AI CV screening in Singapore: a fair-hiring decision log

Under TAFEP’s guidance, AI CV screening leaves the employer accountable, not the algorithm: give the tool job-related data only, let people make the decisions and keep a record of each stage, and tell candidates how the tool is used. A shortlisting decision log is one way to keep that record. This page gives our own five-field version, a worked audit of 12 fictional applicants and five checks to run before each hiring round.

TAFEP says all employers in Singapore are expected to adhere to the Tripartite Guidelines on Fair Employment Practices (TGFEP) now. The Workplace Fairness Act comes later. This is general information, not legal advice.

What do Singapore’s fair-hiring rules say about AI screening?

TAFEP’s “Fair Hiring First, AI Second” (20 October 2025) names risks if an AI tool is not set up and checked properly: it can repeat bias from past data, make it hard to see how a score or recommendation was formed, and possibly filter out qualified candidates with non-traditional profiles. It does not ban the tools; employers remain accountable. Its three practices:

1. Job-related data only. Give the tool only job-related data and criteria, based on the job analysis and job description.
2. People in control. “Let AI inform decisions, not make them.” Recruiters or hiring managers review and approve key stages, broadly understand which factors influence the tool’s recommendations, and record the decision at each stage.
3. Openness with candidates. Describe in simple terms what the tool helps with, assure candidates that people make the final decisions where appropriate, and give a contact point for questions or appeals.

None of the other TAFEP or MOM pages we read mentions hiring tools. Our inference is that a tool is part of the selection process, so the TGFEP’s selection rules reach it too: criteria applied consistently to all candidates, and application forms without fields on age, gender, race, religion, marital status and family responsibilities, or disability, with photographs, national service liability and NRIC numbers left until the job offer or given a stated reason. The Act will complement them: TAFEP says it will cover hiring decisions such as “inviting candidates for interviews”, and MOM lists its protected characteristics. This page does not verify which employers it will cover at the start; see our guide to the Act and check MOM or TAFEP.

What goes in a shortlisting decision log?

One entry per decision, five fields. The fields are our design: the TAFEP and TGFEP pages we read ask for records of hiring decisions but do not prescribe a format.

1. Criterion. The job-related requirement, written before any CV is read. TAFEP offers a job analysis form as a sample template.
2. Given to the tool. What the tool saw and which settings were on, for example “CV text, photograph, date of birth and NRIC number removed; auto-reject off”.
3. Returned. The score, rank or recommendation and any reason text, copied as given.
4. Decided by. A named person and a date, never “the system”.
5. Reason, in job terms. One sentence tying the outcome to a criterion. If you cannot write it, the decision is not ready.

Our guide to the Act suggests a four-line decision record for any hiring; this log adds two fields for when software is involved: what the tool was given and what it returned.

What does an audit of 12 fictional applicants show?

Everything here is illustrative. A fictional 22-person office-supplies distributor in Kallang is hiring a customer service coordinator. Its job analysis gives three CV criteria, fixed before the CVs arrive:

• C1: at least 12 months handling customer enquiries by phone or email, paid or unpaid, full time or part time.
• C2: has used an order system, inventory system or spreadsheet to record orders.
• C3: has resolved customer complaints or returns.

The rule, also written first: meet C1 and C2; C3 breaks ties; four interview places. The fictional tool is given CV text with photograph, date of birth and NRIC number removed but education dates left in, and returns a score out of 100 and a one-line reason. Applicant 6’s log entry would record the tool’s reason (“26-month gap between roles”) and the person’s: meets C1, C2 and C3, and a gap is not a criterion.

All 12 outcomes (criteria met · tool score and reason · decision):

1. C1, C2, C3 · 91, “Strong enquiry and order-system match” · Shortlist.
2. C1, C2 · 84, “Enquiry handling and spreadsheet skills match” · Reserve: no C3.
3. C1, C2, C3 · 78, “Enquiry and order-record skills match” · Shortlist.
4. C1, C2 · 76, “Recent employer is a well-known brand” · Reserve: no C3; the tool’s reason is not a criterion.
5. C2 · 74, “Strong spreadsheet skills” · No: five months of enquiry handling, below C1.
6. C1, C2, C3 · 58, “26-month gap between roles” · Shortlist (override).
7. C1 · 66, “Limited systems experience” · No: C2 not shown.
8. None · 40, “Few matching skills” · No.
9. C1, C2, C3 · 63, “Graduated 19 years ago” · Shortlist (override).
10. C1, C3 · 71, “Strong complaints handling” · No: C2 not shown.
11. C2, C3 · 52, “Limited enquiry-handling history” · No: nine months, below C1.
12. C3 · 35, “Few matching skills” · No.

