Write Job Descriptions With AI Without Bias: Step by Step

Write Job Descriptions With AI Without Bias: Step by Step

To write a job description with AI without bias, start from the real work, not from a text template. List what the role delivers, separate what is essential from what is nice to have, and only then ask for a draft. Next, use the AI as a reviewer of exclusionary language and unjustified requirements, and keep a named person responsible for the final decision. AI speeds up the writing, but it does not know what the job truly demands.

Key takeaways

• The quality of the description depends on the input: without real deliverables and context, the AI reproduces the generic, biased text that already circulates.
• Separate essential requirements from desirable ones, and ask for a justification for every demand about degrees, apparent age, years of experience or location.
• Use the AI in two rounds: one to draft and another, with different instructions, to review for exclusionary language.
• Never put personal data about candidates or colleagues in the prompt, and record who approved the final text.
• Test the description on a real opening, watch who applies and adjust, instead of treating the first draft as final.

Why does a job description carry bias even when nobody intends it?

Most bias in job descriptions does not come from bad faith. It comes from copying. Someone takes the previous posting, adds what they remember about the last person in the role, and publishes it. The result mixes tasks that no longer exist, inherited requirements and a tone that describes the people already in the company.

The most common biases show up in four places:

• Language: terms that suggest a particular gender, age group or personality type, such as "young and dynamic", "aggressive profile" or "good looking".
• Inflated requirements: a degree, a language or years of experience asked for out of habit, with no link to the deliverables.
• Hidden conditions: schedule, commute, travel or physical effort that only come up in the interview.
• Omissions: not saying whether the role is open to people with disabilities, whether adjustments are possible, or whether the work is on site or remote.

A language model trained on large amounts of internet text tends to reproduce these patterns, because they are frequent in the postings that exist. That is why a bare request like "write a job description for a financial analyst" usually returns a plausible text full of cliches. The editor's job here is to stop fluent writing from passing as quality.

What should you gather before asking the AI for text?

The AI writes well about what you tell it. Before opening any tool, put together a short set of notes in plain language:

• Purpose of the role: why the job exists and what changes in the business when it is well filled.
• Recurring deliverables: what the person does every week and every month, with concrete examples.
• Decisions and autonomy: what the person decides alone, what needs approval and who they work with.
• Tools used day to day, without listing what would be "good to know".
• Real conditions: location, work format, hours, travel, physical effort, seniority range.
• The first ninety days: what the person is expected to be able to do by the end of that period, without promising numerical results.

A good source for this material is whoever does the work today, or did it recently. Ask what took the most time, what went wrong at the start and what they wish they had known earlier. Those answers are worth more than any ready-made template, and they remove the temptation to invent requirements just to look thorough.

If your company has a company policy for employee ChatGPT use, check first what it allows people to paste into external tools. For a job description, information about the role is usually enough, with no personal data at all.

Step by step: generating the description with AI

The flow below uses two separate rounds. The first drafts, the second critiques. Mixing both in the same request tends to weaken the review.

• Step one, structure: paste the material you gathered and ask the AI to organize it into role purpose, responsibilities, essential requirements, desirable requirements, working conditions and hiring process stages. Tell it not to invent anything that is not in the material and to flag gaps as questions.
• Step two, answer the gaps: the AI will ask about what is missing, such as seniority range or work format. Answer yourself, together with someone who knows the role. Do not let the model fill the gaps on its own.
• Step three, draft: ask for a draft in direct language, short sentences, action verbs in the responsibilities and no personality adjectives. Also ask that every requirement be tied to a responsibility.
• Step four, justify the demands: ask for a plain-text list pairing each requirement with the task that needs it. Anything with no associated task becomes a candidate to cut or to move to desirable.
• Step five, review the language: in a new conversation, hand over only the finished text and ask for a review of exclusionary language, quoting the problematic passages and suggesting neutral alternatives.
• Step six, human review: the person who leads the area and the person who runs the selection read the text using the checklist in the next section.
• Step seven, publish and observe: record the published version, the date and who approved it, and track who applies and where the process loses people.

Here is an example instruction for step five that you can adapt:

Review this text as an editor focused on inclusion: point out terms that suggest gender, age, appearance or personality type, and requirements unrelated to the listed responsibilities, quoting the passage and proposing an alternative.

How do you review exclusionary language and unnecessary requirements?

