ChatGPT for Recruiters: What Works, What Risks HR Must Manage

ChatGPT for recruiters means using a general-purpose chatbot to draft, summarise, and extract hiring text — job posts, outreach, interview kits, and recruiter notes — while a person still owns the criteria and the decision. It is not an assessment product, and it is not a fair way to rank applicants. Many talent teams already paste a CV and a job description into a chat window because the reply comes back in seconds. That speed is real. Treating the reply as proof of who can do the job is the risk.

Use ChatGPT where the output is a draft a recruiter will edit, and keep it away from any step that selects or rejects a person. Safe work includes rewriting a job description from competencies the hiring manager already agreed, turning those competencies into interview questions, pulling stated experience out of a résumé into a checklist, and summarising notes the interviewer already wrote. Unsafe work is asking the model for a fit score, a rank order, a rejection, or a read on personality, culture, or potential. The model does not watch the candidate do the work, and a second run can order the same CVs differently. Under the EU AI Act, systems intended to filter applications or evaluate candidates are listed as high-risk employment uses. If you need proof of skill, send a job simulation built from the live job description and score the task with a rubric a human can explain.

What ChatGPT does well for recruiters

Treat the chatbot as a writing and reading assistant. It is strong when the source text is already in front of you and a human will check the result before it leaves the team.

• Job posts from agreed competencies. Give it four to six must-haves the hiring manager signed off. Ask for a clear post, not a wish list of every tool the team has ever used.
• Structured interview kits. Ask for questions mapped to those competencies, plus what a strong and a weak answer sounds like. The recruiter still edits tone and removes anything that is not job-related.
• Extraction, not scores. “List what this CV states about X, and quote the line” is useful. “Score this person 0–100” is false precision.
• Note cleanup. After a human interview, a summary of the recruiter’s own notes saves time. It is not a second interviewer.
• Outreach drafts. First-touch messages that a person sends, written from the role and the candidate’s public work, not from inferred personality.

Pair drafting with a real sourcing step. AI sourcing and matching should rank against documented criteria you can show a hiring manager, not against a chat transcript nobody can replay the same way twice.

What should HR never ask ChatGPT to decide?

Do not ask a general chatbot to decide who is worth an interview, who is a “culture add,” or who should be rejected. A CV does not contain a reliable signal for personality, potential, or how someone will behave on the job. When the prompt asks for those anyway, the model still answers. It fills the gap with patterns from similar text, including background markers that have nothing to do with the work.

A University of Georgia study reported in September 2026 found that candidates exaggerate more when they know an AI will score a one-way interview, and that the AI rater missed the exaggeration while human raters did not. Pasting written answers into ChatGPT and asking “who told the truth” repeats that failure mode. You get a confident paragraph about language, not a check on whether the person can do the task.

The same limit applies to stack-ranking a slate. If the order changes when you regenerate the answer, you do not have a procedure. You have a suggestion. Suggestions can help a recruiter prepare. They cannot be the employment decision, and they should not be the only thing in the file.

Why does a pasted prompt fail as an assessment?

An assessment has a task, a rubric, and a record of what the candidate produced. A prompt has a paragraph. Those are not interchangeable.

• It scores talk about work, not work. A fluent description of a migration is not the migration. AI interviews have the same gap: they scale conversation, then miss exaggeration.
• It is unstable. The same CV and the same prompt can come back in a different order. A hiring process has to be explainable on a bad day, not only on the run you liked.
• Candidates use the same tool. Unsupervised take-homes written “with ChatGPT in the other tab” collapse into a contest of prompting. If you cannot tell whose work you scored, you did not score the candidate.
• It is hard to audit. New York City’s Local Law 144 covers automated employment decision tools that substantially assist hiring decisions, and expects a bias audit plus candidate notice. A private prompt that ranks people is difficult to audit because the procedure is not stable. Prefer extraction you can check against the source document.

A job simulation fixes the object of measurement. The candidate does a slice of the role. Reviewers score that artifact against competencies you wrote down before anyone applied. Humans still decide. The model, if you use one, helps build the task from the job description — it does not invent the shortlist from prose.

How should HR use ChatGPT next to a job simulation?

Keep the chatbot on the paperwork. Put the proof of skill in a role simulation, then interview for what neither tool can see: motivation, judgment with incomplete information, and how the person works with others.

1. Write the competencies first. Four to six must-haves, agreed with the hiring manager, before any prompt.
2. Let ChatGPT draft the post, the outreach, and the interview kit. A recruiter edits every line that will be seen by a candidate.
3. Source against those competencies with matching you can explain, not a one-off “who is best?” paste.
4. Send a short job simulation from the live job description via AI role-simulation assessments. Tell candidates what is measured and whether outside tools are allowed.
5. Interview the gaps. Use the kit ChatGPT drafted only as a starting point. Score the simulation artifact, not the chatbot’s opinion of it.
6. Decide as humans. The hiring manager should be able to say which evidence they used, and to override a score.

That sequence is the practical version of “use AI in recruitment” for lean teams. It is also how you avoid paying for a long panel just to discover the polished CV could not do the work. See Dotof for hiring companies if you want sourcing and simulation in one workflow.

What risks should HR manage before anyone pastes a CV?

• Candidate data. Do not paste names, contact details, or assessment answers into a personal chatbot account. Use a workspace your privacy review has approved, keep training off when the product allows it, and paste only the minimum text the task needs.
• Disclosure. Greenhouse’s 2026 U.S. survey found that 63% of job seekers had faced an AI interview, 38% had walked away from a process because of one, and 70% said they were not clearly told AI would evaluate them. If a model helps evaluate, say so. Hidden use is a trust problem even when the draft itself is harmless.
• High-risk use, not just high-risk products. EU AI Act Annex III lists recruitment and selection systems that filter applications or evaluate candidates as high-risk. Drafting a job post is not the same act as scoring applicants. Label the step honestly so legal and HR are looking at the right one.
• Integrity of the task. If candidates can quietly use a chatbot on an unsupervised test, your signal collapses. Prefer a supervised or defend-your-work simulation, and read how to prevent AI cheating in hiring assessments before you scale.
• Human override. Anyone who can be rejected should have a person who can explain the criteria. A chat log is not that explanation.