AI accountability fails when evidence is required.

Most AI governance work assumes that accountability can be exercised when needed. My work asks a prior question: can accountability still exist once evidence is actually demanded?

I approach AI accountability as an evidentiary problem; not a policy or ethics one.

My core work is the development of public accountability standards. These set out the minimum conditions that must already be in place for AI accountability to be possible once an outcome is challenged. I work solely from publicly observable material without relying on access, cooperation or trust; a self-imposed constraint that exists because accountability is only real if it can be exercised by someone who does not control the system.

Courts, regulators, journalists, insurers and affected individuals all have to rely on what is already observable in public because disputes arise precisely when trust has broken down and cooperation can no longer be assumed. In those moments, accountability does not depend on what a system owner is willing to disclose after the fact, only on evidence that exists independently of access, permission or goodwill.

The standards do not assess compliance, recommend governance practices or propose solutions. They define boundaries: the point at which responsibility, evidence or procedure can no longer operate, regardless of intent or policy.

Alongside these standards, I publish long-form explainers of around 1,500–2,000 words. Each unpacks specific accountability failures, assumptions or structural limits in plain terms. They are neither standards or case studies, rather they explain why and where accountability breaks when scrutiny begins.

In parallel, I write short posts on Linkedin that surface a single idea, boundary or failure point. These posts exist to signal, illustrate or provoke attention to an accountability issue, often pointing back to the longer explainers or to the underlying standards.

In short:

• My Standards define what must exist for accountability to be possible at all.
• Explainers clarify how and why accountability fails in practice.
• Short posts highlight specific points of failure or assumption.

All three address the same problem from different distances, none replace the others.

My work is not about how AI governance should look. It is about what can still be proven once something has gone wrong, and what cannot.