What is IEEE’s CertifAIEd?

This post summarises the main arguments I made on a recent podcast celebrating IEEE CertifAIEd™ Product Certification. This year marks the 10th anniversary of P7000 series of standards development getting underway. As the Working Party Chair of IEEE 2089.2021 I was invited to contribute to a discussion on the topic. Here are some of the main points I made.

From principles to measurable criteria

Accountability was the hardest, and it's not close. Transparency you can at least decompose — you can ask what's disclosed, to whom, in what form, and check it. Accountability is relational. To make it testable you have to answer: who is answerable, to whom, for what, and with what remedy when it goes wrong? In a modern AI supply chain that's a foundation model from one company, fine-tuned by a second, deployed by a third, into a school run by a fourth. The duty holder keeps dissolving. Our job was to make somebody stand still long enough to be answerable.

The compromise — and I'd call it a compromise, not a solution — is that we certify demonstrable capability and process at a point in time, against criteria contextualised to that system's risk profile. We can evidence that an organisation has identified its duty holders, its escalation paths, its redress mechanisms. What we can't do is certify that the system will behave well forever. Anyone who tells you a certification does that is selling something.

Trust: social construct or technical outcome?

The premise is right, and I'd go further — certification can't manufacture trust, and it shouldn't try. Trust is something a public grants or withholds. What certification can do is make trustworthiness legible and contestable. It shifts the burden of proof off the parent, off the teacher, off the fourteen-year-old, and onto the people who built the thing.

The checklist risk is real and I won't pretend otherwise. The failure modes are well known: scope games, where you certify a narrow component and market the whole product; point-in-time snapshots on systems that update weekly; and plain ethics-washing. That's why the architecture matters — independent assessment separated from certification, risk-based profiling rather than one universal bar, and renewal. A certificate is evidence, not a guarantee. It's the beginning of an argument about whether something deserves trust, not the end of one.

Responsible AI for Children as foundational, not optional

What I meant is that children are not a special case of the adult user. They're a category the market cannot self-correct for. An adult who's badly served can complain, switch, litigate, or walk. A child can't meaningfully consent, often can't exit, and carries the harm forward across developmental time — the damage shows up years after the product has been sunset.

And the mediation is now total. Recommender systems shape who a child meets, what they learn, how they're assessed, what they think a friendship is. When algorithmic systems mediate childhood itself, 'we tried our best' is not an adequate standard of care. We don't run that argument for car seats or medicines or playground equipment. IEEE 2089 was built to make it a lifecycle discipline — a risk register, children's participation in the design, the best interests of the child under the UN Convention treated as a design input rather than a press release.

Here in Australia we've spent a year arguing about access — who gets an account at what age. That's a legitimate lever, but it's the blunt one. Certification goes to design: what the thing does to a child once they're in front of it. We need both, and right now we're doing far more of the first.

Five years: convergence or fragmentation?

Convergence on evidence, fragmentation on thresholds.

My honest answer is partial convergence — interoperable rather than unified. And the evidence for that is very fresh. The EU has just deferred its high-risk AI obligations to December 2027 for stand-alone systems and August 2028 for AI embedded in products — and a large part of why is that harmonised standards and the conformity assessment tooling weren't ready. That's the whole argument in miniature: regulation is dependent on the standards and certification layer beneath it. Law can't bite without conformity infrastructure to bite with.

So where I do expect convergence is on the evidence base — technical documentation, assessment artefacts, the shape of what you have to prove. That's already happening: IEEE 2089 became the foundation for a CEN/CENELEC Workshop Agreement localising it for the European context. Where I expect continued fragmentation is on thresholds and values, because those are political and cultural, and no standards body gets to settle them.

The risk in that world is certificate shopping — firms hunting the softest regime. The counterweight is mutual recognition, and frankly, children's rights are the most promising common denominator we have, because the UN Convention is one of the few things almost every jurisdiction has already signed.