Give ChatGPT the job's real requirements first, tell it which kinds of biased wording to avoid, then run a separate audit prompt on its draft in a fresh chat and check the result with a second tool or a colleague. ChatGPT can introduce biased phrases itself, so the audit step matters more than the drafting prompt.
Most bias in small-business adverts isn't a stray adjective. It's inflated requirements: a degree for a job that doesn't need one, "five years' experience" where two would do, a driving licence for a role that never leaves the building. Those cut out more capable people than any single word. Word choice still matters, but treat word lists as a prompt for judgement. The best-known research, a 2011 study by Gaucher, Friesen and Kay, found masculine-coded wording made jobs less appealing to women because they expected to belong less. A 2024 replication found that effect in start-ups but not in established firms. The lesson is to fix wording that plainly signals "not for you", not to chase a perfect score.
Five kinds of wording that quietly narrow the field
| Type | Examples | Why it narrows | Say instead |
|---|---|---|---|
| Age-coded | "young, energetic team", "digital native", "recent graduate" | Signals a preferred age, and can be evidence in a discrimination claim | "a small team that moves quickly", "confident with spreadsheets and email" |
| Gender-coded | "rockstar", "ninja", "dominant", "crush your targets" | Tends to put off applicants who don't picture themselves in that culture | "you'll own a sales target and we'll back you to hit it" |
| Unneeded physical requirements | "must be able to stand all day", "must lift 25kg" | Excludes disabled applicants when the task is occasional or adjustable | "the role involves lifting boxes up to 25kg a few times a day; tell us if you'd need adjustments" |
| Culture and class markers | "native speaker", "polished", "good cultural fit" | Vague standards that filter on background, not ability | "fluent written English: you'll write to clients daily" |
| Inflated requirements | "degree required", "10+ years", "clean licence" when not needed | Screens out people who could do the job well | The skill itself: "has run events for 200+ guests" |
Pronouns belong here too. Write "you" throughout; it's warmer and avoids "he or she" entirely.
What ChatGPT produces when you don't give it rules
Ask a general assistant for "a job advert for an events coordinator" and you'll typically get something like this (illustrative, abridged):
"Are you a young, dynamic rockstar who thrives under pressure? We're a work-hard, play-hard events company looking for a superstar coordinator to join our energetic team. The ideal candidate will have a degree in events management, 5+ years' experience and a clean driving licence. He or she will be a natural leader who can dominate a busy schedule..."
Count the problems: "young", "rockstar", "superstar", "work-hard, play-hard" (which reads as long unpaid evenings and after-work drinking), a degree nobody checked was needed, "5+ years" chosen by the model rather than the business, a driving licence the job may not need, "he or she", "dominate". None of it came from the employer. Models reproduce the patterns in the adverts they learned from, and a lot of those adverts were written this way. That's why the method below starts with facts and ends with an audit, rather than trusting a better first prompt.
Step 1: a requirements sheet before any prompting
Write the job's requirements yourself, and test each one with a single question: would you reject an otherwise excellent applicant who lacked this? If not, it's a nice-to-have or it goes. Filled in for a machine operator at a print shop:
Role: Digital print machine operator, full time, 37.5 hrs,
day shifts Mon-Fri
Pay: [your range], plus overtime at time and a half
Must-haves (would reject without):
- Can follow a job ticket and check print against a proof
- Careful with colour and alignment (we'll test with a sample)
- Can lift paper boxes up to 20kg a few times a shift
(adjustments possible)
Trainable (we'll teach):
- Our specific presses and finishing kit
- Our job management system
Removed after the test:
- "Degree or equivalent" (nobody here has one; not needed)
- "3 years' print experience" (two of our best operators
started with none)
- "Driving licence" (deliveries are done by a courier)The "removed" list is worth keeping. Next time someone copies last year's advert, it shows why those lines went.
Sometimes the test says keep it. A domestic cleaning firm whose clients are spread across villages with a bus or two a day genuinely can't use cleaners who don't drive, so "driving licence and access to a car" survives the question. Keep it, but state the reason and what you cover, so it reads as a fact of the job rather than a filter: "You'll drive between three or four clients' homes a day, 5 to 15 minutes apart, and we pay mileage at [rate]." People without a car rule themselves out knowingly, and people with one know the running costs are covered. Mark requirements like this "confirmed" on the sheet so the audit in Step 3 leaves them alone.
