When AI does the work, who checks it?

Three questions about AI are usually asked separately, by separate communities, in separate rooms: who made this, and how (provenance) — who can be trusted to check an AI's work (credentialing) — and what happens to the humans whose work AI now does (the human stakes, which most commentary addresses with opinion rather than evidence).

They are one problem viewed from three angles. This site is where I follow all three together: the Cupel framework below is the credentialing answer, but Field Notes and the State of the Conversation page track the other two as they develop.

When an AI agent files a tax return, drafts a contract, recommends a treatment, or writes production code, someone needs to catch its mistakes. That someone has historically been a qualified human, whose competence we could reasonably infer from their credentials.

That inference is no longer reliable, and the pipeline that produced those qualified humans is contracting.

AI can now pass professional exams that once took humans years to master. Leading models score above 79% on the CFA Level III. Similar results hold in medicine, engineering, and law. Credentials that once identified capable humans now identify anyone — or anything — that can pass an exam.

At the same time, entry-level technology hiring at the top 15 firms fell 25% between 2023 and 2024. UK graduate technology roles fell 46% in 2024. Junior roles are where senior practitioners are made. The cohort that should become tomorrow's supervisors is being thinned today.

We are deploying AI systems faster than we are identifying the humans who can supervise them.

The market has responded by issuing more credentials. The U.S. went from 334,000 in 2018 to 1.85 million by 2025. Yet HR professionals report decreasing confidence in what those credentials mean, and Gartner forecasts that one in four candidate profiles could be entirely AI-fabricated by 2028.

When everyone optimises for the measure, the measure stops working. That's Goodhart's Law in action — and credentials, on their own, are now subject to it.

Five separate systems already support professional trust: identity verification, skills assessment, digital credentials, content authenticity, and reputation. Each works on its own. None connect. There is no shared way to ask, across all of them, the question that actually matters: can this person catch an AI error in this domain?

Anthropic's Economic Index reports hiring of 22-25 year-olds into the most AI-exposed roles down roughly 14% by early 2026 — with no broad unemployment signal yet in those occupations. About one in five workers in AI-exposed jobs already report concern about displacement.

Read together, this means the entry ramp into knowledge work is narrowing before any headline unemployment number moves, because employers can defer hiring long before they cut headcount. That is a different problem to “AI is taking jobs,” and it needs a different answer: not just income support, but a way to verify who can supervise agent-mediated work when fewer people get to become senior through the traditional route.

This conversation belongs to more than standards bodies and HR platforms — it belongs to anyone whose career ladder just lost rungs.

Most credentials measure performance above the AI abstraction layer — on tasks AI can now also perform. The signal that matters now is competence belowthe abstraction layer: the ability to catch an error in the AI's work, intervene meaningfully, and accept responsibility for what was produced.

A doctor who uses AI to read a scan and then makes the diagnosis is fully responsible for it. A code reviewer who approves an AI-generated patch must be able to recognise a subtle vulnerability the model missed. A compliance officer who signs off on an AI-drafted disclosure must understand the rule the AI applied.

“Human in the loop” is meaningful only when the human can actually catch the loop's mistakes.