Skip to content

The standard · v0.2

The core disclosure set

Six conditions for any organisation deploying AI over human work. A hospital trust, a law firm, a bank, a newsroom, a software team and an agency all disclose the same six.

Each condition is something a stranger could verify without having an opinion about the output. None of them asks whether the work is any good, because that is Principle One, the one principle that isn’t negotiable. Each is set as a floor rather than a target, so there is no upside to exceeding it and no incentive to fake it.

It is meant to be dull. The set should read like a labour standard, from the same family as working-time rules and health and safety, rather than like a manifesto.

Condition 1: Hours

Discloses: Paid time actually available for the task, not tasks closed.
A floor looks like: A minimum share of contracted hours left for the core task once meetings, admin and supervising machine output are taken out, disclosed per role.
What it catches: Intensification hidden inside a productivity gain: the same people closing more tasks in the same week, with less time for any one of them.
For example: A hospital trust discloses how much of a consultant’s week is left for patients after AI-drafted notes have been checked and signed off.

Conditions, not merit. Floors, not targets.

Condition 2: Training

Discloses: Paid hours spent getting better at the craft, not at this quarter’s tool.
A floor looks like: A minimum number of paid hours per person per year on the underlying discipline, reported separately from tool onboarding.
What it catches: Deskilling: an organisation can train everyone on the new tool and nobody in the judgement the tool is meant to support.
For example: A law firm reports the hours its associates spent on legal reasoning and drafting, separately from the hours spent learning the contract-review platform.

Conditions, not merit. Floors, not targets.

Condition 3: Junior headcount

Discloses: Whether anyone is still being brought into the profession.
A floor looks like: A minimum ratio of entry-level hires and trainees to senior staff, disclosed year on year.
What it catches: The empty pipeline: AI eats the entry rungs first, and the damage only shows up a decade later, when nobody is ready to be senior.
For example: A bank discloses its graduate and apprentice intake in analyst roles against the number of senior analysts it employs.

Conditions, not merit. Floors, not targets.

Condition 4: Human authorship

Discloses: Who made it, disclosed but never assessed.
A floor looks like: Not a quota but a label: every output states whether a person made it, reworked machine output, or approved machine output unchanged.
What it catches: Responsibility laundering: work that carries a person’s name and sign-off but that no person materially shaped.
For example: A radiology department records whether each report was written by a radiologist or generated by a system and signed off.

Conditions, not merit. Floors, not targets.

Condition 5: Rework

Discloses: What was scrapped and redone, which is the real price of speed.
A floor looks like: Rework published next to any speed or productivity claim for the same period, so the saving can be read net of the redo.
What it catches: Speed that only exists on paper: the time saved at the start of the job, spent again at the end.
For example: A software team publishes how much generated code was reverted or rewritten within ninety days, alongside its delivery figures.
How to read it: Rework is the only condition that cannot be read on its own. A falling rework figure means either that the systems improved or that the people stopped checking, and those are opposite worlds. Rework must therefore be published alongside human authorship and exposure for the same period: rework down with authorship up is a real improvement; rework down with authorship down is automation bias with a productivity story attached. Read in isolation the figure can be quoted in either direction, which is why the core is a set and not a menu. Why this was added →

Conditions, not merit. Floors, not targets.

Condition 6: Exposure

Discloses: AI-mediated adverse events landing on people, by position and harm class.
A floor looks like: No person may receive an adverse AI-mediated decision without a named human with the authority to reverse it. Outcome counts are disclosed with their trend and never targeted.
What it catches: The thing none of the other conditions can count: harm that lands on a person.
For example: Hospital and bank resolve trivially: position three is the patient and the applicant, and adverse decisions are the obvious harm class.
Working definition v0.1: Three positions, a matrix of harm classes, and three open flags. Read the full definition →

Conditions, not merit. Floors, not targets.

Where these came from. The six conditions were first set out in Part I, inside the case for Principle One, and in version 0.2 they became the core of the standard. What changed, and why →

What they look like in one sector. The sector annexes work the core through a real P&L. Annex A, for creative, marketing and media, is written; the other five are open.

The list

Get new principles, revisions and the arguments that killed them, roughly once a month and never more often.