Step 2: the drafting prompt
Write a job advert from the requirements sheet below.
Rules:
- Address the reader as "you". No "he/she", no "the ideal
candidate".
- Use only the requirements given. Do not add qualifications,
years of experience, licences or personality traits.
- Avoid age-coded words (young, energetic, digital native,
recent graduate), hype words (rockstar, ninja, superstar),
and aggressive words (dominate, crush, killer).
- Describe physical tasks as the task and how often, and say
adjustments are possible.
- Include the pay range exactly as written.
- Plain English, short sentences, under 350 words.
- Structure: what you'll do; what you need; what we'll teach
you; pay and hours; how to apply.
Requirements sheet: [paste]Even with those rules, read the first draft for additions. An illustrative extract from the print operator draft:
You'll join a fast-paced production team with a passion for
print. You'll need a keen eye for detail, a can-do attitude
and the ability to work under pressure. Previous experience
on digital presses is an advantage.Three things crept in that the sheet never asked for. "A passion for print" and "a can-do attitude" are personality traits, which the rules banned; the model added them anyway because nearly every advert it learned from has them. "Previous experience on digital presses is an advantage" quietly reverses a decision on the sheet, where the presses were listed as trainable, and it will put off exactly the career changers the owner wanted. "Fast-paced" and "under pressure" may be true, but say what they mean: "most jobs go out within 24 hours, and on a busy day you'll run 40 short jobs". Pasting the sheet's "removed" list into the prompt under "do not reintroduce" usually stops the reversals.
Step 3: the audit prompt, in a fresh chat
Run the audit in a new conversation so the model isn't defending its own draft. Paste the advert after any edits you've made, since human edits reintroduce bias too.
Audit this job advert for wording that could discourage
qualified people from applying. Check for: age-coded words;
gender-coded words; physical or health requirements stated
more strictly than the task needs; culture, class or language
markers; requirements that may be inflated.
For each issue: quote the phrase, name the type, explain in one
line, suggest a replacement.
Then list any phrase you considered but decided is fine, and why.
Do not rewrite the whole advert.
Advert: [paste]An illustrative audit of a draft advert for an events coordinator:
1. "join our lively team" - possibly age-coded, mild. Suggest
"join a small, friendly team".
2. "must be available evenings and weekends" - may exclude
carers; stated more strictly than needed if only some
events run late. Suggest "most events run 9-6; around one
weekend event a month, with time off in lieu".
3. "strong leader" - masculine-coded in some word lists.
Suggest "you'll lead suppliers on the day".
Considered and fine: "lead time" (scheduling term, not
leadership); "competitive salary" (vague, but not biased;
better replaced by the actual range).What you'd do with it: accept item 2 without hesitation, because it's the change most likely to widen the field; carers read "evenings and weekends" as a closed door. Take item 1. Item 3 is a judgement call: "lead suppliers on the day" is clearer anyway, so change it for clarity rather than coding. And add the real pay range, which the audit rightly nudged. The "considered and fine" section is there to stop the model over-flagging; without it, audits tend to object to every other word.
Word lists cut both ways. A nursery advertising for a room leader will see "nurturing", "caring" and "supportive" flagged as feminine-coded, and an audit that follows the list may suggest dropping them. Don't: caring for two-year-olds is the job. Balance them instead with the parts of the role that word lists miss, such as planning the room's weekly activities, writing observations for parents and leading two assistants, so that men considering childcare see responsibility in the role as well as warmth. The question is the one that sorted the sales clichés later on: does the word describe the work, or a type of person?
Step 4: a second check that isn't ChatGPT
One model checking another model's work shares its blind spots. Add one independent check:
- Word's Inclusiveness checks. In the Microsoft 365 version of Word, the grammar settings include an Inclusiveness section covering age, disability, gender, race and other references. It's off by default; tick the boxes you want, then run the Editor over the advert.
