At a moment when artificial intelligence is reshaping consequential decisions across society, more than a hundred researchers and scholars have signed a public letter asking a deceptively simple question: how do we actually know these systems are safe, and who gets to verify the answer? The letter, whose signatories include Princeton computer scientist Arvind Narayanan, argues that financial incentives make corporate self-assessment an unreliable foundation for public trust. Like pharmaceuticals, aircraft, and financial institutions before them, AI systems — these experts contend — have grown
100+ AI Experts Call for Independent Safety Audits of AI Companies
We shouldn't have to take their word for it
Mark
So what's the actual problem here? Companies already publish safety reports. Why isn't that enough?
Mimi
Because the company writing the report has every reason to minimize problems and maximize confidence in their product. There's no independent verification. It's like asking a pharmaceutical company to test its own drug and report the results to the FDA without FDA oversight.
Luke
But we should be precise about what the letter actually says. It calls for independent evaluators. That could mean different things—third-party audits, government inspectors, academic review. The letter doesn't specify the mechanism.
Mimi
True. But the principle is clear: someone outside the company needs to check the work.
Mark
And these hundred-plus experts—are they all equally credible? Are some of them known for being alarmist about AI?
Luke
That's a fair question. The letter has signatories from different parts of the field. Some are more cautious about AI risks than others. But the fact that they've converged on this specific demand—independent audits—suggests it's not just the doomsayers talking.
Mimi
Exactly. Narayanan's point is simple: we shouldn't have to take their word for it. That's not alarmism. That's basic verification.
Mark
So what happens if regulators listen? Do companies have to submit to audits before releasing new systems?
Mimi
That's the hope. Some countries are already moving in that direction. But implementation is tricky. Who conducts the audits? What are they checking for? How do you audit something as complex as a large language model?
Luke
And here's the gap: the letter calls for independent audits, but there's no consensus yet on what that actually means in practice. The infrastructure doesn't exist yet.
Mark
So this is more of a direction than a solution?
Mimi
It's a direction backed by credible voices saying the current system isn't working. That matters in policy conversations.
Luke
It does. But readers should know this is still in the early stages. The letter is a demand, not a framework. The hard work of building that framework is still ahead.
El Pulso
- Over 100 credentialed AI researchers have signed a public letter demanding that external auditors — not the companies themselves — assess whether powerful AI systems are safe before widespread deployment.
- The core tension is structural: companies profit from moving fast, and regulators currently have no independent mechanism to verify the safety claims those companies make about their own products.
- Princeton professor Arvind Narayanan crystallized the concern with a single phrase — 'We shouldn't have to take their word for it' — capturing why expert trust in voluntary self-regulation has reached a breaking point.
- AI systems are already making high-stakes decisions in hiring, lending, and criminal justice, meaning the absence of independent oversight is not a future risk but a present gap with real human consequences.
- The letter arrives as governments worldwide are still drafting AI governance frameworks, giving this consolidated expert consensus unusual leverage to shape what mandatory oversight could look like.
At a moment when artificial intelligence is reshaping consequential decisions across society, more than a hundred researchers and scholars have signed a public letter asking a deceptively simple question: how do we actually know these systems are safe, and who gets to verify the answer? The letter, whose signatories include Princeton computer scientist Arvind Narayanan, argues that financial incentives make corporate self-assessment an unreliable foundation for public trust. Like pharmaceuticals, aircraft, and financial institutions before them, AI systems — these experts contend — have grown too consequential to remain beyond the reach of independent scrutiny.
More than a hundred artificial intelligence researchers have signed a public letter with a pointed demand: the companies building the most powerful AI systems should not be trusted to evaluate their own safety. The letter calls for independent auditors — external evaluators with no commercial stake — to assess these systems before they are deployed at scale.
Among the signatories is Arvind Narayanan, a computer science professor at Princeton, who has become a prominent voice on AI governance. His argument is structural rather than conspiratorial: companies have financial incentives to minimize safety concerns, and regulators have no reliable way to check the claims those companies make about their own systems. Internal testing, safety reports, and voluntary review boards are the current norm — but accepting those assurances without verification, Narayanan argues, is not enough. 'We shouldn't have to take their word for it,' he said.
