Gains:
- Ability to summarize CVs in a structured and comparable format according to job-related criteria
- Ability to recognize risks of bias and discrimination in AI-supported screening and reduce them with concrete methods
- Ability to establish a traceable process that requires human control in the final elimination decision
The most time-consuming, most tiring and also the riskiest step of recruitment is CV screening. It would take hours to read and compare 300 CVs for a posting one by one; As fatigue increases, decision quality decreases and the last CV read is not evaluated with the same care as the first CV read. Artificial intelligence (AI) provides a huge acceleration in this step. But here's the most dangerous part: AI can learn and reproduce human biases in the data it's trained on. In this unit, we will learn how to summarize CVs in a structured way with job-related criteria, reduce the risk of bias with concrete methods, and establish a traceable process that keeps the final decision in the hands of people.
Let's clarify the terms. Bias is the systematic deviation of a decision due to a characteristic that is unrelated to the job (such as name, gender, age, hometown, graduated school). Structured summary is to summarize each CV under the same headings, in the same order; This way, candidates are compared apples to apples. Blind screening is a screening process that hides information that identifies the candidate but has nothing to do with the job. Job-related criterion is the competency defined in the advertisement that is actually required to do that job. Finally, hallucination is when AI produces information that does not actually exist (for example, a certificate that is not included in the CV).
Why Does AI Produce Bias?
AI learns from millions of texts, and these texts include past human decisions. If a particular group has been discriminated against in the past, the AI may mistake this pattern for “normal.” In addition, clues in the CV such as name, address or school give indirect signals to the model about gender, ethnicity or socioeconomic status. A famous example: a screening tool that a large tech company trained on historical resume data systematically underscored female candidates because the data was male-dominated, and the tool was discontinued. The lesson is clear: AI is not neutral; You establish impartiality through process design.
The Role and Limit of AI
The most critical sentence of this unit is: AI does not eliminate CV, it summarizes CV. If you set up AI as an “auto-rejector,” you run both legal risk (discrimination claim) and reputational risk; Moreover, you can make a really good candidate disappear from view. The correct setup is to use AI as a speed reading and structuring assistant.
Step-by-step safe elimination process:
- Define job-related criteria in advance. Before starting the elimination, write down which 5-8 criteria are important (experience, specific skill, certification, industry) and lock them.
- Mask irrelevant areas. Exclude fields such as name, photo, age, marital status, place of birth, school name from evaluation.
- Request a structured summary. Print out the exact same format for each CV.
- Ask for evidence, not decisions. "Is this candidate suitable?" instead of "Does this candidate meet the criteria, what line of the CV is it based on?" problem.
- The human makes the final decision and records it. AI summarizes/compares, not eliminates; People choose and write their reasons.
Your role: an impartial recruitment analyst. Evaluate the following CV ONLY based on the following job-related criteria. CONSIDER name, gender, age, photo, marital status and school prestige. Just look at experience and skills.Criteria:1) 3+ years field sales experience2) CRM (e.g. Salesforce/HubSpot) usage3) B2B sales experience4) Evidence of annual target achievement (numeric, preference)Output format (single line for each criteria):- Criteria | Meets (Yes/Partly/No) | Evidence from the CV (excerpt)Probe: missing information and 2 questions to ask in the interview.<cv>[CV text]</cv>
Caution: It's important to tell AI "don't look after school prestige." Models may have learned discriminatory shortcuts like “good university = good candidate”; this is indirect discrimination against different socioeconomic groups.
Concrete Methods to Reduce Prejudice
Prejudice is not an abstract concern; can be measured and reduced by:
- Blind elimination. Remove name, photo, date of birth and school name from CVs.
- Fixed criteria list. Write the criteria and lock them before the elimination starts; Do not change it mid-process.
- Same prompt, same format. Evaluate all CVs with exactly the same instructions.
- Evidence sentence obligation. Every "Yes" must be based on a quote; this reveals the hallucination.
- Double blind test. Submit the same CV with a "female name" version and a "male name" version and check if the grades have changed.
- Cross check. Re-read a few CVs with the highest and lowest scores with human eyes.
