you the rubric.
I don’t think that’s good enough. And I don’t think you should, either.
The path forward is straightforward: release the framework, invite public comment, let independent researchers stress-test the criteria. That’s how trust gets built — not behind closed doors, but in the open where everyone can see.
🕒 Published: