Chapter 1. Job Postings and Resume Screening Inputs

Learning objectives

• Identify discriminatory expressions related to age, gender, and disability in job postings and revise them into sentences focused on job requirements.
• Distinguish information that the Fair Hiring Procedure Act prohibits requesting as basic screening materials, and distinguish cases where the same information should be collected after hiring.
• Explain the difference between direct discrimination and indirect discrimination, and explain the process by which proxy variables create indirect discrimination in AI resume screening.
• Classify and select input items for AI resume screening into job requirements, information required after hiring, and job-irrelevant information.
• Design a review procedure for inspecting job postings and screening criteria written by AI before publishing and applying them.

Introduction

Suppose a recruiter at a logistics company asked generative AI on a Friday afternoon to "write a job posting for a store manager position." A polished job posting appeared in 30 seconds. It also contained expressions such as "young and dynamic talent," "person with neat appearance preferred," and "industry-top salary." Because the writing was natural, the recruiter was about to post it as was, but read it once more. Each of these three expressions can run afoul of a different law. The first is an issue of age, the second is an appearance condition unrelated to the job, and the third is an issue of displaying working conditions that differ from reality.

AI readily reproduces expressions that were common in its training data. If those expressions reflect long-standing practices, the practices are reproduced as well. Therefore, the first capability an organization using AI in hiring must possess is not the ability to select advanced models, but the ability for humans to filter job postings and criteria produced by AI in light of laws and job requirements.

This chapter covers the entrance of the hiring process. First, it examines how to identify discriminatory expressions in job posting wording, and organizes the information that the Fair Hiring Procedure Act prohibits collecting through applications. Next, it covers what to include and what to exclude in AI resume screening, as well as proxy variables, which explain why discrimination persists even after removing protected attributes. Finally, it establishes a procedure for reviewing job postings and criteria generated by AI.

Key concepts

Three inputs handled by hiring AI

There are broadly three entrances where AI intervenes in hiring. The first is the job posting, which determines who applies. The second is the application form, which determines what information is collected. The third is the document screening criteria and input items, which determine what among the received information is evaluated.

These three influence the process in sequence. If a job posting states "Age 30 or younger," people beyond that age give up on applying altogether. If the application form includes a field asking for family members' occupations, that information enters the reviewer's view and the AI's input. If the screening criteria favor candidates who resemble past successful applicants, past biases are perpetuated directly. Consequently, errors at the entrance are difficult to fix later. This is because people who gave up on applying do not even remain in the data.

Discriminatory expressions in job posting wording

Job posting wording is read alongside three laws.

Article 4-4 (1) of the Act on the Prohibition of Age Discrimination in Employment and the Promotion of Employment of the Aged prohibits employers from discriminating against workers or persons seeking to become workers on the grounds of age without reasonable grounds in recruitment and hiring. Article 7 (1) of the Equal Employment Opportunity and Work-Family Balance Assistance Act prohibits discrimination based on gender in recruitment and hiring, and paragraph (2) prohibits presenting or requiring physical conditions such as appearance, height, and weight, or conditions such as being unmarried, which are not necessary for the performance of the job. The Act on the Prohibition of Discrimination against Persons with Disabilities regards advertisements that express or encourage unfavorable treatment of persons with disabilities without justifiable grounds as discrimination (Article 4 (1) 4), and prohibits employers from discriminating against persons with disabilities in recruitment and hiring (Article 10 (1)).

The reason "Newcomers to the workforce preferred" is placed in the age row in the table is explained in the next section. In job postings and application guides, it is also advisable to provide a channel to request accommodations needed for tests and interviews. At this time, ask for the required accommodations rather than the medical diagnosis of the disability, and manage that information separately from evaluation materials. Even if numbers are removed, expressions that effectively select a specific age group still pose age discrimination issues.

Direct discrimination and indirect discrimination

Direct discrimination refers to treating someone differently on the grounds of protected characteristics (such as age, gender, or disability). "Men preferred" is a representative example. Indirect discrimination occurs when a seemingly identical standard is applied to everyone on the surface, but satisfying that standard is significantly more difficult for a specific group, resulting in an unfavorable outcome for that group.

Korean law explicitly addresses indirect discrimination.

Consider an analogy. Having all applicants take the same test of "climbing five flights of stairs within two minutes" appears fair on the surface. However, if it is an office job where there is no need to climb stairs, this criterion filters out specific groups without relation to the job. Conversely, if the job involves climbing up and down high-rise facilities daily, the same test can be a legitimate requirement. The determination of indirect discrimination examines both the disparity in outcomes and the legitimacy of the criteria.

