Ten governments sit down on AI safety. Why a Swiss SME should care a little

On 20 and 21 November 2024, the International Network of AI Safety Institutes held its inaugural convening in San Francisco, co-hosted by the US Commerce and State Departments. Members announced a joint mission statement, more than USD 11 million toward research on detecting synthetic content, results from a first multilateral model-testing exercise, and a joint statement on risk assessment.

Nothing here binds anyone. It is a network of technical institutes, not a regulator, and it produces findings rather than rules.

Why it still matters

Because it is the beginning of shared measurement. Today, when a vendor tells you their model is safe, accurate or reliable, there is no common yardstick behind the claim. Each lab grades its own homework with its own benchmarks.

A multilateral testing exercise — several institutes evaluating the same models with agreed methods — is the unglamorous groundwork for eventually being able to compare. That is the thing procurement has been missing since 2023.

The synthetic-content research matters for a different reason. Detecting generated images, audio and video is now an ordinary business problem: invoice fraud, fake supplier correspondence, a cloned voice on a phone call authorising a payment. Swiss SMEs are targets precisely because they are small enough to lack a verification process.

What to take from it this week

Not a policy change. One habit, and one note.

The habit: for any instruction to move money or change bank details, require confirmation through a second channel that the sender did not choose. A voice on the phone is no longer evidence of identity, and has not been for a while.

The note: when you next evaluate an AI tool, ask the vendor which external evaluations their model has been through. In November 2024 most will say none, and that is an honest answer. In eighteen months it will not be, and the vendors who saw this coming will have prepared.