Appendix B. Self-check answers and explanations

Chapter 1

1-1. Job postings, application forms, screening criteria, and input fields. The most difficult to correct is the job posting. Because individuals who decided not to apply due to the posting do not remain in the data, there is no way to measure or rectify disparities in subsequent stages.

1-2. Direct discrimination is treating someone differently on the grounds of a protected attribute, while indirect discrimination occurs when applying the same criteria produces a significantly disadvantageous outcome for a specific group and the justification for those criteria cannot be proven. The statutory justification phrases are "proving that it is justifiable" (Article 2 of the Equal Employment Opportunity and Work-Family Balance Assistance Act), "reasonable grounds" (Article 4-4 of the Act on the Prohibition of Age Discrimination in Employment and Elderly Employment Promotion), and "justifiable grounds" (Article 4 of the Act on the Prohibition of Discrimination against Persons with Disabilities, Remedy against Infringement of Their Rights, etc.).

1-3. Physical conditions of the applicant not required for performing the job; the applicant's region of origin, marital status, and property; and the academic background, occupation, and property of direct ascendants, descendants, and siblings. The first case falls outside the scope of application of the Act: this Act does not apply to workplaces with fewer than 30 regular employees (as well as civil service hiring by the State or local governments) (obligations under other laws remain). The second case is an issue regarding the statutory requirements of the provision: information genuinely necessary for job performance is not subject to the prohibition. However, necessity must be supported by job analysis and must not exceed the necessary scope, and cases where region of origin or family information is necessary for the job are extremely rare.

1-4. Graduation year can proxy for age, club names for gender (or religion and origin), and career gap periods for gender related to childbirth and childcare, or disability and illness.

1-5. Even when grouped into broad intervals, a substantial correlation between the variables and protected attributes remains. Above all, if a variable is irrelevant to the job, there is no reason to refine and use it in the first place, so the principle is to remove it; if its relevance is unverified, it is proper not to use it for rejection decisions until verified.

1-6. Compare against internally approved working conditions documents, correct any figures that differ from reality, and then post. Publishing conditions that differ from reality from the outset constitutes false job advertising and false recruitment conditions under Article 34 of the Employment Security Act, and Article 4, Paragraphs 2 and 3 of the Fair Hiring Procedure Act prohibit altering the contents of a recruitment advertisement unfavorably to job seekers without justifiable grounds, or unfavorably changing the working conditions stated in the advertisement after hiring.

1-7. The columns of the table are Item, Classification (Job Requirement / Needed Post-Hire / Job-Irrelevant), Potential Correlation with Protected Attributes, Rationale, and Action. Ambiguous items are decided based on job analysis findings (whether requirements cannot be evaluated without that information), validation using internal company performance data, and group-by-group outcomes; prior to a decision, they are not used for rejection decisions.

Chapter 2

2-1. The establishing elements are that it must be a completely automated system, that it must process personal information, and that it must be a decision; automated administrative dispositions by administrative agencies are excluded from application. Because the Act only states that systems applying artificial intelligence technology are "included" rather than limited to them, rule-based programs are also subject to it if they are completely automated.

2-2. Exceptions apply only to the right to refuse. Because the right to request explanation and review (Paragraph 2) and the obligation to disclose (Paragraph 4) do not have provisos based on the grounds for processing, data subjects may request an explanation and the company must disclose the criteria and procedures even if the decision was based on consent or a contract. However, individual requests may be refused if there are justifiable grounds (Paragraph 3, Article 44-3, Paragraph 4 of the Enforcement Decree), in which case the reasons must be notified.

2-3. It is whether the decision was made without substantial and meaningful intervention by a person with legitimate authority. Even if there is a person who clicks an approval button, if they lack the authority to change the outcome or do not actually review it, this intervention is not substantive, increasing the risk of it being deemed an Automated Decision.

2-4. Take measures in writing, etc., within 30 days from the date of receiving the request. If there are justifiable grounds, the period may be extended up to twice, each within the scope of 30 days, upon notifying the reasons. When refusing, the reasons must be notified in writing, etc., without delay; the Commission's notification requires that, unless there are justifiable grounds, the grounds for refusal and the methods and procedures for raising an objection be notified within 10 days from the date of receiving the request.

