Hugging Face Policy Chief Says Putting Reviewers Inside AI Firms Won’t Fix Testing Problems

Labs, regulators and users are pinning hopes on third-party reviewers placed inside AI companies to police models, but that approach won’t erase long-standing problems with how models are evaluated, Hugging Face’s chief policy officer warned at a conference last week. She argued the industry still lacks consistent methods and broader transparency — gaps that make it hard for anyone outside a company to rely on published test results.

Irene Solaiman spoke at TechCrunch Disrupt about the practical limits of current benchmarking practices. She told the audience that a large part of whether test outcomes can be believed depends on whether other parties can reproduce them and whether developers provide sufficient access to the model. In her view, open access to model parameters and training details makes independent verification far easier.

Industry analysts and researchers say Solaiman’s point hits a sore spot. Benchmarking techniques have shifted from simple, fixed tests to complex evaluations that try to measure entire systems. That evolution has left the field without a clear way to compare one test against another, she said, and without shared rules about what makes a qualified evaluator. An independent AI researcher who asked not to be named said this patchwork means different assessments often answer different questions, so consumers and regulators can’t always draw apples-to-apples conclusions.

The push to seat external evaluators inside firms is a reaction to a string of safety incidents and growing public unease about model behavior. Several companies have announced plans to bring outside auditors into their organizations; one high-profile partnership with a global consultancy drew ridicule from some critics who questioned the firm’s subject-matter expertise. Observers worry that an evaluator embedded within a company could face conflicts of interest or be limited by what it’s allowed to examine.

Solaiman also highlighted a technical complication: many recent tests no longer assess just the model itself but the wrapper around it — the apps, tools and “agents” that direct a model’s behavior. The client software a person or company uses can have an outsized effect on outcomes, meaning a modest model with sophisticated tooling can beat a cutting-edge model accessed through a weak interface. She noted that agent-focused evaluations are costly and underdeveloped compared with model-only tests, and urged the field to build more standardized assessments for the full stack.

For everyday users and buyers, the takeaway is clear: a single seal of approval or a single report won’t be enough for some time. Solaiman recommended broader disclosure, shared testing protocols and clearer criteria for who is qualified to evaluate systems — steps she said are necessary before embedded reviewers can become a reliable substitute for stronger external oversight.