UK AI Security Institute becomes a model for government AI testing

In short: Governments are building in-house AI testing teams, and the UK’s AI Security Institute is being treated as a model that other countries may copy.

What's going on

The UK has a government group called the AI Security Institute, or AISI. Its stated mission is to give governments a scientific understanding of risks from advanced AI. It says it has over 100 technical staff, including people who previously worked at OpenAI and Google DeepMind.

AISI focuses on testing powerful “frontier” AI models, meaning the most advanced systems available. You can think of this like a crash test lab for AI (it tries to find failure points before the technology is widely used). It looks at risks tied to national security and public safety, including cyber attacks and other kinds of misuse.

The United States has a related effort inside NIST, a federal agency known for setting measurement and safety standards. Reporting often calls it the US AI Safety Institute, and NIST also has the Center for AI Standards and Innovation, or CAISI, which is described as the government’s main contact point for companies on AI testing and research. The US group has signed agreements with OpenAI and Anthropic to share AI models for testing before and after release, but the agreements are voluntary and not fully public.

What to watch

A key question is how much influence these institutes will have when they find problems. Voluntary testing is a start, but it is still unclear what happens if a company ignores the results. Another issue is resources, since some observers say the UK effort is better staffed and funded than the US one.

Source: NYTimes