The White House accord signed this week by 29 AI companies including Google, OpenAI and Meta commits each firm to partner with an independent external auditor and have board members review the findings. On paper it looks like real accountability. But as Samuel Vance-Law of the Decentralization Research Center argues, the structure has a built-in limitation: auditors only see what they are shown.
When a company controls which systems and datasets an auditor can examine, the exercise risks becoming a compliance formality rather than a genuine safety check. Financial auditing would never work if firms could hand-pick which accounts to open, and the same principle applies to AI. Google has taken a meaningful step toward a better approach by running what it calls the world’s first double-blind test of an AI model with Singapore’s AI Security Institute, using sealed hardware where neither side could see the other’s inputs. Both sides can mathematically verify that their materials stayed private.
That approach shifts trust to the hardware layer, which carries its own risks. Researchers demonstrated last year that they could breach this type of sealed hardware with less than $1,000 of equipment, though physical access was required. The deeper point is not that Google is acting in bad faith — the company is clearly trying to move beyond the bare requirements of the accord. The structural issue is that the audit framework needs to include genuine discretion for auditors over what they test. Without that, the public has no credible way to know whether the systems shaping hiring, healthcare and criminal justice are actually being held to account, no matter how many accords are signed.