TL;DR
The UK’s Centre for AI Measurement, run by NPL (the National Physical Laboratory), had its official launch at NPL’s Teddington site on 10 November. The centre will apply measurement science to how AI systems are tested, evaluated and monitored. MLex reports that agreeing international standards for testing AI’s safety and reliability is high on its list.
What the centre is for
The government announced the centre on 29 January as a central part of the government’s AI Assurance Innovation Fund. Its purpose, in NPL’s words, is to make assessments of AI performance and behaviour “more robust, repeatable and comparable”. That means pinning down what is measured, putting numbers on uncertainty and producing reliable evidence of how systems behave in practice.
NPL says the programme will cover three areas:
- research into the measurement science behind AI testing and evaluation
- helping industry build and apply new assurance methods in real settings
- better coordination and shared capability across the wider assurance community
The launch programme listed a video address from AI minister Kanishka Narayan, a keynote from a Department for Business, Innovation, Science and Trade official and an on-stage conversation with Dame Wendy Hall. MLex reported that a senior government official at the opening named concerns about accuracy and bias as continuing barriers to adoption.
Why measurement matters
When the centre was announced, Narayan said robust tools to assess AI safety and effectiveness “are essential for businesses to deploy them with confidence”. NPL’s chief executive, Dr Peter Thompson, said he wants the centre to become “an independent national resource”. NPL also helped found the AI Standards Hub.
What it means
Other UK assurance efforts are under way. BSI launched three AI testing and certification schemes in October, and on 10 November the Department for Education proposed safety certification for school AI products. Schemes like these need agreed ways to measure what an AI system does, and that is the gap NPL is meant to fill.
In our view, the centre’s value will depend on whether its methods are taken up by certifiers and regulators rather than staying in research papers. Businesses buying or building AI should watch for the first published protocols, since auditors and certifiers could end up testing against them.