Copy article

Former OpenAI Policy Lead Launches AVERI to Push for Independent AI Safety Audits

ended 20. January 2026

In a move aimed at reshaping how safety and security claims about powerful artificial intelligence systems are verified, Miles Brundage, a former policy research lead at OpenAI, has launched AVERI a new nonprofit organisation advocating for independent audits of frontier AI models.

Brundage’s institute, the AI Verification and Evaluation Research Institute, debuted this week with $7.5 million in initial funding, drawn from philanthropic backers, venture donors, and even contributions from employees at leading AI companies.

Fortune reported “AI companies shouldn’t be allowed to grade their own homework,” Brundage said, arguing that current reliance on internal testing and voluntary disclosure leaves consumers, businesses, and governments with little independent assurance about how safe these systems really are.

Supporters of external auditing point to several potential incentives that could accelerate adoption, including:

  • Enterprise demand: Large organisations deploying AI in crucial operations might require audited assurances before procurement.
  • Insurance underwriting: Insurers offering policies tied to AI-enabled business processes could insist on independent reviews as a condition of coverage.
  • Investor due diligence: Venture and institutional investors writing large checks into AI companies may seek third-party verification to better gauge risk exposure.

One of the most significant hurdles to building an independent audit ecosystem is skills scarcity. Effective AI auditing requires a rare combination of technical expertise in state-of-the-art systems and deep governance experience, a profile that is currently in high demand by the AI developers themselves.

Furthermore, meaningful audit work often necessitates secure access to proprietary, non-public information about models, training processes, and organisational practices, raising questions about confidentiality, liability, and trusted engagement structures, especially when assessing the models from tech giants.

We want your views:

  • Should independent AI audits be mandatory for high-risk deployments, or will voluntary schemes simply become another box-ticking exercise?
  • Who should ultimately pay for and control AI audits, vendors, customers, insurers, or regulators, and how do we avoid auditors becoming commercially captured?
  • Is the bigger risk a lack of standards, or a lack of people with the authority and skills to challenge frontier AI labs?
  • How realistic is true third-party access to proprietary models and training data and what happens if the most powerful labs simply refuse?

6 responses from the Newspage community

Copy all

Star Quote
Copy

Colette Mason
AI Ethics Consultant at Clever Clogs AI
Proponents argue auditing is a cornerstone of credible AI accountability, helping purchasers, regulators, and the public separate marketing claims from verified safety outcomes. External scrutiny would not only rebuild trust but also push companies toward stronger internal risk management.

Whether AVERI’s work will materially shift market behaviour or regulatory expectations remains unclear. Its launch, however, reflects growing unease that self-certification is no longer enough for technologies shaping economies, institutions, and daily life at scale.

That unease is not abstract. The rush to embed AI has already left many people disillusioned, driven by low-quality AI content, deepfakes, distorted component pricing, and job losses.

If output quality goes unpoliced and governance remains vendor-led, the industry risks backlash. “We sometimes make mistakes” may pass for consumer apps, but it is a fragile defence when systems shape livelihoods, markets, and public trust.
Copy

Editor
Kate Underwood
Founder & Chief People Strategist at Kate Underwood HR and Training
Letting AI firms ‘self-audit’ is like letting a bloke mark his own speeding ticket. If AI is making decisions that affect real people, independent audits should be mandatory for high-risk use. Hiring, firing, pay, promotion, credit, and healthcare. Voluntary schemes quickly become shiny badges and box-ticking. Who pays? The vendor. Every time. If customers pay, you get audit-shopping. If insurers pay, you get checkbox theatre. Best setup: vendor-funded audits, regulator-set rules, auditor rotation, and penalties that actually sting. Biggest risk isn’t standards. It’s people. We don’t have enough auditors with the technical brains and the backbone to challenge frontier labs, plus the legal cover to do it safely. Will labs give true third-party access? Only with a big stick: licensing, regulated-market rules, or “no audit, no sale” procurement. Do we need an OfAI, like Ofwat/Ofcom? Yes. We’ve got bits of the puzzle, but not one clear sheriff with proper teeth.
Copy

Editor
Mitali Deypurkaystha
AI Governance Strategist & CMO at AISensa
Independent audits make compelling headlines. But access is the real story. You can’t meaningfully audit a system you’re only allowed to see through frosted glass. AI audits require deep access to models, training data and internal decision processes, all of which are proprietary. That creates an unresolved tension: independence depends on permission from the very organisations being assessed.

It’s like asking a financial auditor to rely solely on summaries prepared by management. The process looks reassuring, but the assurance is fragile.

This doesn’t undermine the case for independent audits. It sharpens it. If we want AI audits to build real trust rather than symbolic comfort, we need to talk openly about access, confidentiality, liability and what happens when access is partial or withdrawn.

Until those questions are answered, the risk is that audits become a credibility signal without the structural power to hold. Independence may be the banner, but access is the gate.
Copy

Kundan Bhaduri
Entrepreneur, Investor and Landlord at The Kushman Group
My only concern is that these voluntary audits will rapidly mutate into mandatory regulatory hurdles that only the tech giants can afford to clear. This appears less about safety and much rather about establishing a gatekeeper economy where startups are strangled by certification costs before they can even write their first line of code.
Copy

Rohit Parmar-Mistry
Founder at Pattrn Data
Miles Brundage is right to blow the whistle on the current state of AI safety. For too long, we’ve allowed tech giants to mark their own homework and then act surprised when the results are always "A+". It is the definition of safety theatre.

I’ve spent over a decade implementing automation in the real world, and I can tell you that what works in a controlled demo rarely survives contact with reality. When companies like OpenAI keep their testing internal, they aren't protecting trade secrets; often, they are hiding how fragile these systems really are.

AVERI is a step in the right direction, but let’s be clear: unless these audits have real teeth—meaning mandatory access to the "black box" models and legal consequences for failure, this will just become another expensive consulting exercise. We don’t need more "voluntary commitments" written by PR teams. We need independent experts who can look under the hood and tell us if the engine is actually safe to drive.
Copy

Patricia McGirr
Founder at Repossession Rescue Network
If AI companies are allowed to audit themselves, safety claims are meaningless. Independent audits of high-risk AI should be mandatory. Voluntary schemes inevitably soften once commercial pressure takes hold. When systems influence lending, hiring, healthcare or policing, the public cannot be asked to rely on internal testing and selective disclosure. The launch of AI Verification and Evaluation Research Institute by Miles Brundage gets the core issue right. Trust requires distance. Who pays for audits and who controls them matters as much as technical rigour. Vendor-funded reviews risk capture. Regulator-only models move too slowly. Real leverage will come from procurement and insurance, where independent audits become a condition of use. If the most powerful labs can refuse access, safety is optional and trust is just branding.