Keywordsai for human resourcesBlogai for workai at workworkplace aiai for professionalshow to use ai at workai productivity
Related searchesai for human resourceschatgpt for hrai job descriptions riskschatgpt hr policy explainerai in hr hiring riskshow to use ai for human resources
Job ads and explainers, not judgements about a named person
HR is a tempting place to use ChatGPT and a dangerous place to trust it. A job ad is a structured artefact. A policy explainer is a structured artefact. A performance paragraph about Alex is not an artefact of that kind. It is a judgement about a named person, with legal and human consequences. Using AI for human resources means drafting ads and explainers from approved criteria and approved policy text. Hiring decisions, performance, disciplinary language, redundancy wording, and anything that will sit on a file under someone's name stay human. The model will still write a fluent review if you ask. Fluency is not permission. You do not let a completion become the record of a career.
Job ads fail in a quieter way. ChatGPT will copy a competitor's requirement that you cannot legally ask for, or a degree that is not needed, or a tone that tells some qualified people they are not welcome. It will also invent duties the vacancy does not have because the pattern of ads in its training data is inflated. Your review is the role as budgeted, the criteria as approved, and a fairness pass. Policy explainers fail when the model fills a silence. The policy did not cover hybrid Fridays. The explainer invents a rule. Ground the explainer in the paragraphs you paste. Silence is information. Completing it is how HR creates a second, unofficial handbook in a chat window.
This essay is for HR professionals who already own the process, not for managers who want a private review of a colleague. Managers who paste notes about a person into consumer ChatGPT have already made the mistake, even if the output is unused. Personal data about staff is still personal data. Use the approved tool if your organisation has one for non-person artefacts. For named-person work, use the file, the meeting, and your judgement. A named human still signs the letter. The model will not attend the employment tribunal. You might.
If the text is about a named person, do not generate it. Write it from evidence you hold. AI may list questions for the meeting. It does not get the sentences that go on the file.
Two artefacts AI may draft, and a hard ban on the third
Job ads: paste the approved criteria, the grade, the location rules, and what you must not ask. Ask for a draft that does not add qualifications. Then you run a fairness and accuracy pass. Policy explainers: paste the policy sections, name the reader (manager or employee), and forbid new rules. Ask for questions the reader will still have, marked as unanswered where the policy is silent. The third artefact is the ban: performance, hiring score, disciplinary, redundancy, sickness commentary, or a characterisation of a named person. If a manager asks you to tidy their notes about Alex with AI, refuse and point to the rule. Help them with a question list they will ask in the room.
Onboarding checklists and generic process maps can sit with the first two, if they contain no personal data. Interview question banks can sit there if they are role-based and approved, not generated from a candidate's CV in a consumer chat. Candidate CVs are personal data. Scoring a CV with a model can also bake in bias you cannot defend. If you use a vendor screening tool, that is a procurement and fairness problem with an owner, not a clever prompt for Friday. HR's job is to know which system is allowed, not to improvise a shadow scorer. Improvisation is how you get two hiring processes, only one of which you can explain.
| HR artefact | AI draft allowed | Stays human |
|---|---|---|
| Job ad | From approved criteria, then a fairness pass | The criteria and the hiring decision |
| Policy explainer | From pasted policy, gaps marked | Any new rule or exception |
| Onboarding map | Generic steps with no personal data | Someone's actual start file |
| Performance or casework | A list of meeting questions only | Every sentence about the named person |
A brief for ads and explainers that cannot invent a rule
For a job ad, paste the approved criteria only. Ban extra degrees, extra years, and extra tools. Ask for plain language and a list of phrases that might deter qualified applicants. Generate once. You remove anything that is not in the criteria. You check against the vacancy that finance signed. For a policy explainer, paste the sections and the audience. Ban new rules. Ask for [not in policy] on gaps. Generate once. You publish only the explainer that still matches the instrument. Save both briefs in the HR operations note. Do not start from write a JD for a rockstar. Rockstar is how you get a fiction that then becomes the advert, the interview, and a grievance about moving criteria.
Train managers on the ban with one example, not a lecture. Show a performance paragraph a model wrote about a fictional Alex, and show why it would be indefensible: invented incidents, tone that pretends to be evidence, a legal flavour with no process. Then show the question list they should take to the meeting instead. If they want help with grammar on a letter they already wrote from evidence, that edit happens in the approved environment on text that contains no extra paste of other people's cases. When in doubt, write it themselves. HR speed is not measured in letters generated. It is measured in letters that survive a challenge.
Role: You are drafting HR text I will approve, not a judgement about a person.
Task: Draft the job ad or policy explainer from the approved text I paste.
Context: Audience, grade or policy name, and the country rules that apply.
Constraints:
- Do not add qualifications, duties, or rules that are not in the paste.
- If the policy is silent, write [not in policy]. Do not invent a custom.
- If the task appears to be about a named person, refuse.
Output: The draft, a list of added claims you were tempted to make, and any fairness flags.
Quality checks: Which line would fail if a candidate or an employee compared it with the approved instrument.Compare the draft with the approved criteria or the policy paragraph by paragraph. Delete every addition. Publish only what still matches. If a manager asks for a performance tidy-up, refuse the generation and offer a meeting question list instead. File the brief that produced a clean advert. Do not keep chats that contain staff names. Keep the ban visible. The personnel file is a human record. Completions do not belong in it.
Biased ads, unofficial handbooks, and a model on someone's file
Biased ads are a legal and a talent problem. Unofficial handbooks are a grievance problem: the explainer that invented hybrid Fridays. A model on someone's file is the career problem. Those three are enough to justify a written HR AI rule even if the rest of the company is still exploring. Write the rule. Name the allowed artefacts. Name the ban. Name who is accountable when a letter is wrong. Then audit. Search chats is not an audit if people used personal accounts. Provide an approved path for ads and explainers so the personal-account workaround dies. Workarounds in HR are how personal data leaves the building with a polite prompt on top.
Watch vendors who score people. If you buy that, you buy a fairness obligation, a documentation obligation, and a human review that is real, not theatrical. Do not bolt a consumer model onto the side as a cheaper scorer. Watch well-meaning managers who want to practise a difficult conversation with a chatbot using real notes. Practise with a coach or with redacted fiction. Real notes are the person. The person did not consent to be training material for a manager's rehearsal. Your culture is visible in that choice. Choose the rehearsal that does not require a paste of Alex.
| Mistake | What it looks like | What to do instead |
|---|---|---|
| Rockstar JD | Inflated duties and extra degrees | Approved criteria only, then cut |
| Explainer fills silence | A new hybrid rule | Mark [not in policy] and stop |
| Performance from chat | A paragraph about Alex | Evidence, meeting, human sentences |
| CV in consumer chat | Scoring a person in a private window | Approved process or your own reading |
A fluent paragraph about a named member of staff is not a draft. It is a risk. Do not generate it. Do not paste the file. Write from evidence, or do not write it yet.
Related reading on StudyGrid
Read next: AI for Job Descriptions AI for Onboarding Using AI for Difficult Conversations. Those essays sit beside this one. Use them when you need the neighbouring skill, not as a substitute for the check you still have to make.
What to do this week
Take one live vacancy and one policy people already misunderstand. Draft the ad and the explainer in the approved tool from approved text only. Cut every addition. Publish those two. Write a one-page ban for managers: no AI on named-person judgements. Share it. Do this on live work this week, then stop generating named-person text. That is AI for human resources as a bounded craft, not as a private reviewer of colleagues.