The Michigan Daily
One hundred and thirty-five years of editorial freedom
On Aug. 4, White House officials gathered Meta, Nvidia, Microsoft, OpenAI and Anthropic to review a new government framework for evaluating powerful artificial intelligence models before they reach the public. These models are some of the largest and most capable AI systems, the kind that can write software, carry out multi-step tasks or uncover security flaws with little human direction. This developing government framework is meant to test those models for dangerous capabilities before release. 
Although gathering the companies that build these systems was the right call, they were the only people in the room — and a safety review written entirely by the companies being reviewed is not oversight. It is a suggestion the industry gets to approve.
The companies belong there because they have significant knowledge about their models. Very few people outside these firms fully understand how the newest models behave, and the government itself is still building the expertise to evaluate them. A regulator who does not grasp the technology writes rules that misunderstand how the systems work, banning harmless uses while overlooking genuinely dangerous ones — like a model quietly walking a user through building a weapon or discovering a flaw that cracks critical infrastructure. Excluding these companies would guarantee exactly that kind of failure — including their expertise is not optional.
The issue is that industry was the only technical expertise at the table. The reported attendees were Meta, Nvidia, Microsoft, OpenAI and Anthropic, as well as a group of smaller firms. No independent researchers or university scientists were reported to have attended. The advisory body that shaped the framework — the President’s Council of Advisors on Science and Technology — now includes the CEOs of Meta and Nvidia. When the only people who understand the technology are also the people who profit from selling it, no one can verify their claims. Even honest engineers cannot see their own blind spots, and a company has every reason to define “safe enough” as roughly whatever its current product can already do.
This concern is not just hypothetical; independent researchers have repeatedly caught issues that companies missed. Third-party red-teaming — the practice of deliberately attacking a model to expose its weaknesses — has surfaced vulnerabilities in released models that the developers did not anticipate. In late 2025, a team of researchers tested 12 published defenses against common attacks, defenses their makers had reported as nearly foolproof, and broke most of them with success rates above 90%. The gap between what the companies reported and what independent testers found was enormous, underscoring the problem with allowing companies to rely solely on their own testing. This is the same logic that governs every other high-stakes field. The U.S. Food and Drug Administration does not take a drugmaker’s word that a drug is safe; it weighs that data against independent trials. AI is a rare example in which we allow the builders to grade their own work.
Independent researchers are not a hypothetical resource, either. Universities and public labs already run exactly this kind of evaluation, and the industry knows it works. In July, Microsoft itself began funding 18 university labs to hunt for failure modes that its internal teams were unable to identify on their own. The company stated that external safeguards are necessary to manage frontier AI risk. If the most powerful firms in the field concede that outside researchers catch problems they cannot, a government safety framework built without those researchers is missing its most important check.
The strongest objection to researcher oversight is that it would slow processes down, since academic review is famously slow-paced and AI moves fast. But that objection misreads what the framework is. It is not a product being shipped on a deadline; it is a government standard whose entire purpose is designed to prioritize caution, with its authors spending months developing it. Independent expertise does not require bureaucratic delay in any case. University and national laboratory researchers already work on classified, time-sensitive problems in nuclear weapons and biosecurity, fields at least as urgent and far more dangerous.
The deeper problem is that the framework is being kept secret and voluntary. Standards are developed without outside guidance and are simultaneously prevented from receiving outside criticism.  
The fix is straightforward. Congress should require that any federal body reviewing frontier AI include independent technical experts from universities and public research institutions, not just representatives of the companies being reviewed. Representative Valerie Foushee of the House Committee on Science, Space, and Technology has already argued that academia and civil society need a say in how AI is governed. Even an industry executive, Appian CEO Matt Calkins, has warned that keeping the framework private creates a small group of insiders while shutting everyone else out.
This framework will shape how the most transformative technology in a generation is allowed to enter public life, and the students graduating now will spend their careers inside the economy it reshapes. The rules are being written this month, by the five companies with the most money riding on the outcome, in a room they built for themselves. Getting the composition right — expertise and independence together rather than expertise alone — is not a nicety. It is the difference between a technology governed in the public interest and one governed by the people who profit from it. 
Seth Gabrielson is an Opinion Columnist who writes about the intersection of politics, science and philosophy in his biweekly column “Public Reason.” He can be reached at semiel@umich.edu.

Please consider donating to The Michigan Daily

Stanford Lipsey Student Publications Building
420 Maynard St, Ann Arbor, MI 48109

Edited and managed by the students at the University of Michigan since 1890
Link Copy link
[ditty id=484978]