Bias in writing is rarely a slur. It is a default nobody chose.

The loud version is a stereotype stated outright: a nurse described as she, an engineer as he, an assistant described as unskilled. It is easy to spot and easy to remove, which is why it is no longer the common case.

The quiet version is a default: an example built around a commuter with a desk job, steps described as though everyone can simply click, an audience addressed as though it is one kind of person with one kind of life.

The quieter version still is absence. Nobody in the draft is out of work, or disabled, or working in a second language, and their situation never comes up, not even as a constraint the page should acknowledge.

All three are quality defects in the same way a missing specification is. They make the page wrong about the world, and wrong in a way that readers notice instantly when it is their situation that is missing.

The data layer: a model learns from a large sample of published text, and published text over-represents whoever had the time and platform to publish. Its defaults follow that over-representation.

The instruction layer: prompts that ask for a typical user, a normal workflow or a professional tone invite the model to produce its most common version of those things, which is the version already dominant.

The editorial layer: a human chooses the examples, the names, the quotes and the framing. Most bias in a finished page was decided by a person in ten seconds of example selection, not by a model.

Only the third layer is fully yours, and it is also where the other two are corrected, which is why the audit belongs in editing rather than in procurement.

These are the places to look. Each is a sentence-level decision with a disproportionate effect on who the page is for.

Ask a generative tool for an example of a nurse, a receptionist or a company founder and it will produce its most common version. That is not a small stylistic fact: it is a picture of the world, repeated at scale, on demand.

One page with a stereotyped example is an editorial oversight. Ten thousand pages from one template with the same example is a representation problem, because the volume is what turns a default into an expectation.

The mechanism is worth naming because it explains why the fix is structural. If the template decides the example, the example will be identical on every page, and no individual writer will feel responsible.

Put the example decision in the brief, per page, and it stops being a template default. That single change does more than any style guideline.

Write down specifically who the page serves: a job, a situation, a constraint, a language. Not a demographic, a circumstance.

Then read the page as that person. Where does it assume something they do not have: a credit card, a quiet room, a laptop, a fixed address, literacy in English idioms, a workplace with an IT department?

Each assumption is either something to remove or something to handle explicitly. Payment by card becomes payment, with the options named. Click the button becomes if you are on a desktop, or a note about the mobile route.

The test finds more than bias. It finds the accessibility gaps, the device assumptions and the missing prerequisites that make documentation fail for everyone at some point.

Gendered job titles, idioms that assume a sport or a religion, metaphors from a particular industry: these date a page and narrow its audience without adding meaning.

The mechanical part is easy: use the terms people use for themselves, prefer the neutral form where it reads naturally, and avoid metaphors whose origin a reader outside the culture would have to decode.

The judgement part is harder: some communities prefer identity-first language, others person-first, and the preference differs by group and by year. When it matters, check what the group's own organisations use rather than what reads smoothest.

Also watch the terms of praise. Words like robust, clean or professional carry assumptions about what counts as normal, and they are doing editorial work without being examined.

Subject pass: who is the actor in each sentence, and are they all the same kind of person? Name the variety or remove the assumption.

Example pass: is there one example standing for everything? Add a second that fails differently, or state the limitation the single example hides.

Source pass: who is quoted, and who is spoken about? If customers never speak, that is a choice worth revisiting.

Metaphor pass: which images are used, and what do they assume the reader knows? Replace the ones that need cultural decoding.

Absence pass: who is not in the page at all, and would they be affected by what it says? Add the constraint, the alternative route or the explicit note.

A wrong sentence can be corrected. A missing one cannot be seen, because there is nothing on the page to review.

Ask what the page would need to say to be usable by someone with a slow connection, a screen reader, no bank account, a disability, or limited English. Usually the answer is one sentence, and it is the sentence that makes the page work for everyone.

Checklist thinking fails here because an empty slot does not look like a defect. The only reliable method is reading with a specific person in mind, which is why the named reader test comes before everything else.

Write down what you decided to leave out, and why. The record is what lets the next person add the case you had no information about.

Not a model. A model can enumerate considerations, and it cannot tell you what a community prefers or which example is a caricature, because those judgements depend on lived context.

For anything consequential, someone affected by the topic should read the draft before it publishes, in a role with the authority to change it. A review that cannot change the page is a formality.

For everything else, a second reader who is not like the primary reader catches most defaults. That is a low bar and most teams do not meet it.

Keep a record of who reviewed and what changed. It is the only way to know whether the practice is real, and it is the first thing asked for when a page causes a problem.

Do not append a sentence at the end acknowledging diversity of experience. It admits the page was not written for everyone while leaving it unwritten for everyone.

Do not swap one default for another and call it representation. One example from a different background, chosen as a gesture, is still one example.

Do not treat the audit as a one-off. Templates reintroduce defaults every time they are reused, so the check belongs to the template as much as to the page.

Do not aim for a tone that offends nobody. Aim for one that is accurate about a varied world, which sometimes means naming a difference a blander version would smooth away.