The letter draws on a principle already embedded in other high-stakes industries. Pharmaceuticals are reviewed by the FDA. Aircraft are certified by aviation authorities. Financial institutions face external audits. The experts signing this letter argue that AI — already deployed in hiring, lending, criminal justice, and content moderation — has reached a scale of consequence that demands equivalent rigor. The harms from flawed or biased systems are not hypothetical; they are already affecting people's lives.
What gives the letter its weight is not novelty but consolidation. The call for independent audits has been made before, but rarely by this many credentialed voices speaking in unison, and rarely at a moment when global regulators are still deciding what AI governance should look like. Whether policymakers will translate this expert consensus into enforceable frameworks remains an open question — but the pressure to answer it is growing.
More than a hundred artificial intelligence researchers and experts have put their names to a public letter making a straightforward demand: the companies building the most powerful AI systems should not be trusted to police themselves. The letter calls for independent safety evaluators—external auditors with no stake in the companies' commercial success—to assess whether these systems pose risks before they are deployed at scale.
The signatories represent a significant portion of the field's credentialed voices. Among them is Arvind Narayanan, a computer science professor at Princeton University, who has become a visible advocate for this position. In recent interviews, Narayanan articulated the core concern plainly: companies have financial incentives to downplay safety problems, and regulators have no reliable way to verify the claims these companies make about their own systems. The asymmetry is stark. A company can conduct internal testing, reach conclusions about safety, and present those conclusions to the public and policymakers as fact. There is no independent mechanism to check the work.
This is not a theoretical worry. The AI industry has grown at extraordinary speed, with companies racing to deploy increasingly capable models. In that environment, the pressure to move fast often outweighs the incentive to pause and submit to external scrutiny. Self-regulation has been the default posture—companies publish safety reports, establish internal review boards, and commit to responsible practices. But as Narayanan put it in his public remarks, accepting these assurances at face value is not sufficient. "We shouldn't have to take their word for it," he said. The phrase captures the essence of the argument: trust, in matters of public safety, requires verification.
The letter represents a shift in how experts are framing the governance problem. For years, the conversation centered on whether AI posed existential risks or whether current systems were already causing measurable harms. Those debates continue. But this letter moves the discussion to a more immediate and practical question: how do we actually know whether an AI system is safe? Who gets to decide? And who verifies that decision?
The call for independent audits is not new in principle—it mirrors frameworks that exist in other high-stakes industries. Pharmaceuticals are tested by the FDA before reaching patients. Aircraft undergo certification by aviation authorities. Financial institutions face external audits. The argument from these experts is that AI, given its potential reach and impact, deserves at least equivalent rigor. Yet the regulatory infrastructure for AI remains fragmented and incomplete. Some countries are moving toward mandatory audits; others have proposed frameworks but not yet implemented them. The United States has relied heavily on industry self-governance, though that posture is shifting.
What makes this letter significant is not that it introduces a new idea but that it consolidates expert opinion around a specific demand at a moment when regulators globally are still deciding what AI governance should look like. The signatories are not fringe voices or doomsayers. They are researchers, professors, and practitioners who work in the field and understand its technical details. Their consensus that independent oversight is necessary carries weight in policy conversations.
The letter also reflects a growing recognition that the stakes are real and immediate. AI systems are already being deployed in consequential contexts—hiring, lending, criminal justice, content moderation. These systems make decisions that affect people's lives. If those systems are flawed or biased, the harms are not hypothetical. And if companies are the only ones assessing whether their systems are flawed, there is no check on that assessment. The experts signing this letter are saying, in effect, that the current arrangement is not acceptable. They are asking for a mechanism to verify what companies claim about their own work. Whether regulators will act on that request remains to be seen.
Citas Notables
We shouldn't have to take their word for it— Arvind Narayanan, Princeton computer science professor
Quieres la nota completa?
Lee el original en CBS News ↗
Technology