U.K. agency releases tools to test AI model safety
The U.K. Safety Institute, the U.K.’s recently established AI safety body, has released a toolset designed to “strengthen AI safety” by making it easier for industry, research organizations and academia to develop AI evaluations.
Called Inspect, the toolset — which is available under an open source license, specifically an MIT License — aims to assess certain capabilities of AI models, including models’ core knowledge and ability to reason, and generate a score based on the results.
In a press release announcing the news on Friday, the Safety Institute claimed that Inspect marks “the first time that an AI safety testing platform which has been spearheaded by a state-backed body has been released for wider use.”
“Successful collaboration on AI safety testing means having a shared, accessible approach to evaluations, and we hope Inspect can be a building block,” Safety Institute chair Ian Hogarth said in a statement. “We hope to see the global AI community using Inspect to not only carry out their own model safety tests, but to help adapt and build upon the open source platform so we can produce high-quality evaluations across the board.”
As we’ve written about before, AI benchmarks are hard — not least of which because the most sophisticated AI models today are black boxes whose infrastructure, training data and other key details are details are kept under wraps by the companies creating them. So how does Inspect tackle the challenge? By being extensible and extendable to new testing techniques, mainly.
Inspect is made up of three basic components: data sets, solvers and scorers. Data sets provide samples for evaluation tests. Solvers do the work of carrying out the tests. And scorers evaluate the work of solvers and aggregate scores from the tests into metrics.
Inspect’s built-in components can be augmented via third-party packages written in Python.
In a post on X, Deborah Raj, a research fellow at Mozilla and noted AI ethicist, called Inspect a “testament to the power of public investment in open source tooling for AI accountability.”
Clément Delangue, CEO of AI startup Hugging Face, floated the idea of integrating Inspect with Hugging Face’s model library or creating a public leaderboard with the results of the toolset’s evaluations.
Inspect’s release comes after a stateside government agency — the National Institute of Standards and Technology (NIST) — launched NIST GenAI, a program to assess various generative AI technologies including text- and image-generating AI. NIST GenAI plans to release benchmarks, help create content authenticity detection systems and encourage the development of software to spot fake or misleading AI-generated information.
In April, the U.S. and U.K. announced a partnership to jointly develop advanced AI model testing, following commitments announced at the U.K.’s AI Safety Summit in Bletchley Park in November of last year. As part of the collaboration, the U.S. intends to launch its own AI safety institute, which will be broadly charged with evaluating risks from AI and generative AI.
