How to Spot Bias and Stereotypes in AI-Written Content

Spot bias in AI-written content by reading for who the text assumes the reader is. Check four things: default people (every customer "he", every planner "she"), stereotyped descriptions, who is missing from examples and images, and loaded wording about age, disability, money or background. Then run a swap test: change the group and see whether the sentence still works.

The reason this needs a deliberate check is that AI bias rarely looks like bias. Language models learn from vast amounts of existing writing, so they reproduce its most common patterns, including who usually gets described doing what. Mainstream assistants are tuned to avoid openly offensive output, so what gets through is quieter: a default, an assumption, an omission. Each sentence looks harmless on its own.

Follow me on Instagram@sagnikteaches

The main example is an illustrative print shop with six staff that uses AI to draft its website pages, social posts, customer emails and the occasional job advert. Along the way are examples from an events company, a software reseller, a clothing shop and a jewellery seller, because bias shows up differently in each.

Connect on LinkedInSagnik Bhattacharya

A word on scope. This covers the words and images you publish. If you use AI to make decisions about people (screening job applicants, setting prices, approving credit), the stakes are higher and the checks are different; the tutorial on AI bias in small business decisions covers that side.

Subscribe on YouTube@codingliquids

Six patterns bias takes in everyday AI drafts

Most of what you'll find falls into six patterns. The examples in this table are illustrative lines of the kind AI assistants produce for small businesses every day.

There is a seventh pattern that doesn't fit in a table because it is an absence: who is missing. Ten AI-generated example customers who are all young founders, or case-study ideas that are all about offices and none about trades, say something about who you think your customers are.

The swap test, with worked sentences

The fastest single check is to swap the group in a sentence and see whether it still reads as natural. If the swapped version sounds odd or insulting, the original was relying on a stereotype.

• Original: "Our jewellery makes the perfect gift for her birthday." Swap: "the perfect gift for his birthday". A jewellery seller whose rings sell to everyone just told half its buyers the shop isn't for them. Fix: "A birthday gift they'll wear every day", with the gift guide sorted by style and budget rather than by gender.
• Original: "Even technophobes over 60 will find our portal simple." Swap: "Even technophobes under 30 will find it simple." The swap sounds strange, which shows the original linked age to incompetence. Fix: "The portal has four buttons, and there's a two-minute video if you want a walkthrough."
• Original: "Our friendly female staff will help you choose." Swap: "Our friendly male staff". Staff sex is irrelevant to choosing a font. Fix: "Our staff will help you choose."
• Original: "Clothing for real women with curves." Swap: "Clothing for real women without curves." The swap exposes the problem: it implies some women aren't "real". Fix: for a clothing shop, "Sizes 6 to 26, with fit notes and measurements on every product."

The swap test doesn't catch everything, and it can over-flag. "Women's sizing runs small in this brand" swaps awkwardly but is product information, not bias. Use it as a prompt to think, then keep the specifics your product genuinely requires.

A print shop's reading routine for AI drafts

Here is the worked example. The print shop's owner noticed the problem when a customer, a woman who runs a building firm, replied to an AI-drafted email about trade flyers with one line: "I'm the builder, not his secretary." The draft had opened "Hi, could you pass this on to your husband about the site boards?" because the assistant had assumed from her business type who was in charge.

So the shop audited everything it had drafted with AI in the previous six months: 30 pieces including 12 web pages, 10 social posts, 5 email templates, 2 job adverts and a "who we help" page. The owner and one colleague read each piece with the six patterns and the swap test in mind. Illustrative results:

The first audit took about two hours for two people. After that, the shop added the check to its normal sign-off, which adds roughly two minutes per piece. The quick sum: at about 25 AI-drafted pieces a month, that's under an hour a month, which is cheaper than one lost trade customer.

The personas were the worst offender

The "who we help" page had been built from three AI-generated personas. The prompt asked for "three typical print shop customer personas", and the assistant returned:

Sample AI output (illustrative): 1. Busy mum, 35, runs a craft business from her kitchen table and needs affordable stickers. 2. Tech-savvy founder, 28, launching a start-up and ordering sleek business cards. 3. Traditional sole trader, 62, who struggles with technology and prefers to order in person.

Every persona was built on a demographic stereotype, and none reflected the shop's real customers, who were mostly trades, local charities, event organisers and schools. The fix was to write personas from order data instead: what people buy, how often, what they ask. "Trade customer ordering 500 site-board flyers and vehicle magnets twice a year, needs proofs by phone photo" is more useful than any age and gender, and it can't be stereotyped because it describes behaviour.

