The Moment a Candidate Asks "Why Was I Rejected?"
Korea's Ministry of Employment and Labor has been pushing hiring-process reforms centered on disclosing AI use in advance and prohibiting discrimination, and has signaled that guidelines for AI use in recruitment are on the way. The current Act on Fair Hiring Procedures does not spell out a duty to disclose AI use. But Article 37-2 of the Personal Information Protection Act has been in force since March 2024: where a decision is fully automated, candidates may refuse it and demand an explanation, and the criteria must be disclosed in the privacy policy. There is no reason to wait for the hiring-procedures amendment.
The starting point is not buying a new system. It is counting the ones already running. Document screening scores, AI video interview analysis, performance-review recommendations, even the generative model someone uses to polish job postings — once companies actually list these, most find AI touching more decision points than they expected. Without that inventory, you cannot even draft a disclosure notice.
What Actually Counts as Discrimination
Removing gender, age, school, and region from the input is the minimum condition, not a sufficient one. Those attributes come back through proxy variables. Graduation year reconstructs age, commuting distance reconstructs residence, and certain clubs or military-service categories reconstruct gender with meaningful accuracy. Unintended channels exist too, such as cover-letter length or vocabulary distribution.
The more structural problem is training data. Train a model on five years of successful hires and you reproduce five years of interviewer preference. The model never learns what a good employee looks like; it learns who got picked before.
Measurement starts with comparing pass rates across groups. The U.S. EEOC's practical benchmark, the four-fifths rule (investigate when a group's pass rate falls below 80% of the highest group's), is not a Korean legal standard but works well as a monitoring metric. If men pass document screening at 24% and women at 15%, the ratio is 0.63 — under 80%, and a cause analysis is warranted.
Tuning to hit the metric, however, backfires. Applying different cutoffs by group to equalize pass rates creates fresh discrimination exposure. Use the metric as a signal that points to a cause, and make the actual fixes on the feature and training-data side.
Translating Explainability into System Requirements
"Explainable AI" sounds abstract, but the questions you actually have to answer compress into three: at which stage was AI used, on what criteria did it judge, and what data did it reference.
Translated into system requirements, that becomes log design. If you store only the score, you will be unable to explain anything six months later. At minimum, these items need to be retained per candidate.
When the model changes, the explanation changes with it. Swap models mid-season and the same résumé gets a different outcome, with no way to explain the difference to the candidate. The safer practice is to freeze the model version within a single hiring round, and if replacement is unavoidable, switch at a round boundary and record that fact.
Define the Human Decision Point Like a Contract
Using AI as a screening aid and using it to reject carry very different legal and operational weight. The recommended design has AI produce rankings and rationale while a human finalizes every rejection. To keep "a human reviewed it" from becoming a rubber stamp, the review itself has to leave a record.
An appeals process needs to exist as actual screens. Intake → reviewer assignment → review outcome → notification, each with a deadline and an owner. Without that, an inquiry simply disappears into one recruiter's inbox.
In the evaluator UI, sequencing is an effective way to reduce anchoring. Show the AI score first and evaluator judgments converge on it. Practical alternatives: reveal the AI score only after the evaluator submits their own rating, or present the supporting factors without the score at first.
What to Verify When Using a Vendor Solution
When adopting an external AI hiring solution, the materials to demand during evaluation are clear. The source and composition of training data, group-level bias test results along with the methodology used, the scope of log export (can you pull the records into your own database?), and reproducibility — whether the same input yields the same result.
In the contract, explicitly pin down candidate data retention periods, whether data is transferred overseas, and a clause barring reuse of your applicants' data to train the vendor's models.
The most important premise is this: the duty to explain, to candidates and to regulators, stays with the hiring company, not the solution vendor. "We don't know the internal logic because it's a vendor model" is not an available answer.
A Minimum Framework Within Six Months
| Timeline | Task |
|----------|------|
| Month 1 | Inventory AI use (stage, purpose, vendor) |
| Month 2 | Add disclosure language to postings and application forms |
| Months 3–4 | Build decision-log and review-history retention |
| Month 5 | Define the bias-audit cycle (per round or semiannual) |
| Month 6 | Launch the appeals channel and handling procedure |
Much of this is achievable without replacing your existing HR system. Attaching a decision-log table and an appeals workflow as a separate module, integrated with the current ATS and HR system over APIs, is the realistic option in both cost and timeline.
Drawing on our experience building internal enterprise systems and AI integrations, POLYGLOTSOFT implements decision logging, bias-audit reporting, and appeals-handling screens on top of existing HR and recruitment systems. Our subscription development model also supports expanding this in stages as the regulatory picture develops. If you would like a review of the explainability posture of the hiring and evaluation systems you run today, please get in touch.