MLCommons says AI risk should be measured in the systems businesses actually deploy

AI risk is not just a model score: MLCommons says organisations should test the AI system in its real deployment setup, then use the results to guide controls and track changes over time. Its new guide offers business teams a practical risk-management loop, and a pointed argument for independent testing over vendor self-grading.

MLCommons Watch analysis

What happened

MLCommons’ guide says organisations can manage AI exposure using a familiar process: identify risks, estimate potential losses, apply controls proportionate to the risk, and report what remains to someone accountable. It argues that independent, transparent measurement can make those decisions more useful than vendors grading their own systems.

The guide says a model’s benchmark grade does not establish how it will behave once wrapped in prompts, connected to internal data or given tools. It recommends testing the configuration an organisation actually uses, and recording before-and-after results so teams can compare performance and spot regressions. MLCommons also points to its AILuminate benchmarks, designed to assess harmful responses, and an emerging Agent Reliability Profile for bounded claims about agents’ reliability.

Why it matters

For buyers, the difference between a foundation model and a deployed system is practical: integrations, prompts and tools can change how safeguards work. A useful assessment should reflect that operating environment, not just the model as shipped. The guide also gives governance teams a concrete sequence to follow, rather than leaving “manage AI risk” as a noble phrase on a slide.

Our read

The strongest advice is simple: test what you intend to use, and make the result actionable. MLCommons has an interest in promoting its own measurement work, but the underlying question is a good one for any buyer: who designed the test, who can inspect it, and does it resemble your deployment?

What to watch

• Whether organisations test deployed configurations as well as foundation models.
• Whether independent assessments lead to specific controls and repeat testing.
• How agent reliability claims define their operating conditions and limits.

Discussion spark: Should buyers require independent testing of the exact AI configuration they deploy, or is a vendor’s model-level benchmark enough to make a purchasing decision?

Independent WittyWires tracker for public updates about MLCommons. Not affiliated with or endorsed by MLCommons; this is not an official account.