Lesson 2 · AI Governance Basics
Responsible AI principles and accountability
- Length
- 26 minutes across 7 sections
- You will be able to apply
- The Refusal Test
- You will produce
- Commit Statement
- You will work
- 3 gated questions
Personalize the practice
Apply this to your environment
These details adapt the application prompts and coach questions. They do not affect your score.
Name the last thing this principle stopped you doing.
Core question
find the flaw
Spot the justification that really survives the Refusal Test
Required practiceA stakeholder is drafting OmniCorp's principles statement and wants to claim one principle is truly operative. Choose the statement that would still stand up once you ask what the principle actually refused.
Almost every organisation now publishes AI principles. Fairness, transparency, human oversight, accountability, safety. The words are broadly the same everywhere, which should be the first clue that they are doing very little work. A principle that no organisation would decline to sign cannot distinguish between organisations, and cannot distinguish between decisions either.
This lesson is not an argument against principles. It is an argument that a principle is not a statement of value: it is a record of things refused. Until it has refused something, it is a description of how you would like to be seen.
The pathology: Principle Drift
Principles are agreed at a moment of low pressure, usually in a workshop, by people who are not currently trying to ship anything. They are then applied at moments of high pressure, by people who are. The gap between those two moments is where the drift happens, and it happens in a specific way: the principle is not abandoned, it is reinterpreted until the decision in front of you satisfies it.
Drift cannot be prevented by writing the principles more carefully. It is prevented by attaching each principle to a decision it has actually changed, and reviewing whether it has changed one recently.
The commitment is written at a moment of low pressure by people not currently shipping.
- Principle agreed
- The commitment is written at a moment of low pressure by people not currently shipping.
- Pressure arrives
- A decision is needed and the principle appears to stand in the way of progress.
- Reinterpretation
- The principle is read more broadly until the decision in front of you satisfies it.
- Constraint evaporates
- The words survive intact but the practical obligation has disappeared.
The Refusal Test
Definition
A principle is operative only if you can name a specific thing it caused you to refuse, when the refusal happened, and who bore the cost of it. Three elements, all concrete. A principle with no refusal on record is not yet a principle: it is an aspiration, and it should be labelled as one so nobody relies on it.
The cost element is the one that matters. A refusal that cost nothing was not a refusal, it was a preference that happened to align with the easy path. Principles are only visible in the cases where they were expensive.
When to Use It
Run it annually against every published principle, and immediately after any decision that felt uncomfortable. The uncomfortable decisions are where the record comes from: if you cannot remember being uncomfortable, you have no refusals to record, which is itself the finding.
Run it also when adopting a new principle. Before publishing it, ask what you would have to stop doing today if it were true. If the answer is nothing, do not publish it.
How to Apply It
- Take each principle and name one decision it changed in the last twelve months, with a date.
- State what the refusal cost, in revenue, delay, effort or someone's goodwill, and who absorbed it.
- Mark any principle with no refusal as aspirational, in writing, where the same audience can see it.
- For each aspirational principle, either name the decision that would test it next quarter or remove it.
| Principle | Operative when | Aspirational when |
|---|---|---|
| Human oversight | A launch was delayed because no reviewer could be resourced | A reviewer is named in the design document |
| Fairness | A feature was cut after a subgroup performed worse | Fairness metrics are collected and reported |
| Transparency | A vendor was declined for refusing to document its logic | Model documentation is available on request |
| Safety | A profitable use case was withdrawn after a near miss | An incident process is documented |
Worked example 1 of 3
OmniCorp Health publishes five AI principles. Dr. Naomi Ellery ran the refusal test across all five in a single ninety-minute session with the clinical AI group.
- Dr. Naomi Ellery
- Human oversight. What has it stopped us doing?
- Clinical AI lead
- We require clinician sign-off on every AI-drafted letter.
- Dr. Naomi Ellery
- That is a control we designed in, not a refusal. What did we decline, and what did it cost?
- Clinical AI lead
- We turned down the auto-send option on the referral tool. It would have saved about forty clinician hours a month.
- Dr. Naomi Ellery
- That is a refusal. Date, cost, and who absorbed it. Now do transparency.
- Clinical AI lead
- Nothing. We have never declined anything on transparency grounds.
- Dr. Naomi Ellery
- Then it is aspirational, and we say so, in the same document, in the same size type.
Two of the five principles had refusals on record. Three did not. Marking those three as aspirational was uncomfortable and took a further two weeks to agree. Within a quarter, one of them had produced its first refusal: a vendor was declined for treating its ranking logic as confidential. The label had done what the principle alone had not.