The arithmetic: 6 of 12 applicants meet both required criteria (1, 2, 3, 4, 6, 9) and 4 of those meet C3 (1, 3, 6, 9), which fills the four places without a judgement call. The tool’s top four by score were applicants 1 to 4, so it matches the final shortlist on 2 of 4 places and would have dropped applicants 6 and 9, who met every criterion.

Three of the 12 tool reasons (applicants 4, 6 and 9) cite something that is not a criterion: employer brand, a gap between roles and years since graduation. Without the reason field, scores of 58 and 63 would just look weak. Our inference: a gap can reflect caregiving, which MOM lists among the Act’s protected characteristics, and years since graduation can stand in for age; the TGFEP make a similar point about NRIC numbers, which they say can reveal age. Our suggested fix: remove education dates before the tool sees the CV, and ask the vendor whether gaps between roles affect the score (our AI vendor questions may help). The numbers are invented; they are not typical rates and say nothing about any real product.

Where does the PDPA come in?

PDPC’s Advisory Guidelines on Use of Personal Data in AI Recommendation and Decision Systems, issued on 1 March 2024, are advisory and not legally binding (paragraph 2.2). They were written mainly for AI in products and services: the deployment part is labelled “business to consumer”. The only hiring example we found, paragraph 5.4(c), concerns an internal HR system giving a first cut in matching candidates to vacancies, as a purpose for which the Business Improvement Exception could be relevant to developing an AI system. We found no paragraph on notices to job applicants, so reading paragraphs 9 and 10 across to a CV tool is our inference. With that caveat:

• Consent and notification: unless deemed consent or an exception applies, consent is required to collect and use personal data to provide recommendations, predictions or decisions (9.1), and people must be told the purposes (9.2). Whether an exception covers your use is a legal question this page does not answer. Notifications should be proportionate to the risks (9.4).
• Written policies: organisations using AI systems should include in their written policies the practices and safeguards for fairness and reasonableness, in proportion to the risks (10.3). For higher-impact outcomes, PDPC says organisations may wish to consider whether it is useful to provide information on how human agency and oversight have been implemented (10.6). In our reading a hiring decision is such an outcome, and a decision log could help show that oversight.

If staff also paste CVs into a general chat assistant, see our explainer on PDPC’s generative AI guidelines and the one-page staff AI policy.

What should you tell candidates about AI screening?

TAFEP says employers using AI tools can apply the TGFEP’s transparency principle by clearly explaining how the tools are used in hiring. The wording below is ours, not TAFEP’s; edit it so every sentence is true for your firm:

If the tool rejects applicants automatically, the notice cannot say people decide. The PDPA’s notification duty is separate, so ask whoever handles PDPA compliance whether your application form’s privacy notice covers the tool.

What to check before each hiring round

These five checks are our own design, built on TAFEP’s three practices and the TGFEP. Sign them off each round. TAFEP’s interview evaluation form can support the human stages that follow.

1. Inputs. List every field the tool receives. Each must trace to a criterion. Remove photograph, date of birth and NRIC number and, as we suggest, education dates unless a criterion needs them. Pass: “Given to the tool” fits in one sentence.
2. Criteria. Written before the first CV is read, the same for every applicant, matching the advertisement. Pass: a second person applies them and gets the same answer.
3. Control. Auto-reject is off. A named person approves the shortlist and reads every rejection before it is sent. Pass: nobody is turned away by software alone.
4. Reasons. Every outcome has a reason in job terms; tool reasons that are not criteria are ignored and logged. Before the round, submit one invented CV twice, changing only a detail that is not a criterion, such as education years. Pass: the score does not move.
5. Candidates and records. The notice is live, the contact point answers, and the log is kept with the other recruitment records for at least one year, stored securely. Pass: someone can answer “how was this applicant assessed?” from the log alone.

Next action: run check 1 on your current tool today. More in our AI topic and glossary of AI and automation terms.

Common questions