- Gender Decoder. A free web tool built on the 2011 word lists. Paste the advert and it counts masculine- and feminine-coded words. It matches word stems without context, so "lead time" will register as "lead"; read its output with that in mind.
- A person. Ask someone from outside the team, ideally from a group you're worried about putting off, to read it and say which line would make them hesitate. Two minutes of that is often the most useful check of all.
The person check catches what tools can't, because not every barrier is a coded word. Picture a wholesaler's warehouse advert that has passed both the audit and Gender Decoder. The outside reader, a parent who works school hours, stops at one line: "overtime as required." To her it means she can't plan an evening. The owner checks the rota and finds overtime is optional and runs to about two evenings a month, almost all in the busy season. The line becomes "overtime is optional; it's usually two evenings a month in our busy season, paid at time and a half." The original contained nothing a word list would flag. It was the open-endedness that would have put people off.
Sales roles: where coded wording hides in plain sight
Sales adverts attract the heaviest coding, because the clichés are so familiar nobody notices them. A software reseller's draft for an account executive might ask for "a hungry hunter with a killer instinct, thick-skinned, who thrives on aggressive targets". Every phrase there describes a personality, not a job, and an audit will flag most of them.
The fix isn't to pretend the job is gentle. Sales involves rejection, and an honest advert says so. The difference is describing the work rather than the person:
"You'll call around 30 IT managers a day, run product demos over video, and own a quarterly target for new customers. Most calls end in a no, so you'll need to be comfortable hearing it and moving on. Commission is paid monthly, and you'll spend your first two weeks shadowing before taking your own calls."
That version tells a candidate far more, including the hard parts, and lets people who'd be good at the job but wouldn't describe themselves as a "hunter" see themselves in it. It also gives you something to test in an interview: the call volume and the demo are concrete, so you can ask about them.
The same approach works for any role that has picked up a macho or youthful stereotype: warehouse, events, trades, tech support. Ask ChatGPT, in the audit chat, to list every phrase that describes a personality trait rather than a task, then replace each with the task behind it. It's a narrower instruction than "remove bias" and tends to produce more useful changes.
Before and after: an online clothing shop's customer service advert
An online clothing shop with six staff is hiring its first dedicated customer service person. The owner's first version, drafted quickly with ChatGPT and lightly edited:
Before: "We're a young, fast-growing fashion brand looking for a bubbly customer service superstar. You'll be a digital native who lives on social media, a native English speaker with a flawless eye for style, and able to work flexibly including evenings and weekends. Degree preferred. Must be a great cultural fit for our fun team!"
After the requirements sheet, the drafting rules and the audit:
After: "You'll answer our customers by email, chat and Instagram messages: order questions, sizing advice, returns and the occasional complaint. You'll need to write clear, friendly replies quickly and keep track of where each order is using our order system, which we'll teach you. An interest in clothes helps, because customers ask about fit and fabric. Hours are Monday to Friday, 9 to 5:30, with one Saturday morning a month on a rota, paid at time and a half. Pay is [range]. To apply, send your CV and a reply to the sample customer email below; no cover letter needed."
What changed, and why. "Young", "bubbly", "digital native" and "cultural fit" went, because none described the work. "Native English speaker" became "clear, friendly replies", which is the actual skill. "Flexibly including evenings and weekends" became one Saturday morning a month, because that's what the rota needed once the owner checked. "Degree preferred" went because nobody could say what it was for. And the application asks for a sample reply instead of a cover letter, which tests the job's core skill and removes a hurdle that favours confident writers of cover letters over good writers of customer emails. For the wider question of making adverts attractive as well as fair, see writing job descriptions with AI that attract good hires.
How to tell whether the rewritten advert widened the field
Compare the new advert with the last one for the same role on three numbers: total applicants, how many met every must-have on the sheet, and how many of those you'd interview. As an illustration, the print shop's old advert might have drawn 12 applicants, 3 of whom met the must-haves; the rewritten one draws 26, and 10 meet them. The second number matters more than the first. More applications that miss the must-haves just means more reading. More that meet them means the rewrite removed barriers rather than adding noise.