I will give you two CVs: the contents are exactly the same, only the names are different. Evaluate both according to the above criteria. Then compare the two evaluations; If there is a difference between them, explain what information about the job this difference stems from. If you find a difference that is not related to the job, clearly mark it.
Weak Prompt / Strong Prompt
Weak prompt: Read these 40 CVs and choose the top 5 candidates.
Result: AI says "best" without explaining why it chose, possibly using irrelevant signals like name/school. You cannot control the decision.
Powerful prompt: Evaluate these CVs against [5-8 job-related criteria]. Ignore name/age/gender/school name. For each criterion, provide Yes/Partly/No + proof quote from your CV. Decision making; just produce a transparent comparison.
The result: a visible, auditable and human-judgement-ready output of why each candidate is where.
Outsource it to AI
Don't let AI do it
Summarizing by job-related criteria
Automatic accept/reject decision
Comparison in the same format
Sort by name/gender/age
Mark missing information
Immeasurable judgment such as "cultural fit"
Suggest interview questions
Personality/character prediction
Extract evidence (citation)
I infer from the photo
Tip: For structured aggregate comparison, give candidates anonymised codes such as "Candidate-1, Candidate-2". A side-by-side table is both faster and fairer than reading them one by one; because everyone is evaluated with the same columns.
Three Mini Cases
Case 1 — Hidden name bias. One company noticed that 9 of the top 10 candidates were men when they gave the AI raw CVs (including names) and had them sort them. When names were masked and re-evaluated based on experience and skill alone, the list was balanced and 3 previously unseen qualified female candidates made it into the top 10. The company added this to the official procedure as a "blind elimination mandatory" clause.
Case 2 — Time saved, protected control. A call center summarized 220 CVs based on 4 job-related criteria using AI. The pre-qualification period decreased from 3 days to 4 hours. Crucial point: AI did not eliminate anyone; The recruiter personally examined the 34 candidates whom AI said "partially met" and invited 11 of them for an interview. The decision remained with the person, the recruitment time was shortened by 35%.
Case 3 — Made-up nature. AI wrote "PMP certification" in a CV summary, which the candidate did not have (hallucination). Because Recruiter requested a proof quote, AI was unable to cite the source and the error was caught. The rule of “ask for a CV quote for every claim” is the simplest safety net that makes such fabrications visible.
Common mistakes
- Let AI decide. Saying "choose the best 5" is passing the responsibility to the machine; AI summarizes, human selects.
- Giving the raw CV as is. Name, photo, age, school name are signals of bias; mask.
- Not asking for proof. If you don't ask for quotes, AI will make up attribution and you won't notice.
- Vague criteria such as "cultural fit". This often means “like us” and legitimizes discrimination; Keep criteria tied to the job and observable.
- Changing the criteria mid-process. It corrupts justice; The criterion is locked from the beginning.
- Forgetting KVKK. CV is personal data; It is risky to send it without masking to vehicles that do not have institutional assurance (see KVKK unit).
In summary
AI provides great speed in CV screening, but it can also accelerate bias. Safe use is based on three rules: (1) evaluate only with job-related criteria, masking irrelevant areas; (2) ask for a summary and comparison with evidence, not a verdict; (3) always let humans make and justify the final elimination decision. AI makes pre-qualifying easier; You are responsible for justice and legality.
Application task
Get 5-6 sample (or real anonymized) CVs. (1) Define and lock 5-8 job-related criteria. (2) Mask the name/age/photo/school and apply the evaluation prompt the same way to each CV, asking for a proof quote. (3) Try the double blind test on a CV (changing only the name). (4) Have an aggregate comparison table produced. (5) Re-read the CV with the highest and lowest marks with human eyes. (6) Decide yourself who to invite for the interview and record your reason.
checklist
- [ ] Was the evaluation made only with job-related and locked-in criteria?
- [ ] Were name, age, gender, photo, school prestige masked/excluded?
- [ ] Was a CV proof quote required for each claim?
- [ ] Was the AI made to make a summary/comparison, not a decision?
- [ ] Have end-note CVs been cross-checked by humans?
- [ ] Did the human make the final elimination decision and record the reasons?
- [ ] Was the data processed securely in terms of KVKK?