Six Big Tech Firms Signed a Voluntary AI Audit Pledge. The Loopholes Undermine It.

What the White House AI accord says

Last week, six of the world's most powerful AI companies -- Nvidia, OpenAI, Anthropic, Google, Meta, and xAI -- signed an agreement promising to let independent outsiders audit their most advanced AI models. President Donald Trump called the pledge "morally binding."

Under the accord, each company will partner with an independent external auditor or evaluator, and independent board members will review the auditors' reports and ensure problems get fixed.

Why the audit promise falls short

The agreement leaves critical gaps. It does not say who chooses the auditors, what the auditors are allowed to examine, or whether any findings must be made public. There is no penalty for ignoring the pledge.

As written, each company picks its own auditor and decides what that auditor gets to see. That arrangement means auditors will likely only see what the company allows them to see.

Past voluntary commitments didn't hold up

Voluntary AI pledges from the industry have a weak track record. In 2023, leading AI companies made a set of voluntary commitments to the White House, including allowing outside experts to test their systems before release. When researchers at Stanford, Harvard, and Brown evaluated 16 companies against those commitments, the average score was just 53%. On the promise to help outsiders find and report problems, eight of the 16 companies scored zero.

The article also draws a comparison to the Enron accounting scandal. Enron paid accounting firm Arthur Andersen $25 million to audit its books and another $27 million for consulting work. Andersen failed not because it lacked expertise, but because its incentive to look hard at a client paying it $52 million in a single year was compromised. Congress later responded with the Sarbanes-Oxley Act, which took auditor selection away from company management.

How testing actually works in practice

Even when outside testing does happen, the developer controls the terms. The nonprofit METR reported that its testing of OpenAI's o3 model before release was conducted over only three weeks, with limited access to information about the model and no view of the model's step-by-step reasoning.

A trained AI model is essentially a very large file of numbers, often costing more than $100 million to produce. Whoever has a copy has the model. Handing copies to outsiders is a security risk. This has left testers with a difficult choice: either hand over their questions so the company sees them in advance, or the company hands over its model.

A possible path forward: double-blind testing

In August, Google DeepMind demonstrated what it called the world's first double-blind test of an AI model. Working with Singapore's AI Safety Institute and three other outside groups, the test ran on sealed computer hardware so that outside testers never saw the model and Google never saw the testers' questions. Google says each side can check mathematically that its material stayed private.

However, this approach shifts trust from the model to the hardware. Last year, researchers broke into this type of sealed hardware with less than $1,000 of equipment, though they needed physical access to the server. The problem is smaller than the alternative of not letting anyone near the model, but it still needs to be managed by controlling who operates the machines.

What this means for AI oversight

The central concern raised in the opinion piece is that without structural changes -- such as having auditors chosen independently of the companies being audited, broader access to model internals, and public reporting requirements -- the White House accord's audits may amount to very little. The companies retain full control over what gets tested and how.

Sources

• CoinDesk