AI at work

AI Resume Screening: How to Shortlist Candidates Fairly with ChatGPT

By Sauravjit Singh · Updated 2026-11-10 · 7 min read

Short answer: AI resume screening works best as a first-pass sorter, not a decision-maker. Remove names, photos, ages and other identifying details, give the AI a written scoring rubric built from the job's must-haves, ask it to quote evidence for every score, and have a recruiter review every shortlist and rejection. Test the setup for bias before you rely on it.

Used this way, a general assistant such as ChatGPT, Claude or Gemini can cut the time it takes to read a pile of 150 applications for a sales executive role from most of a day to an hour or two of review. Used carelessly, it can quietly reject good people for reasons nobody would defend out loud.

This guide gives you the workflow we recommend to HR teams: an anonymisation step, a rubric prompt you can copy, a hedged note on India's data protection law, and the swap test we use to catch bias. If you are new to AI in HR generally, start with our wider guide to using ChatGPT across HR work.

What should AI do in screening, and what should it never decide?

AI should read, extract and score against criteria you wrote. People should decide who moves forward, who is rejected and why. The table below is the split we suggest.

TaskFine to give AIKeep with a human
Pull skills, tools and years of experience from a resumeYesSpot-check a sample
Score each resume against a written rubricYes, with evidence quotesReview every score near the cut-off
Final shortlistSuggest onlyYes, always
Rejection of a candidateNoYes, with a reason you could explain to the candidate
Judging culture fit, attitude or 'potential'NoYes, in interview
Explaining career gapsNoYes, by asking the candidate

A useful rule: if you would be uncomfortable telling a candidate 'a chatbot decided this', the step belongs with a person.

How do you anonymise resumes before pasting them into AI?

Strip anything that identifies the person or hints at gender, religion, caste, age, region or family status, then replace the name with a code. This protects the candidate's data and removes the details most likely to bias a score.

  1. Export each resume to plain text (copy from the PDF or use your applicant tracking system's text view).
  2. Replace the name with a code such as C-014 and keep the code-to-name mapping in a separate sheet that never goes into the AI tool.
  3. Delete phone number, email, LinkedIn URL, home address and PIN code, photo, date of birth, age, marital status, father's or spouse's name, religion and nationality lines.
  4. Decide on college names in advance. For most non-specialist roles, replace the institution with the degree only ('B.Com, 2022') so a big-name college does not outweigh actual work.
  5. Keep job titles, employers, dates, tools, projects and measurable results. That is what the rubric scores.
  6. Read two anonymised resumes yourself before batching. Hobbies, sports teams and language lines often leak gender or community, so trim them if they are not job-relevant.

This takes a few minutes per resume by hand. For volume, many applicant tracking systems now offer a 'hide personal details' or blind-review view; check whether yours does before building a manual process.

Is it legal to put candidate resumes into ChatGPT in India?

It can be, but it depends on how you do it, and you should confirm your setup with a lawyer. A resume is personal data, and India's Digital Personal Data Protection Act, 2023 applies to digital personal data your company handles. The Rules under the Act were notified in 2025 with obligations phasing in over time, so check the current position on the official Ministry of Electronics and IT website or with counsel.

Practical pointers that most lawyers will raise, offered as orientation, not legal advice:

  • Tell candidates in your job post or privacy notice that applications may be processed with software tools, including AI, for shortlisting.
  • Use a business or team account whose terms say your inputs are not used to train models, rather than an employee's personal free account. As of November 2026 most major assistants offer such plans; read the current data terms yourself.
  • Anonymise first, as above. Less personal data in the tool means less risk.
  • Keep resumes only as long as you need them, and delete them from chat history if the tool stores conversations.
  • Write the rule down. Our guide to drafting a company AI usage policy includes where candidate data sits in a green, amber and red table.

Data protection is not the only legal angle. Existing employment law already restricts some kinds of hiring discrimination, for example on gender under the Code on Wages, and using an AI tool does not shift that responsibility away from the employer. If a shortlist is ever challenged, you need to be able to explain how it was made, which is another reason to keep the rubric and the swap-test record below.

The scoring-rubric prompt (copy and adapt)

Say you are hiring an inside sales executive for a furniture showroom in Kirti Nagar. Write the rubric yourself first, from the job description, before any resume touches the AI. Then use a prompt like this:

'You are helping a recruiter do a first-pass screen. Score the resume below against this rubric only. Criteria and weights: (1) B2B or B2C sales experience of at least 1 year, 30 points; (2) evidence of meeting a target, with a number, 25 points; (3) CRM or Excel use mentioned, 15 points; (4) Hindi and English communication shown through customer-facing roles, 15 points; (5) furniture, interiors or retail experience, 15 points. For each criterion give a score, quote the exact line from the resume that supports it, and write NOT FOUND if there is no evidence. Do not infer anything that is not written. Do not consider college, location, age, gender or gaps. End with a total out of 100 and one sentence on what a recruiter should ask in a phone screen.'