Scope of application of the Fair Hiring Procedure Act and basic screening materials

The Fair Hiring Procedure Act applies to recruitment procedures of businesses or workplaces that ordinarily employ 30 or more workers (Article 3). Cases where the State and local governments hire public officials are excluded.

This Act divides documents at the application stage into several categories (Article 2). Basic screening materials refer to application forms, resumes, and personal statements. Supporting materials refer to materials that prove matters stated in the basic screening materials, such as degree certificates, employment certificates, and qualification certificates. In-depth screening materials refer to materials that demonstrate competency, such as portfolios and research achievements. These three combined are called recruitment documents. This distinction is used again in Chapter 3 when discussing the return and destruction of recruitment documents.

Most inputs entered into AI resume screening tools come from basic screening materials. Therefore, what candidates are required to enter in basic screening materials directly determines what the AI sees.

Information that cannot be requested as basic screening materials

Article 4-3 of the Fair Hiring Procedure Act prohibits recruiters from requiring job seekers to state the following information in basic screening materials or collecting it as supporting materials, where it is not necessary for the performance of the relevant job.

Two points must not be missed when reading this provision.

First is the condition "not necessary for the performance of the job." If the information is genuinely necessary for performing the job, it is not subject to the prohibition under this provision. However, the justification for necessity must be backed by job analysis, and it must not exceed what is necessary. In particular, cases where place of origin, marital status, or family members' educational background, occupation, or property are necessary for performing a job are extremely rare, so this proviso should not be read broadly. For a job involving lifting heavy equipment and performing work at heights, whether an applicant can perform that work may be asked; however, asking for height and weight is an indirect guess at physical capability and exceeds the necessary scope.

Second is timing. This provision governs basic screening materials—that is, documents at the application stage. Information required for the employment relationship once hiring is confirmed, such as calculating family allowances, can be collected during the onboarding procedure within the scope necessary for that purpose. Collecting it in advance on the application form on the grounds of allowances creates an unnecessary risk of mixing that information into the screening process.

The Fair Hiring Procedure Act is not the only standard. The Personal Information Protection Act (PIPA) also requires collecting only the minimum personal information necessary for the purpose and places the burden of proof on the controller (Article 16), and resident registration numbers cannot be collected based solely on applicant consent without statutory grounds (Article 24-2). Both laws are reviewed together when examining application forms. Detailed discussions are covered in Chapter 3.

According to government guidance (Korea Policy Briefing Policy Wiki, "Fair Hiring Procedure Act"), place of origin here refers to birthplace, registered domicile (formerly ancestral home), and similar records; current place of residence and alma mater are not prohibited from collection; and an identification photograph for identity verification may be attached. However, because residence and school location can serve as variables acting as a proxy to infer place of origin or age, and photos directly reveal physical appearance, it is safer to exclude them from AI evaluation inputs and screening screens even if collected.

Job postings and actual working conditions: Prohibition of false job advertisements

AI sometimes plausibly fills in salary or work arrangements by referencing public market data. If those figures differ from the conditions actually approved by the company, problems arise. Publishing conditions that differ from reality from the outset can constitute false job advertising or presentation of false job offer conditions prohibited under Article 34 of the Employment Security Act, and Article 34 of its Enforcement Decree includes advertisements where working conditions differ significantly from those at the time of application and advertisements where important details differ from the truth within that scope. Article 4 of the Fair Hiring Procedure Act prohibits false job advertisements disguised as recruitment (paragraph 1), changing the contents of a job advertisement to the disadvantage of job seekers without justifiable grounds (paragraph 2), and unfavorably changing the working conditions presented in a job advertisement after hiring without justifiable grounds (paragraph 3). It also prohibits coercing job seekers into assigning recruitment documents and related intellectual property rights to the employer (paragraph 4).

The practical rule is simple. All working conditions in an AI draft must be posted only after cross-checking against internally approved documents. If approval has not yet been granted, write "Negotiable" and do not fabricate numbers. If conditions change after publication, update both the job posting and applicant notices together.

Selecting input items for AI resume screening

Items to be entered into AI resume screening tools are evaluated across three categories.

Three points are checked when making this classification: (1) Without this item, is it impossible to evaluate job requirements? (2) Is this item strongly correlated with protected attributes (age, gender, disability, place of origin, etc.)? (3) Even after removing this item, does the same information leak in through free-form text? If the answer is not (1) and is (2), excluding it is the default. If the answer is both (1) and (2), document the rationale in writing, monitor outcomes by demographic group, and proceed with use.