Why This Works
Refusals are checkable and intentions are not. Two people can hold irreconcilable views about whether an organisation is committed to fairness and both be arguing honestly. Nobody can argue about whether a launch was delayed in March. Anchoring principles to events converts a values debate into a records question, and records questions have answers.
Labelling the unrefused principles as aspirational works because it removes the reward for publishing them. Once a principle has to declare its own emptiness, the cheapest way to fix the label is to start applying the principle.
Worked example 2 of 3Optional depth
Trevor Okafor at OmniCorp Retail tested a fairness principle against a recommendation system. The refusal on record was real: a personalisation feature had been cut eight months earlier because it recommended lower-value products to customers in three postcodes. The cost was a projected margin improvement that never arrived, absorbed by the merchandising team's own targets. The principle was operative. What Trevor found next was the subtler problem: the refusal had never been communicated, so three teams were rebuilding equivalent features in ignorance. An operative principle that is not visible produces the same outcome as an aspirational one, more slowly.
Worked example 3 of 3Optional depth
Jo Halvorsen at OmniCorp Studio has no published principles and did not want any. Asked what she had refused, she named it immediately: she had declined a retainer from a client who wanted AI-generated case studies presented as staff-written, and the refusal cost her about a fifth of that year's projected revenue. That is one operative principle. She wrote it as a single sentence, on the studio's internal page, with the date and the cost. One sentence with a refusal behind it outperforms a page of five without.
Edge Cases and NuancesOptional depth
A new principle has no refusals yet and should not be penalised for it, provided it is dated and a testing decision is named. A refusal can also be recorded dishonestly: declining something you were never going to do, then filing it as evidence, is drift wearing the test's own clothes, and the guard is the cost element, since a fabricated refusal has no cost to name. And a principle can be operative in one function and aspirational in another, which is worth recording separately rather than averaging, because the average conceals exactly the part that needs attention.
A specific refusal with a date, a cost in recognisable units, and a named person who absorbed it.
- Operative
- A specific refusal with a date, a cost in recognisable units, and a named person who absorbed it.
- Tested once
- A single refusal exists but has not yet been repeated or communicated beyond the deciding team.
- Aspirational
- The principle is published but has never refused anything or cost anyone anything.
Knowledge check
A team cites a human-oversight control as evidence that their oversight principle is operative. Is this sufficient?
Common Failure Modes
The test end to end
Alan Brixmoor at OmniCorp Public applied the refusal test to a four-principle set that had been published for two years and cited in every AI business case.
Accountability produced a refusal immediately: a workflow had been held for eleven weeks in the previous autumn because no directorate would accept the named-owner role, and the cost was eleven weeks of delayed benefit, absorbed by the sponsoring directorate. Human oversight produced one too, though a weaker one: a proposed automation threshold had been lowered, costing throughput. Explainability produced nothing in two years. Proportionality produced something that looked like a refusal and was not, on examination: a low-risk use had been exempted from review, which is the principle permitting rather than refusing.
Two operative, one aspirational, one miscounted. Alan published all four states with the dates and the costs, in the same document as the principles themselves. The uncomfortable line was explainability, which the organisation had been citing to citizens for two years. Within a quarter it had its first refusal: a supplier was declined for treating its scoring logic as commercially confidential, and the cost was a four-month procurement restart. Alan's view afterwards was that the label had done the work, because nobody wanted to defend the word aspirational in a public document twice.
Decision point
Marisa Delgado at OmniCorp Financial reviews the annual AI principles statement. Every principle has a paragraph describing the controls that implement it. Asked for refusals, her team produces one genuine case for safety and, for the other four, descriptions of controls. The statement is due to be published in nine days and last year's version was well received. What do you do?
Self-check
Mark the level that describes you today. Nothing is submitted.
| Behaviour | Ready | Developing | Not yet |
|---|---|---|---|
| Distinguishing refusal from control | |||
| Handling an unrefused principle | |||
| Recording cost |
Commit
Commit Statement
Complete every line in your own words, then sign and date it. Attach your refusal list, including the blank lines.
| Window | Field application |
|---|---|
| Days 1 to 7 | Write the refusal, date and cost against every principle you are subject to, and leave the empty ones empty. |
| Days 8 to 21 | Get the aspirational ones labelled where the same audience can see them, and name the decision that will test each. |
| Days 22 to 30 | Find one refusal that was made and never communicated, and publish it so nobody rebuilds what was declined. |
The refusal test tells you which of your principles are real. It does not tell you how to make a contested refusal stick when the person proposing the work outranks the person raising the objection, how to record refusals so they survive a change of leadership, or how to demonstrate to an external reviewer that the record is complete. Building principles into enforceable decision rights is the capability the paid programs develop next.
Learner feedback