What to check in the output: every score should have a quote next to it. If you see a score with a paraphrase instead of a quote, the model is guessing. In practice the most common error is generous inference, such as giving CRM points because the person worked at a large company. The 'quote or NOT FOUND' instruction reduces this but does not remove it, so a recruiter reads every resume scoring within about 10 points of your cut-off.

Run each resume in a fresh chat or a clearly separated message. Long chats with 40 resumes tend to drift, and later candidates get compared with earlier ones rather than with the rubric.

How do you test AI shortlisting for bias?

Run a swap test: score the same resume several times with only one detail changed, and see whether the score moves. If it does, the setup is not neutral and needs fixing before real use.

  1. Pick five real anonymised resumes that span strong, average and weak.
  2. Make copies that change one thing at a time: a name that reads as male versus female, names that suggest different communities or regions, a well-known college versus a state university, a one-year career gap versus none, a Delhi address versus a small-town address.
  3. Score every version with the identical prompt, each in a new chat, three times each, because AI output varies between runs.
  4. Average the three runs per version and compare. A consistent swing of more than a few points, or a change in shortlist status, is a red flag.
  5. Fix the cause: tighten the rubric wording, strip the field that caused the swing, or move that criterion to human review.
  6. Repeat the test whenever you change the prompt, the tool or the model version, and keep a one-page record of results.

The point is not to prove the tool is perfectly fair; no one can promise that. It is to find the obvious skews in your own setup and document that you looked.

Mistakes that make AI shortlisting unfair or useless

  • Asking 'is this a good candidate?' with no rubric. The model then applies its own idea of good, which you cannot audit.
  • Letting AI auto-reject. Keep a human on every no, especially for borderline scores.
  • Penalising career gaps. Gaps often mean caregiving, illness or exam preparation, and asking is the fair approach.
  • Copying the JD's wish list into the rubric. Ten 'must-haves' produce low scores for everyone; three to five real requirements work better.
  • Trusting keyword matches. A resume stuffed with tool names can outscore someone who actually ran the work. Ask for evidence of outcomes.
  • Forgetting to tell hiring managers how the shortlist was made. Transparency is part of fairness.

Where this fits in an HR team's week

For a small firm hiring a few people a quarter, the manual workflow above is enough. For a team handling hundreds of applications a month, look at the AI screening features inside your applicant tracking system, but run the same swap test on them; a vendor's fairness claims are not a substitute for checking your own results.

Resume screening is one of several HR tasks AI can speed up. For the bigger picture, including where AI saves time across departments, see our main guide on how to use AI at work. If you want structured practice in prompting, verification and safe data handling, the Applied AI Professional course covers these skills across mainstream AI tools.

Frequently asked questions

Can ChatGPT screen resumes accurately?

It can extract skills and score against a clear rubric reasonably well, but it makes inference errors and its output varies between runs. Treat its scores as a sorting aid and have a recruiter check every shortlist and rejection.

Is AI recruitment biased?

It can be, because models learn from human-written text that carries bias, and resumes contain proxies such as names, colleges and addresses. Anonymising resumes, scoring against a fixed rubric and running a swap test reduce the risk but do not remove it.

Do I need candidate consent to use AI for screening in India?

The DPDP Act, 2023 generally relies on consent given after a clear notice, with limited exceptions it calls legitimate uses, and how that applies to candidate screening depends on your setup and the Rules as they apply now. Mention AI-assisted processing in your privacy notice and confirm the specifics with a lawyer.

What are good ChatGPT prompts for recruiters besides screening?

Useful ones include rewriting a job description in plain language, drafting phone-screen questions from a rubric, and summarising interview notes into a standard format. Keep personal details out of prompts and review everything before it goes to a candidate.

Should we tell candidates that AI is used in shortlisting?

Yes. A short line in the job post or privacy notice is honest, helps with data protection obligations and builds trust. Also give candidates a way to ask for human review.

Want to learn this with a trainer?

Leave your details and our counsellor will call you within one working day with batch dates, fees and the right course for you.

  • Free counselling call, no obligation to join
  • Weekday and weekend batches, online or offline
  • Small batches of 15 to 20 students

Or chat on WhatsApp

Enter a 10-digit mobile number.

We use your details only to contact you about Spettro School courses.

Chat on WhatsApp