Proxy variables and indirect discrimination

A proxy variable is not a protected attribute in itself, but is strongly correlated with a protected attribute and consequently acts as a proxy by revealing that attribute. Because AI models identify patterns among input features, even if the gender column is deleted, if other columns correlated with gender exist, the model can produce the same result through those correlations.

Physical appearance in photos or videos is information that directly reveals attributes such as gender, age, and disability rather than acting as a proxy variable. As a rule, it should be excluded from evaluation inputs to prevent job-irrelevant judgments of appearance from intervening.

The NIST AI RMF divides AI bias into three categories: systemic, computational and statistical, and human-cognitive bias. Disparities accumulated in past hiring decisions being captured in data and transferred to models is a textbook example of systemic bias. Furthermore, NIST warns that a system with reduced harmful bias is not necessarily fair. Removing a single variable does not conclude the task; outcomes must be remeasured.

The handling of proxy variables depends on the status of the variable.

The thought that "aggregating postal codes broadly to the province/city level might be fine" is common, but even after aggregation, a substantial portion of the correlation remains. As a rule, irrelevant variables should be excluded rather than refined for use.

Bypass information in free-form text

Even when prohibited items are deleted from the application form, there are cases where the same information enters through alternative channels. Examples include family information being mixed into personal statements, such as "While helping my father's company," or marriage and childbirth timing being revealed in career descriptions. AI can read such information from free-form text and reflect it in scores.

A four-layered response is employed. First, design questions so that free-form questions do not circumvent prohibitions to ask for prohibited information (prompts such as "Describe your upbringing" are representative circumvention pathways). Second, perform preliminary masking before inputting data into AI by concealing names, school names, geographic names, and expressions that reveal family relationships using rule-based filters or named entity recognition (a technology that automatically detects proper nouns such as names, school names, and place names) tools. Because masking is not foolproof, do not rely on it alone. Third, explicitly instruct the AI evaluation prompt not to use information other than job requirements for evaluation, and have it output evaluation rationales as verbatim citations from the source text to verify what basis was used. Because prompts cannot completely control model behavior, a fourth step is required. Fourth, compare scores of applications that disclose prohibited information against those that do not in pilot evaluations to identify leakage points. There are also prompt injection attacks (OWASP LLM01:2025) attempting to manipulate scores by hiding instructions for AI inside applications out of sight of human eyes; therefore, feed the application text designated as data rather than instructions, and verify the original text for any cases with abnormally high scores.

Review procedure for AI-generated job postings and screening criteria

Review AI drafts in the following order:

1. Cross-check job requirements. Compare every requirement sentence in the posting and criteria line by line against the job analysis results, and flag requirements that lack corresponding evidence.
2. Check protected attributes. Search for direct and indirect expressions (phrases that allow guessing age, words evoking a specific demographic group) related to age, gender, marital status, physical conditions, disability, and place of origin.
3. Inspect application form items. Verify that the three clusters under Article 4-3 of the Fair Hiring Procedure Act do not enter through form fields or free-form essay questions.
4. Cross-check working conditions. Compare salary, working hours, work location, and employment type against internally approved documents.
5. Record the review. Record revised contents and reasons, reviewer name, and date. If the same type of issue recurs, add an item to the review checklist.

When asking AI to assist with review, it is safer to have it search for candidates rather than delegating final determinations to it.

Review the following draft job posting. Do not make determinations; compile a list by quoting verbatim the sentences that may correspond to the items below:
1) Expressions directly or indirectly indicating age, gender, marital status, disability, physical conditions, or place of origin
2) Qualification or experience requirements not connected to requirements in the attached job analysis table
3) Numbers or conditions that differ from the attached internally approved working conditions
If there are no such sentences, write "None".
<Draft Job Posting> ... </Draft Job Posting>
<Job Analysis Table> ... </Job Analysis Table>
<Approved Working Conditions> ... </Approved Working Conditions>

Documenting evidence of job relevance

The exit route in indirect discrimination determinations is documentation that can "prove that it is justifiable." For each requirement, it suffices to record the following five items on a single sheet: the requirement statement, the underlying job analysis results (what tasks are performed and how frequently), review of alternatives (whether less exclusionary methods existed), group-level outcome audit results, and the date for the next review. Article 30 of the Equal Employment Opportunity and Work-Family Balance Assistance Act places the burden of proof on the employer in resolving disputes relating to this Act. This means that without supporting documentation, there is no way to prove the legitimacy of the conditions. This document serves as the basis for answering inquiries when applicants or supervisory authorities ask for justifications, as well as a mechanism for recognizing when requirements have become outdated and no longer fit.