2-5. What it demonstrated relevance to (whether it was actual job performance), what jobs and applicant pools it targeted, what the sample size was, and what the group-by-group outcomes were. If the target differs, re-verify using internal company data.

2-6. A developer enterprise is a person who develops and provides artificial intelligence; a deployer enterprise is a person who provides products or services using artificial intelligence provided by a developer enterprise; a user is a person provided with products or services; and an affected person is a person whose life, physical safety, or fundamental rights are significantly affected thereby. An HR solution company that sells recruitment evaluation services to client companies is an artificial intelligence enterprise (a developer enterprise if it developed it directly, or a deployer enterprise if it created the service using another company's AI).

2-7. When there is sufficient time, have reviewers evaluate using the original text first without looking at the AI score, and only then re-examine together cases that differ significantly from the AI. When time is insufficient, require decisions to be made only after cross-referencing and linking each AI rationale to the exact location in the original text, and implement sample re-reviews. As effectiveness metrics, one can use the original text view rate, the rate of judgments differing from the AI along with records of their reasons, and the error rate discovered in sample re-reviews.

Chapter 3

3-1. It is the principle that if the purpose can be achieved anonymously, it must be processed anonymously, and if it cannot be achieved anonymously, it must be processed pseudonymously. While payroll calculations requiring payment on an individual basis inevitably identify individuals, survey analysis to identify department-level improvement tasks does not require the identity of respondents, so it must be analyzed using anonymized data.

3-2. Sensitive information may be processed and security measures taken only with separate consent or when required or permitted by statutes. Unique identification information also requires separate consent or specific statutory requirements/permissions, along with measures such as encryption. Resident registration numbers may be processed only where specifically required or permitted by statutes, in cases of urgent interests, or in unavoidable circumstances specified by notification; consent cannot serve as legal grounds. Therefore, they are not collected at stages lacking statutory grounds, such as the application stage, and are processed within that specific scope and encrypted only for processing specifically required by statutes after employment.

3-3. The criterion is for whose purpose the data is processed. If it handles work on behalf of the company, it is entrustment; if the recipient uses it for its own purposes, it is provision to a third party. The entrustor must entrust processing through a document containing statutory details, disclose the trustee and the entrusted tasks, train and supervise the trustee, and decide whether to consent to re-entrustment.

3-4. An example is when the entrustment of processing and storage is necessary for concluding and performing a contract with the data subject. In this case, the items transferred, destination country, timing, method, recipient, purpose of use and retention period, and methods and effects of refusal must be disclosed in the privacy policy or notified via email, etc.

3-5. AI-generated scores, summaries, embeddings, and analysis logs, as well as copies remaining in the vendor's (trustee's) systems. Because outcomes and pseudonymous information that can identify an individual or easily be combined with other information to identify an individual are personal information, when they become unnecessary, they must be destroyed without delay along with the originals (records subject to retention under other statutes are stored separately), and the destruction obligation applies mutatis mutandis to the trustee.

3-6. Statistical purposes, scientific research, and preservation of public records, among others. This special exemption is for analysis that does not identify individuals (Article 28-5 prohibits processing pseudonymous information to identify specific individuals), and because pseudonymous information is also personal information, it cannot serve as the legal basis for processing aimed at singling out specific employees for interviews or reassignment. Such personnel actions must be conducted within the scope of general processing grounds, such as performing employment contracts, and within the notified purposes.

3-7. Processing status, training use, and retention are found in the terms of service, data processing addenda, and privacy policies; location and re-entrustment in sub-processor lists and regional configuration documents; safeguards and supervision in security documentation and contract clauses; and Automated Decision processing in entrustment documents. Cells that could not be verified are marked as "Unverified", and as a rule, personal information is not entered until resolved.

Chapter 4

4-1. They can be divided into general document drafts (low), regulation guidance (medium), and individual evaluation statements, individual metrics, and behavioral monitoring (high). High-risk use cases require inputs restricted to source materials, procedures where humans verify against facts before finalizing, and records that can explain judgments later.