Prompts that ask the AI to audit its own assumptions

Assistants are reasonably good at spotting these patterns when asked to audit rather than write. It won't replace a human read, but it gives you a list to check, and it's quick. The print shop now runs this on anything going to more than a handful of customers:

Review the text below for bias and stereotypes. Do not rewrite it yet.
Our customers include people of all ages, genders, abilities, backgrounds,
family set-ups and incomes, and businesses of every kind.
List in a table: Exact phrase | Assumption it makes | Who it might put off |
Suggested neutral wording.
Look specifically for: default genders or roles; assumptions about age,
disability, money, family, religion or culture; examples or names that all
come from one group; anything that treats a group as a joke or a problem.
If you find nothing, say so; don't invent issues.
Text:
[paste draft]

An illustrative result for an events company's wedding FAQ:

What it missed is worth noting. All six example names in the FAQ's sample timeline came from one culture, which the audit didn't mention even though the prompt asked about names. And it flagged "stairs to the mezzanine" as a potential issue before concluding it was fine, which is the right conclusion but shows how these audits over-flag. The human reader decides.

Stop the defaults at the source with saved instructions

Audits catch problems after the draft. Rules in the assistant's saved instructions, such as a ChatGPT or Claude Project or a Gemini Gem (becoming a skill from November 2026), stop many of them appearing at all. These are the six lines the print shop added:

- Refer to customers and staff as "you", "they" or by role, never a default he or she.
- Don't assume a customer's age, gender, family, religion, income or ability.
- When giving example names, mix cultures, or use [first name].
- Offer an alternative whenever you mention seeing, visiting, calling or clicking.
- Don't make jokes about any group of people.
- Keep product facts such as sizes, access and compatibility exactly as given.

An illustrative test from a software reseller shows the difference. Before the rules, a support reply draft read: "When your office manager calls, she should have her login ready, and if she gets stuck, her IT guy can reset it." After the rules went into the project, the same request produced: "When you call, have your login ready. If you get stuck, whoever manages your Microsoft 365 account can reset it, or we can do it with you on the phone." The second version is also clearer, which is the usual side effect of removing assumptions.

Rules reduce the work; they don't remove the read. In the print shop's first month with the rules in place, the reading routine found 4 issues across 25 new pieces, down from 19 across 30 before. All four were in content where someone had pasted wording from an old template that the rules never saw, which is a reminder to check the templates themselves, not only the fresh drafts.

Images and alt text carry stereotypes too

Text isn't the only place. AI image generators have strong defaults: ask for "a business owner" and you'll often get a middle-aged man in a suit; ask for "a happy customer" and you'll get someone young and slim. The print shop's first attempt at AI-generated images for its "our customers" page produced six people who looked like they worked at the same design agency, and none who looked like the builders, teachers and charity volunteers who actually order from it.

The fixes are practical:

• Be specific in image prompts about the job and setting ("a builder in hi-vis holding a printed site board outside a half-finished house"), and vary age and appearance deliberately across a set.
• Use real customer photos where you can, with written permission. They are more convincing than any generated image and represent your actual customers.
• Don't let alt text guess about people. AI-written image descriptions often add age, ethnicity or gender the tool inferred. The tutorial on generating and checking AI alt text explains why to leave those out unless they matter and are known.
• Write image rules into your brand guidelines, alongside colours and style. The tutorial on brand guidelines for AI images shows where they fit.

Where fixing bias goes wrong

Well-meant edits can create new problems. These are the realistic ones:

• Inventing representation. A clothing shop asked its assistant to "make the testimonials more diverse", and it rewrote real reviews with new names and details. That turns genuine reviews into fabricated ones, which can breach consumer and advertising rules. Never edit what customers said; ask more kinds of customers for reviews. The tutorial on turning customer reviews into marketing copy covers how to use real reviews properly.
• Tokenism. One diverse stock image among twenty identical ones, or a single line about accessibility at the bottom of a page, reads as box-ticking. Representation should be spread through the content, not bolted on.
• Stripping useful specifics. A software reseller's AI edit removed "screen-reader compatible" from a product description because the audit prompt flagged "disability-related terms". That was the most useful line on the page for some buyers. Neutral wording isn't vague wording.
• Clumsy language. "Individuals of all ages, genders and backgrounds are welcome to order business cards" is inclusive and unreadable. Usually the fix is simpler: remove the assumption and write naturally.
• Over-correcting a real quote. Customer quotes and testimonials stay as the customer said them, even when their wording isn't how you'd put it.

Job adverts and HR text need a stricter check

Everything above is about tone and fairness in marketing. In job adverts and staff documents, biased wording can also be unlawful: in many places, wording that signals a preference for a particular age, sex or other protected characteristic can breach equality or anti-discrimination law. Check with an HR or employment adviser if you're unsure. The patterns AI tends to produce in job adverts:

• Age-coded words: "young, energetic team", "digital native", "recent graduate" (unless the role genuinely is a graduate scheme).
• Physical requirements stated as defaults: "must be able to stand all day" when the real need is "the role involves operating a guillotine at a standing workstation for most of the shift". Describe the task, not the body.
• Culture words that exclude: "work hard, play hard", "beers on Friday", which suggest who will fit in.

