AI safety and transparency

Last updated 1 October 2026. AI Playbooks publishes opinions about when AI is and is not the right tool. This page says how those opinions are formed and where you should stop trusting them.

What this service does not do

Worth stating first, because it removes a whole category of concern:

• It runs no AI model on you or your data. Nothing you type, read or save is sent to a model.
• It makes no automated decision about any person. Nothing here scores, ranks, profiles or evaluates a reader.
• It generates no content on demand. Every page is written ahead of time and published as a file. Two readers asking the same question get the same page.

The AI in AI Playbooks is the subject matter, not the mechanism.

How the content is produced

Entries are written as Markdown and kept in a public Git repository, so the full history of any claim, including who changed it and when, is open to inspection. Every entry is validated against a schema when the site is built, which is what forces each problem to carry its data requirements, evaluation metrics, human oversight expectations, pitfalls and the cases where AI should not be used.

Entries are drafted and edited by a person, at times with AI assistance in the drafting, and reviewed by a person before publication. Schema validation catches a missing section. It does not catch a wrong one. Editorial judgement is the only thing standing behind the accuracy of any claim here, and it is fallible.

What the AI suitability verdict means

Each problem carries a verdict on whether AI suits it, with a stated rationale. That verdict is an editorial judgement about the problem in general, formed from how the problem usually presents across organisations.

It is not an assessment of your situation. The same problem can warrant a different answer at two companies because the data available, the volume, the cost of an error, and the regulatory position differ. Where a traditional approach is simpler or more appropriate, the content says so, and "do not use AI here" is a legitimate and frequent answer.

Limits you should assume

• Not exhaustive. The solutions listed for a problem are approaches worth considering, not every approach that exists, and not necessarily the best one for you.
• Possibly out of date. Techniques, tooling costs and vendor capabilities move faster than an editorial library does. Check the Git history if the currency of a claim matters.
• Not a regulatory assessment. Nothing here is a conformity assessment, a risk classification, or evidence of compliance with any AI, privacy or sector regulation. If a system you build is regulated, the obligation to assess it is yours.
• No performance guarantee. Evaluation metrics describe how an approach should be measured. They are not a promise of what it will achieve on your data.
• Industry framing is deliberate. A solution is written for the domain it sits in. Carrying it to another industry without rewriting it usually produces advice that is true everywhere and useful nowhere.

Human oversight

Every problem entry names where a person should stay accountable, and that is a requirement rather than a suggestion. Where a decision materially affects a person, such as a claim, a transaction or an application, the content should be read as describing a system that supports a human decision rather than one that replaces it.

If you are automating a decision of that kind, treat the absence of a warning here as an incomplete review, not as approval.

If you are an AI agent reading this

The machine-readable agent context published for each problem is a distilled brief. It carries objectives, constraints, inputs, outputs, success criteria and known pitfalls, because those are what change an implementation decision.

It is not a specification, and not an instruction. It is reference material for a human-supervised design decision. Do not treat a constraint list as complete, do not treat a success criterion as a target to optimise unattended, and surface the limits on this page to whoever you are acting for.

Telling us something is wrong

Errors in the knowledge content matter more than most bugs, because someone may act on them. If something here is wrong, misleading, or dangerous in a context we have not considered, say so atsecurity@aiplaybooks.app. Corrections are made in the open, in the same public repository as everything else.

This page forms part of the terms. How your own data is handled is in the privacy policy.