4-2. Examples include external knowledge restriction ("Use only facts present in the provided observation records as evidence") and citation verification ("Attach the supporting record number to every sentence"). The technique of allowing it to admit when it does not know can be incorporated together as "If there is no supporting evidence, mark as 'Insufficient Records'".

4-3. If facts in the work log were verified against objective records such as issue management systems or performance data, they may be used as grounds for evaluation. Using the log as-is without verification cannot be permitted because self-narrative would directly become the evaluation. If installing a system that newly collects or integrates work records, the consultation and notification checks of Chapter 4 must also be conducted.

4-4. Excessive functionality: Attaching HR database modification capabilities to a chatbot intended for regulation guidance. Excessive permissions: An employee-facing chatbot querying using an account capable of reading the salaries of all employees. Excessive autonomy: A chatbot approving leave requests or altering HR records without human verification.

4-5. Query logs, decision logs, and configuration change logs. Because the logs themselves contain personal information, unless the retention period necessary for the purpose and access permissions are defined, risks of indefinite retention and use beyond the intended purpose arise.

4-6. This is because whether consultation is required is the first gateway that establishes the adoption schedule and procedures, whether processing on an individual unit basis is necessary determines what will be installed and collected, and the items to be collected and retention periods are specified on top of that decision. The answers in earlier stages define the scope of subsequent stages.

4-7. The data to compare are metric values against actual performance evaluations and work outcomes, as well as results by age group and disability status. Judgment criteria are the degree of correlation between metrics and performance, and the magnitude and significance of differences between groups. Action options are modifying or excluding metrics (weak correlation), documenting grounds and continuous monitoring (strong correlation), and adjusting the evaluation method for individual employees (reasonable accommodation).

Chapter 5

5-1. Systemic, computational and statistical, and human-cognitive biases. This means that one should not conclude fairness has been established simply because a single metric exceeded a threshold, but should instead treat the results as material for judgment to examine both the root causes and job-relatedness together.

5-2. The selection rate for the former group is 30% and for the latter group is 20%, so the selection rate ratio is 20 ÷ 30 ≈ 0.67. Because it falls short of 4/5 (0.8), this is treated as a signal of adverse impact; after checking the sample size (600 people is on the larger side) and statistical significance, causal analysis begins. The 4/5 rule of thumb is a benchmark from U.S. federal guidelines; it is used not as a legal standard under Korean law, but as a signal for internal checks.

5-3. Even a smaller difference can constitute adverse impact if it is statistically and practically meaningful. Conversely, even a larger difference may not constitute adverse impact if it is not statistically significant due to a small sample size; in such cases, data spanning a longer period can be examined together.

5-4. Looking only at the final pass rate makes it impossible to know whether the disparity arose from the AI document evaluation or the interview conducted by humans, leading to fixing the wrong stage. If narrowed down to the interview stage, one can choose structured interviews (identical questions and scoring rubrics), blind interviews that mask personal background information, independent scoring by multiple interviewers, and interviewer training along with reviewing results by interviewer.

5-5. First, the underlying causes that created the disparity (proxy variables, unverified metrics, biased evaluations) remain intact, causing the same problem to recur. Second, applying different standards based on protected attributes simply to match audit numbers itself constitutes treating people differently based on those attributes. This is distinguished from preferential systems whose legal grounds and purposes are prescribed by law, such as affirmative action.

5-6. Periodic checks, change-triggered checks, and anomaly-signal checks. A vendor's model update is a trigger for a change-triggered check, which calls for advance notice, comparative data on results by group before and after the change, and the authority to roll back to the previous version if a problem arises.

5-7. For example, the responsible party can be linked to GOVERN 2.1; change control to MANAGE 4.1 (change management) and GOVERN 2.1 (roles of personnel responsible for changes); shutdown authority to MANAGE 2.4; inspection frequency to GOVERN 1.5 (post-deployment monitoring to MANAGE 4.1, fairness evaluation to MEASURE 2.11); and appeal channels to MEASURE 3.3. The record retention period is determined by applicable laws and organizational policies rather than NIST items, and data handling when decommissioning a system is linked to GOVERN 1.7. What is important is whether specific individuals, signals, and deadlines are specified for each element.