An illustrative before and after from the print shop's finishing assistant advert:

Before (AI draft): Join our young, dynamic team! We're looking for an energetic go-getter who can keep up with a fast-moving print floor and join in with our after-work drinks.

After (edited): We're hiring a finishing assistant to cut, fold and pack print orders. The work is mostly standing at a workstation, lifting boxes up to 12kg, on weekday shifts from 8am to 4pm. We'll train you on the machines. Tell us if you need any adjustments to apply or to do the job.

The after version tells applicants what the job involves, which is fairer and also gets better applications. The tutorial on writing job adverts with ChatGPT without biased language goes further on adverts specifically.

A one-page bias check to keep beside your style guide

This is the print shop's check, filled in with the notes it uses. Keep it short enough that people actually use it.

1. People: are customers, staff and roles described without a default gender? (Use "they", "the couple", "whoever looks after...")
2. Age: does anything treat age as a joke, a limit or a selling point it doesn't need to be?
3. Ability: can someone who can't see, visit, hear or use a mouse still do what the text asks? Is there another way in?
4. Money and family: does it assume a car, a partner, children, a holiday or a budget?
5. Culture: does it assume one religion, one set of holidays, one naming order, one kind of name?
6. Who's missing: do the examples, images and case studies reflect the customers you actually have?
7. Swap test: for any sentence about a group, does the swapped version still read naturally?
8. Specifics kept: have we kept the product facts (sizes, access, compatibility) that some customers need?

The owner's rule for the team: if two people disagree about whether something is biased, rewrite it in plainer words. The plainer version is almost always better anyway. For content that goes through several hands, add this check to your AI content approval workflow so it happens at sign-off rather than after a customer points it out.

Questions about bias in AI drafts

Is it bias to write copy aimed at one group, such as women's clothing?

No. Describing who a product is made for is product information. Bias is an assumption the copy didn't need, such as implying only women buy gifts or that older customers can't use a website. Keep the specifics your product requires and remove the assumptions it doesn't.

Can I ask the AI to make my testimonials more diverse?

Never invent or alter testimonials, for diversity or any other reason. Fake or edited reviews can breach consumer and advertising rules and destroy trust if discovered. If your real testimonials all come from one kind of customer, ask a wider range of customers for reviews and use what they actually say.

Do AI tools have settings that remove bias?

Not in a way you can rely on for your own content. Mainstream assistants are tuned to avoid openly offensive output, but the quieter defaults described here still appear. Instructions in your prompt and saved project help, and a human read with the swap test remains the check that catches what the rules miss.

How often should we check older content?

Check high-traffic pages and anything customer-facing that AI drafted once a year, and whenever you update your style guide. Job adverts and HR documents should be checked every time they're reused, because the stakes are higher.

Further reads

• Is AI CV Screening Fair? Bias Checks Every Recruiter Should Run — Bias checks when AI helps screen job applications.
• How to Write an AI Content Policy for Your Marketing — Put your inclusive-language rules into a written AI content policy.
• Tone of Voice Examples: AI Copy Before and After Editing — Before-and-after edits that fix tone at the same time.
• How to Write a House Style Guide Your Team and AI Both Follow — Add the bias check to a style guide your team and AI follow.
• How to Set Up Human Review for AI Work Without Slowing Down — Set up human review for AI work without slowing everything down.
• How to Write Fairer Performance Reviews With AI — Keep bias out of AI-assisted staff reviews.
• AI Ethics for Small Businesses: A Practical Checklist — Twenty-five questions in six groups, each with why it matters and how to check, plus red lines and a nursery example run end to end.
• How Coaches Use AI to Build Workbooks, Exercises and Programmes — Turn your coaching method into a programme and workbook with AI doing first drafts, while the method, the edits and the claims stay firmly yours.
• How Estate Agents Use AI to Write Property Descriptions — Build the fact sheet at the valuation, prompt in your house style, get portal, brochure and social copy from one input, and check every claim before listing.
• How Tutors Should Check AI-Made Worksheets Before Using Them — The ten-minute, four-pass check that catches wrong answer keys, two-answer questions, off-level reading and garbled diagrams before a pupil sees them.
• How to Check AI-Generated Images for Errors Before Posting — An eight-point checklist and four viewing tricks for catching AI image mistakes, with a filled-in check and a campaign of eight images put through it.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: examples and checks are illustrative, based on common patterns in AI-drafted business copy. For employment-related wording, readers should confirm requirements with an HR or employment adviser.