AI and Editorial Transparency Policy

Where AI is used

AI tools assist with drafting explanatory prose, suggesting candidate distractors for practice questions, and writing the code that runs this site. We are telling you that rather than hiding it, because a study site quietly publishing unreviewed machine output is precisely the thing a candidate should be wary of.

Disclosure on its own is not much of a commitment, though. What matters is where the machine is not allowed to decide anything, and what has to happen before something reaches you.

Where AI decides nothing

• No answer key is accepted from a model. For Table Reading, Block Counting and Instrument Comprehension the key is computed from the underlying data that draws the figure, so there is nothing for a model to assert. Everywhere else a person verifies it.
• No factual claim about the exam is published on a model’s word. Subtest counts, timings, composite structure and policy statements are checked against a registered source with a verification date, or they are withheld.
• No author, reviewer or credential is generated. Our schema layer throws an error if any page tries to emit Person markup, because no verified contributor exists yet — see authors and reviewers.
• No citation is accepted unchecked. If we cannot point at a source that exists, the claim is marked uncertain rather than dressed up.
• Nothing publishes automatically. Every question requires a named technical reviewer and a named editorial reviewer, enforced by a function in the code rather than by process discipline.

Why this matters more here than elsewhere

Language models are fluent and confident about details they have partly invented. Ask one how many questions are on a subtest, or what the minimum composite is for a career field, and you will often get a specific, well-formatted, wrong answer — delivered with exactly the same tone as a correct one.

On most websites that produces an embarrassing error. On a study site it produces a candidate who revises the wrong material, or who decides when to sit a commissioning exam based on a figure nobody checked. That is the scenario this policy is built around, and it is why our retake page publishes no attempt limits at all rather than a plausible one.

It is also why the validation runs automatically. Human review does not scale reliably across a growing bank, so the structural checks — key matches an option, explanation is substantive, tags present, computed items agree with their own geometry — are enforced in continuous integration where they cannot be skipped on a busy day.

What we record

Every question carries an AI-assistance record alongside its author, both reviewers, its validation status and an immutable version history. Every consequential page records how it was made in the Who / How / Why block at its foot, including whether AI assisted in drafting.

That record is not decorative. If a systematic error is ever traced to a generation method, we need to know which items shared it in order to withdraw them together rather than one complaint at a time.

What we guarantee

• No practice question is served without human technical and editorial approval.
• Every question has a worked explanation; thin or circular explanations are rejected automatically.
• Every factual claim about the exam traces to a registered source with a verification date.
• No invented people, credentials, testimonials, user counts or statistics appear anywhere.
• Errors are corrected publicly on the corrections log.

Where the line sits, concretely

Abstract commitments are easy, so here is the same policy expressed as specific decisions on this site.

The subtest timings on our format page were not written from a model’s recollection. They were read off the delivery provider’s page on a recorded date, and an earlier version of this site had four of them wrong — which is logged on the corrections page precisely because that is the failure this policy exists to catch.

The retake page publishes no attempt limit, despite a model being perfectly willing to supply one. The governing document is unverified in our register, so the figure is withheld.

The authors page lists nobody, despite it being trivial to generate a plausible reviewer with a plausible background. The schema layer throws rather than emit Person markup for a placeholder.

Block Counting answers are computed from geometry rather than asserted, so the correctness of those items does not depend on anything a model claimed.

Each of those cost us something — a thinner page, a missing figure, an empty bank. That is what the line actually looks like in practice, as opposed to in a policy document.

Reporting AI-shaped errors

Confident but wrong text is the characteristic failure, and readers spot it faster than authors do. If an explanation reads plausibly but does not actually follow — the steps do not connect, or the reasoning would justify a different answer — please tell us. We treat that as seriously as a wrong answer key, because it usually indicates a systematic problem rather than a one-off.

Our wider standards are in the editorial policy, and the question pipeline is documented in full on the question methodology page.