Talent Acquisition · 12 min read

TAFEP-Safe AI in Screening and Recruitment

By the FPA team · Published 7 September 2026 · Anchored to Tripartite Guidelines on Fair Employment Practices

AI screening is the highest-leverage use case in Singapore talent acquisition — and the fastest way to end up defending a hiring decision to TAFEP. The models are getting better; the enforcement is getting sharper. This guide is the operating manual for using AI in screening without losing the ability to defend how you shortlisted.

The one rule TAFEP cares about. An employer must be able to demonstrate that the hiring decision was made on merit, not on protected characteristics (age, race, religion, gender, nationality, family status, disability). Any AI in the loop must not undermine that demonstration.

Why AI screening trips the TAFEP wire

Three reasons AI-assisted screening lands employers in TAFEP conversations:

Proxy variables. The model learns that a name pattern, a school, a country of residence, or a career gap predicts "fit". Those proxies correlate with protected classes. The model looks neutral; the outcome is not.
Opaque decisioning. When a hiring manager says "the tool ranked her lower", the employer cannot show what tipped the ranking. The burden of proof under TAFEP inspection sits with the employer, not the vendor.
Automation bias. Humans defer to AI rankings. "The tool put him on top" becomes the reason for the interview slot. That is a decision by the tool, not by a human.

The five controls of TAFEP-safe AI screening

1. Job-relevance gate

Every criterion the AI weighs must be tied back to a bona fide job requirement documented in the JD. If your JD does not name "5 years experience in FMCG", the model must not down-rank a candidate for lacking it.

2. Protected-class stripping

Before candidate data hits the model, strip name, photo, date of birth, nationality, gender, religion, family status, and NRIC. This is table-stakes. If the vendor cannot do it, do it in a pre-processing script.

3. Bias-scan pass

Run every AI-generated shortlist through a second AI pass that scores for concentration on protected-class proxies. If the shortlist has drifted (e.g. all candidates from one country, all under 30), flag for human re-review before the manager sees it.

4. Human final decision with reasoned override

A named human — not the AI — must make the final shortlist call, with a written one-sentence reason for including or excluding any candidate the AI ranked differently. Store the reason with the decision.

5. Candidate transparency notice

Include a plain-English AI-use disclosure on your careers page and application form. Not a legal wall of text — a "we use AI to assist screening, a human makes the final call" statement plus a contact for questions.

The AI screening decision log

Every AI-assisted shortlist needs a decision log with:

Requisition ID and JD version
AI tool name and version
Input fields sent to the tool (i.e. what data classes)
AI ranking output
Bias-scan output
Human reviewer name
Human final shortlist and reasoned overrides
Date/time and retention until

This log is what you show TAFEP if asked. FPA workshops rehearse it with mock investigations. See the AI in Talent Acquisition workshop.

Common bias-inducing patterns to catch

Pattern	Why it breaks TAFEP	Fix
Model prefers CVs written in a specific tone	Correlates with country of education	Score on capability keywords, not prose fluency
Model down-ranks career gaps	Correlates with family status, health	Explicitly instruct model to ignore gaps under 24 months
Model uses "years since graduation" as a proxy for experience	Correlates with age	Score on relevant experience only
Model weights local school names higher	Correlates with nationality	Redact institution names, evaluate credentials generically
Model favours candidates who match past hires	Encodes historical bias	Do not use "similar to top performer" prompts

A TAFEP-safe AI screening prompt

The pattern taught in FPA's TA workshop, simplified:

You are a TAFEP-anchored screening assistant. Evaluate the candidate against the JD criteria listed below.

RULES:
- Only score against the listed criteria. Ignore all other information.
- Do NOT infer or reference candidate age, gender, race, religion, nationality, family status, or disability.
- If a criterion cannot be assessed from the CV, state "insufficient data" — do not guess.
- Return: a score 1-5 per criterion, a short evidence quote, and a plain-English rationale.

JD CRITERIA:
{criteria}

CANDIDATE CV (protected fields already stripped):
{cv}

Full prompt library with variants for different role types is in the ChatGPT prompt library guide.

What TAFEP is likely to ask if they investigate

What AI tool did you use, and what was its role in the decision?
What data did the AI see about the candidate?
How did you validate the tool did not use protected characteristics?
Who made the final decision, and what evidence supports it was based on merit?
How do you handle candidate requests for explanation?

If you cannot answer all five with documents, you are not defensible.

Frequently asked TAFEP-AI questions

Is video-interview AI (facial / voice analysis) allowed in Singapore?

Technically permitted, practically avoid. Facial and voice analysis correlate with protected classes in ways vendors rarely disclose. If deployed at all, it should be advisory only and never determinative.

Can I use AI to summarise candidate LinkedIn profiles?

Yes, with the same job-relevance gate — the summary must exclude protected-class inferences.

What about ATS ranking algorithms?

Most modern ATSs include ML ranking. Ask the vendor for their fair-use documentation and audit the top-of-funnel outcome distribution quarterly.

Get your TA team TAFEP-safe on AI in one day

FPA's AI in Talent Acquisition workshop walks the five controls with your team's real JDs and rehearses a mock TAFEP investigation. Every participant leaves with the decision log template and the screening prompt library.

Book a discovery call Explore FPA HR Desk →