Treat one advert as a hint, not proof: the time of year, the job board and the pay range all move these numbers too. Keep the three figures for each role in a simple sheet, note what changed in the wording, and look again after three or four hires. If the must-have count doesn't move, the barrier is probably elsewhere, in the pay, the hours or the application steps below.
What rewording can't fix on its own
A neutral advert attached to a narrow process still produces a narrow shortlist. Four things to check alongside the wording:
- Pay. A stated range attracts more applicants than "competitive", and removes guesswork that deters people who assume they'll be underpaid.
- Hours and flexibility. If the job could be done in school hours or four days, say so; it's often the single biggest widening you can make.
- The application steps. Timed online tests, mandatory cover letters and long forms each lose people who'd do the job well. Ask for what you'll actually use.
- What happens next. If AI helps screen applications, the same fairness questions apply there; whether AI can screen CVs fairly covers what small employers need to know, and setting up an AI-assisted hiring process links the steps together.
Bias in hiring is one instance of a wider pattern in AI-assisted decisions, explained in AI bias in small business decisions. For adverts, the practical version is simple: facts first, rules in the prompt, an audit in a fresh chat, and one check by something that isn't the model that wrote it.
Follow-up questions on inclusive adverts
Is it illegal to write "recent graduate" or "energetic" in a job advert?
It depends on the employment law where you hire, and on context. In many places, wording that signals a preferred age can be used as evidence in a discrimination claim, even if nobody meant it that way. The safe course is to describe the actual requirement instead. If a phrase is central to the role, such as a graduate scheme, take employment-law advice on how to word it.
Should I let ChatGPT write the salary line?
No. Set the pay range yourself from what you can afford and what similar roles pay, then give it to ChatGPT as a fixed fact. Ask it to include the range as written. A stated range is one of the most useful things an advert can carry, and a model guessing at one can mislead applicants or undercut your own budget.
Is the free version of ChatGPT safe to use for adverts?
For drafting an advert, generally yes, because an advert contains no candidate data and is meant to be public. Avoid pasting confidential material alongside it, such as internal pay bands for other staff or reasons a previous employee left. If your firm has a business plan, use that for anything touching internal HR information.
Further reads
- Is AI CV Screening Fair? Bias Checks Every Recruiter Should Run — Carry the same fairness checks into CV screening.
- How to Test Job Candidates' AI Skills in an Interview — Fair, practical interview tasks once applications arrive.
- How to Redesign Job Roles Once AI Handles Routine Tasks — Check the role itself is right before advertising it.
- AI or a New Hire? How to Decide Before You Recruit — Decide whether the job needs a new person at all.
- AI CV Screening for Recruitment Agencies: Setup and Safeguards — How agencies screen the applicants an advert attracts.
- How to Write a Job Advert for an AI-Savvy Admin Assistant — A worked advert for a specific, AI-heavy role.
- How to Evaluate the AI Features in Your Recruitment Software — Six tests to run on the AI already in your recruitment software, using past placements as the answer key, with pass marks and a filled-in scorecard.
- 12 Custom GPT Examples for Small Businesses, With Setup Notes — Twelve assistants small businesses built as custom GPTs, rebuilt as Projects and Gems after the GPT retirement, each with instructions, files and sharing.
- AI Words to Avoid: A Banned-Phrase List for Your Brand — A ready-made banned-phrase list for AI drafts: the words and sentence patterns that sound generic, what to write instead, and a five-minute check.
- How to Spot Bias and Stereotypes in AI-Written Content — Find the quiet bias in AI drafts: default genders, age and ability assumptions, missing customers and stereotyped images, with before-and-after fixes.
- Questions to Ask a Recruitment Agency About Its AI Screening — Stage-by-stage questions on an agency's AI screening, how to grade the answers, and how a grooming salon found good candidates its agency's filter had rejected.
- How to Hire a Part-Time Marketing Assistant Who Uses AI — Decide what 12 hours a week should produce, write the job description, test candidates on fixing AI output, and set up accounts before day one.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Gaucher, Friesen and Kay (2011), Journal of Personality and Social Psychology; Seong et al. (2024) replication, Strategic Entrepreneurship Journal; Gender Decoder tool notes; Microsoft Word inclusiveness settings documentation.