The AI review helps, but it does not replace the eye of a person who knows the role. Use this checklist as the final filter:

• Does each responsibility describe a concrete action, rather than an abstract quality like "proactive" or "high energy"?
• Are there words that assume gender in the job title or the text, where a neutral form would be possible?
• Is there any reference to age, marital status, children, appearance or place of origin that has nothing to do with the work?
• Is the requested degree truly necessary, or can the same deliverable be shown through experience, a portfolio or a practical test?
• Does the requested length of experience match something the role requires, or is it a number inherited from the previous posting?
• Is the requested language used in real tasks of the role? At what level?
• Are the "desirable" requirements really separated from the essential ones, or does the whole list work as a barrier?
• Are physical, schedule and travel conditions clear and justified, with openness to adjustments where possible?
• Does the text invite people with disabilities to apply and say whom to ask for adjustments in the process?
• Can someone skimming on a phone understand the description, without internal jargon?

A simple test: mentally replace the person who holds the role today with someone who has a different background but can make the same deliverables. If the posting would rule them out for no reason connected to the work, something needs fixing.

Illustrative scenario: an administrative assistant opening

Illustrative scenario: a small distribution company needs someone for its back office. The old posting said "young, proactive, good looking, university degree, fluent English and several years of experience in the same industry".

While preparing the material for the AI, the manager listed the real deliverables: entering invoices into the system, reconciling payments with finance, answering suppliers by email and organizing documents for the accountant. When asked about English, the manager realized it appeared only in rare emails from one foreign supplier, handled with reviewed machine translation. The degree was not used in any task.

In the justification round, the AI pointed out that "good looking" and "young" were not tied to any responsibility, and that "experience in the same industry" did not appear in the tasks. The final text described the deliverables, treated English as desirable and a completed high school education as essential, with organization and attention to detail shown in a small practical reconciliation exercise. The manager still had to decide alone whether the exercise was proportionate to the role, and adjusted its size. The AI proposed the text, but the decision on each requirement was human.

What data and decision precautions should you take?

A job description is a public document and usually contains no personal data. The risk appears when AI use extends from the posting to candidate screening. Two precautions are worth adopting from the start.

The first is not to paste resumes, names or employee histories into the prompt to "calibrate" the text. If the company ever decides to use AI to support candidate screening, first understand what the LGPD says about automated decisions and the right to review and what changes when the tool provider processes data outside the country, a topic covered in international data transfer when using AI providers. The ANPD publishes official data protection guidance that works as a starting point.

The second precaution is conceptual: choosing who moves forward in a hiring process affects people's lives, and that decision should stay with people. Drafting a posting is a low-risk task for AI support. Eliminating candidates is not. The same line appears in what not to delegate to AI in business. To think about risk in a structured way, the NIST AI Risk Management Framework offers a useful vocabulary for mapping, measuring and managing it.

How does the map, limit, pilot and transfer method apply?

The same method used in AI projects for business processes works here, on a small scale.

• Map: record how descriptions are written today, who writes them, who approves them and which old postings get copied. This is the baseline for comparing later.
• Limit: define that the AI receives only information about the role, never candidate data, and that it does not decide who advances in the process.
• Pilot: run the flow on one real opening, with an agreed success criterion, for example the clarity perceived by the manager and HR, and the diversity of profiles that reach the interview, without promising a number.
• Transfer: turn what worked into instructions and a checklist that the HR team maintains on its own, and decide whether to expand to other roles, adjust or stop.

The advantage of treating this as a pilot is that what you learn becomes procedure: the review prompt, the checklist and the approval record are documented, and do not depend on the memory of whoever wrote the first posting.

Frequently asked questions

Can AI eliminate bias from a job description? Not on its own. It helps identify exclusionary terms and propose alternatives, but it can also reproduce the biased patterns present in the texts it was trained on. That is why human review, with a checklist, is still necessary.

Can I use ChatGPT, Claude or Gemini for this? Yes, as long as the company has defined what may be pasted into those tools. For a job description, the ideal is to send only information about the role, with no personal data, and to record which version was approved and by whom.

How many requirements should I list as essential? Only those that, if missing, prevent the person from performing the main deliverables from the start or after a reasonable adjustment period. Everything else goes to desirable or gets cut. If you cannot tie a requirement to a task, it probably should not be there.

Is it worth using AI to screen resumes too? That is a different and higher-risk decision, because it involves personal data and directly affects candidates. Before considering it, assess the LGPD requirements, keep human review of any exclusion and start with a small pilot with clear criteria. The job description is a safer starting point.

Sources

• NIST: AI Risk Management Framework
• ANPD: Autoridade Nacional de Proteção de Dados
• OECD: AI Principles