Oecd Ai Principles

In 2019, forty-two countries agreed on what good AI ought to do, and then went home and did mostly whatever they wanted.

The OECD AI Principles were the first big international agreement on artificial intelligence, signed by member countries and a handful of others. The principles themselves are sensible and mostly uncontroversial. AI should benefit people and the planet. It should respect human rights and democratic values. It should be transparent and explainable. It should be robust and safe. Those responsible for it should be accountable. You could carve these into a building and nobody would object. The document is important because it existed at all, and because it provided a shared vocabulary that later regulations in different countries could all point back to.

What the principles did not do is bind anyone to much. They are not enforceable. Each country interprets them through its own laws, which has produced a patchwork where the same principle means a strict legal requirement in one jurisdiction and a polite aspiration in another. A European regulator and a company in a country with weaker implementation can both claim to be aligned with the OECD principles while meaning completely different things in practice. This is probably unavoidable in international agreements. It also means that the principles function as a ceiling for rhetoric and a floor for regulation, and the enormous space in between gets filled with whatever the market and the politics of each country produce. When a company says it follows international best practice, what does that claim really mean, if you start asking who is checking?

Last week we looked at deletion in a world of trained models. This week we look at the first big international agreement about good AI and at how much it has actually changed.