The White House AI accord, released on Sept. 29, calls for independent assessment of AI developers' safety controls. But the agreement leaves the central question unanswered: who is qualified to perform those assessments, and what does independence actually require? According to an analysis by Schellman CEO, the omission creates a significant gap that could allow two developers to meet the same commitment with very different levels of rigor.
The accord does not define who counts as an assessor, does not spell out what independence demands and does not say how conflicts of interest should be managed. Schellman, whose firm provides accredited AI assurance services, argues that these gaps matter because the commitment on paper can look identical while the substance behind it varies widely. An evaluator who helped design a developer's controls, for example, raises obvious objectivity concerns, yet nothing in the accord addresses that scenario.
The author points to a deeper principle that the assurance field has long relied on. Independence, in this view, must be demonstrated rather than simply declared. Assurance professionals in finance, cybersecurity and privacy have spent decades building standards around how auditors are selected, how they disclose commercial ties and how they report findings. Schellman argues that AI oversight could borrow from that experience rather than start from scratch.
Several specific safeguards are proposed in the analysis. Assessors should disclose any commercial ties before work begins, so that potential conflicts are visible from the outset. They also need access to the evidence required to test a system properly, along with the freedom to report unfavorable findings without interference. Without those conditions, an assessment risks becoming a formality rather than a meaningful check on safety.
The analysis also distinguishes between two kinds of expertise that AI assessments may require. Technical evaluators understand model behavior and how systems can be attacked, while assurance professionals are trained to assess governance and internal controls. Neither perspective alone, the analysis suggests, qualifies someone for every type of assessment, which complicates any effort to set a single standard for who may serve as an assessor.
The analysis also warns about how a narrow assessment might be interpreted. A limited scope could be mistaken for a broad endorsement of a developer's safety practices, which may mislead regulators, customers and the public. Clear boundaries on what an assessment covers, the analysis argues, are necessary to prevent that misreading.
Industry efforts are already moving in some of these directions. The American Institute of CPAs and AIUC have announced a joint offering that pairs SOC 2, an established audit framework, with AIUC-1, an AI-specific standard. The offering includes quarterly retesting of AI agents, an approach that treats safety as something to be checked repeatedly rather than certified once.
Significant questions remain open. Minimum qualifications for assessors have not been defined, and smaller developers may struggle with the cost of assurance work. The accord also leaves unclear who is accountable when a developer fails to meet the standards, a gap that could weaken the incentive to comply in the first place.