Going deeper

The boundary between actual job requirements and discriminatory expressions

Not all requirements are discriminatory expressions. If all personnel at a control center operating 24 hours a day take on night shifts in equal proportion, "eight night shifts per month" is an actual working condition; in fact, specifying it clearly in the job posting allows applicants to make an accurate assessment. Conversely, "young and spirited talent" in the same posting is not a job requirement, but a preference regarding the person.

There are also statutory exceptions regarding age. Article 4-5 of the Act on the Prohibition of Age Discrimination in Employment and Aged Employment Promotion does not consider the following to be age discrimination: cases where a specific age criterion is inevitably required in light of the nature of the duties; reasonable differentiation in wages or welfare benefits taking length of service into account; the setting of retirement ages pursuant to statutes; and support measures to maintain and promote the employment of specific age groups. However, reasons such as "a youthful organizational culture" or "organizational adaptability" are not unavoidable criteria given the nature of the duties.

Three questions mark the boundary: Does this requirement stem from the actual work? Is it applied equally to all applicants? Can evidence be presented if an objection is raised that it disadvantages a specific group? If the answer to all three is "yes," it is fairer to applicants to document the justification and state the requirement specifically rather than deleting it. Well-intentioned suggestions such as "let us remove even actual working conditions because they might disadvantage certain groups" create a different problem: applicants apply without knowing the actual working conditions.

Preferential expressions and affirmative action

Article 2, Subparagraph 1 of the Equal Employment Opportunity and Work-Family Balance Assistance Act excludes "cases where affirmative action measures are taken pursuant to this Act or other statutes" from the definition of discrimination, and Subparagraph 3 defines affirmative action measures as temporary measures favoring a particular sex to eliminate existing employment discrimination between men and women or to promote equal employment. Therefore, not all gender preference expressions receive the same assessment.

However, this exception applies to measures that possess the purpose and nature of being "temporary" and "to eliminate employment discrimination or promote equal employment." Phrases such as "men preferred" or "women welcome" routinely inserted by AI cannot be justified under this exception unless they are part of such measures. To use preferential wording, one must first verify the basis showing which affirmative action measure it belongs to.

The scope of application varies by law

Although the Fair Hiring Procedure Act applies to workplaces with 30 or more regular employees, other laws establish their own separate scopes of application. Even if a business has fewer than 30 regular employees and Article 4-3 of the Fair Hiring Procedure Act does not apply, prohibitions on discrimination based on gender, age, or disability, as well as the principle of minimal collection under the Personal Information Protection Act (Article 16), may still pose compliance issues depending on each law's scope of application. The assumption that "our company is small, so we are fine" is a common error that overextends the scope of one statute to all others. When determining whether company size exempts an obligation, check the applicable provisions of each respective statute.

At work

Job posting review at Zelkova Trading

A recruiter at household goods distributor 'Zelkova Trading' (380 regular employees) received an AI-generated draft of a job posting for an administrative management position at a logistics center. The draft included phrases such as "digitally savvy individuals in their 20s and 30s," "meticulous female applicants preferred," "advanced Excel proficiency," and "annual salary of KRW 40,000,000 or more."

The recruiter proceeded according to the review procedure. First, the draft was checked against the job analysis table. Advanced Excel proficiency was retained because it directly related to preparing inventory reports. "Digitally savvy individuals in their 20s and 30s" was an age-based expression; if "digital competency" was needed, that competency could simply be stated directly. It was revised to "those able to prepare weekly reports using inventory management systems and spreadsheets." "Meticulous female applicants preferred" was a routine phrase unrelated to any company affirmative action measures, so it was deleted and replaced with "experience in identifying and correcting numerical errors." The annual salary was adjusted to KRW 35,000,000 to match the internally approved proposal.

In the review meeting, discarded alternatives were also recorded. A proposal to merely remove "20s and 30s" and replace it with "MZ generation welcome" was discarded because it simply rephrased the same age preference. A proposal to "leave the posting as is and just ignore age during resume screening" was discarded because it would fail to prevent discouraged applicants from abandoning their applications. Finally, "expressions referring to generations or age groups" and "gender preference expressions" were added as items to the review checklist, ensuring that future AI drafts would be filtered at the initial stage.

Input item cleanup at Green Barley Foods

Food manufacturer 'Green Barley Foods' (900 regular employees) received 18 input items proposed by a vendor while introducing AI resume screening for its open recruitment of production management positions. The HR team placed the job analysis table alongside the list and categorized the items into three groups.

Mandatory qualifications (food hygiene-related credentials), relevant years of experience, production facility operation experience, and job assignment scores were retained as job requirements. Salary accounts and emergency contact information were deferred until after onboarding. Registered domicile, marital status, and parents' occupations were prohibited information under Article 4-3 of the Fair Hiring Procedure Act, so they were removed from the form. Identification photos for identity verification were kept on the form, but masked from AI evaluation inputs and reviewer screens.

The difficult part was the gray area. "Graduation year" almost directly revealed age, and because years of experience were already collected separately, it added nothing to job-related evaluation. It was decided to remove it. For "career gaps in the past 3 years," the vendor explained that it "predicts production floor management performance," but this had never been validated with internal company data, and in the pilot data, the majority of applicants with gaps were women. The HR team neither removed this variable immediately nor used it as was. They commissioned a validation analysis using performance appraisal data from internal production management roles and decided not to use scores reflecting this variable to make disqualification decisions until results were available. In the meeting minutes, they also recorded that merely "predicting early turnover" hardly establishes job relatedness, and that the company bears the burden of proving justification for variables skewed toward a specific sex. "Hours of external volunteer activity" lacked any basis connecting it to production management duties, and participation opportunities varied across groups, so it was removed.

Finally, they scheduled a rerun of the pilot evaluation to compare pass rates across demographic groups before and after organizing the items. Unless one verifies that removing variables actually reduced the disparity, the cleanup is not complete.

Common mistakes

• "It is enough to just remove the numbers." Replacing "age 30 or younger" with "entry-level professional" still effectively selects a specific age group. Yielding particularly disadvantageous results for a specific age group based on non-age criteria without reasonable cause is also deemed age discrimination (Article 4-4, Paragraph 2 of the Act on the Prohibition of Age Discrimination in Employment and Aged Employment Promotion).
• "There is no discrimination because protected attributes were removed from the input." Other variables, such as postal codes, graduation years, or career gaps, can reveal those attributes by proxy. Outcomes must be measured by demographic group.
• "Making it optional is fine." If a field exists, information will be submitted. Eliminate the field itself for information unrelated to the job.
• "We collect it in advance on the application because it is needed for allowance calculations." Information needed for the employment relationship should be collected after hiring is confirmed, strictly to the extent necessary for that purpose.
• "Remove actual requirements if they might disadvantage someone." For actual job requirements applied equally to everyone, it is fairer to state them specifically while documenting the supporting rationale.
• "Keep the posting as is and simply do not look at it during evaluation." Discrimination at the recruitment stage has already affected those who were deterred from applying.
• "The salary was written by AI from market data, so it is just for reference." Working conditions in a published job posting constitute an official commitment by the company. They must be cross-checked against internally approved terms.

Summary at a glance

Check the official docs

• Article 4-3 of the Fair Hiring Procedure Act — Check the original text for the three categories of information that cannot be demanded as basic screening materials.
• Article 4 of the Fair Hiring Procedure Act — Check the prohibition of false job advertisements and unfavorable changes to working conditions.
• Article 4-4 of the Act on the Prohibition of Age Discrimination in Employment and Aged Employment Promotion — Read the areas where age discrimination is prohibited and the deemed provisions under Paragraph 2.
• Article 2 of the Equal Employment Opportunity and Work-Family Balance Assistance Act — Read the definition of discrimination, including indirect discrimination and exceptions.
• Article 7 of the Equal Employment Opportunity and Work-Family Balance Assistance Act — Check the prohibitions regarding recruitment and hiring.
• Article 4 of the Act on the Prohibition of Discrimination against Persons with Disabilities, Remedy for Infringement of Their Rights, etc. — Check the seven categories of discriminatory acts and justifiable grounds.
• NIST AI RMF — Characteristics of trustworthy AI — Read the three categories of bias and "Fair – with Harmful Bias Managed".

Self-check

1-1. State the three entry points where AI intervenes in the hiring process, and explain which entry point's error is the most difficult to correct in later stages and why.

1-2. Define direct discrimination and indirect discrimination in one sentence each, and write at least one "exit" expression commonly included in indirect discrimination provisions.

1-3. State the three groups of prohibited information under Article 4-3 of the Fair Hiring Procedure Act, and distinguish and explain the two cases where this article is not an issue (when outside the scope of application of the law, and when the information is necessary for performing the job).

1-4. State which protected attributes graduation year, club name, and career gap period can each reveal as a proxy.

1-5. Explain why the proposal to "group items that could be proxy variables into broad ranges" is difficult to serve as a viable solution.

1-6. When generative AI fills in the annual salary of a job posting by referencing market data, state what must be done before posting and the supporting provision.

1-7. (Design) Select a job role in a company, design a table that classifies ten candidate input items for AI resume screening into three categories (job requirements, needed after joining, job-irrelevant), and state what materials to use to decide on an item whose classification is ambiguous.

Hands-on tasks