Discriminatory Effects and Zip-Code Proxies Are Different Questions
In Public Act 103-0804, the phrase “has the effect of subjecting employees to discrimination” is the hinge of one substantive inquiry, while the separate reference to using zip codes as a proxy creates another. This chapter examines what changes when the law focuses on the consequence of an AI-assisted employment practice rather than only the employer's stated purpose, and why the path taken by a geographic input must be traced separately. The reader should come away able to ask two disciplined questions about the same process—what happened to people, and what role did zip code play—without assuming that one answer settles the other.
“Has the effect” puts the result at the center
Section 2-102(L)(1) prohibits an employer, with respect to the employment decisions and conditions named in the provision, from using artificial intelligence that “has the effect of subjecting employees to discrimination” on the basis of protected classes under the Illinois Human Rights Act. The wording is important because it describes the prohibited use through its effect. The statute does not add a separate sentence requiring proof that the employer wanted to discriminate or that a vendor designed the system for that purpose.
That is an effects-focused rule. In plain language, it asks what the employer's AI-assisted practice does in the employment setting. A system can be built around a neutral business goal—finding qualified applicants, allocating training opportunities, identifying performance concerns, or predicting likely success. The goal may be legitimate as a business description. It does not, by itself, answer whether the use has a discriminatory effect on a protected basis.
Consider a hypothetical hiring screen. An employer asks an automated service to rank applicants by predicted job performance. The employer applies the same ranking process to everyone and does not instruct the tool to prefer or reject a protected group. Those facts describe purpose and procedure. They do not yet describe effect. To examine the effects pathway, a reviewer would need to look at the employment decision the ranking influenced and compare what happened to people in the relevant protected groups. The question is not whether the software used the word “race” or whether the employer intended a particular result. It is whether the employer's use subjected employees to discrimination within the covered employment context.
This does not mean that an unequal result automatically proves a violation. The statutory excerpt does not say what degree of difference is legally sufficient, which population should be compared, or how a causal connection must be shown. It gives the reader an effects-centered prohibition, not a complete statistical protocol. The distinction matters: an outcome inquiry is required by the wording, but the supplied text does not authorize the writer of a policy or a vendor to invent a single threshold and call it the statutory test.
What evidence belongs to the effects inquiry?
The first task is to identify the decision or employment condition. Section 2-102(L)(1) reaches recruitment, hiring, promotion, renewal of employment, selection for training or apprenticeship, discharge, discipline, tenure, and the terms, privileges, or conditions of employment. The same ranking system can therefore have different factual significance depending on whether it determines who is interviewed, informs a promotion recommendation, or influences discipline. A reviewer cannot interpret an outcome without knowing what decision the AI output entered.
The second task is to identify the relevant outcome. A hiring screen might affect who moves forward. A promotion tool might affect who is considered for advancement. A training-selection system might affect access to an opportunity that changes later employment prospects. A disciplinary recommendation might affect a person's conditions of employment or continued status. The Act's list keeps the analysis connected to an employment consequence rather than to an abstract score.
The third task is to establish the comparison that makes an effect visible. The supplied law does not specify the denominator, the comparison group, the time period, the sample size, or the measure to use. Those are not minor technical omissions. A conclusion can change depending on whether a reviewer compares applicants with applicants, selected employees with the eligible pool, recommendations with final decisions, or one period with another. The source text gives no permission to select whichever comparison produces the most convenient conclusion.
A fourth task is to connect the outcome to the employer's use of AI. A difference in results may reflect many parts of an employment process: the data supplied to the system, the design of a scoring rule, the way people interpret an output, or a decision made after the output. The amendment does not state a complete causation test. It does say that the prohibited conduct is the employer's use of AI in relation to the covered employment context. A serious review therefore has to preserve the path from input to output to employment action instead of treating an outcome statistic as self-explanatory.
These questions describe evidence, not a statutory checklist. Public Act 103-0804 does not set an approved fairness metric, a mandatory audit interval, a model-validation procedure, or a human-review safe harbor in the cited provision. An employer may choose to collect and preserve such information as a prudent risk-management practice. That can make the process more explainable and make later factual analysis possible. It should not be presented as though the amendment itself has already dictated one method.
The wording about “protected classes under this Article” creates another boundary. The AI provision does not provide a short list of demographic categories that can be checked once and then set aside. Section 2-101(E-1), in a definition of harassment, lists bases including race, color, religion, national origin, ancestry, age, sex, marital status, order-of-protection status, disability, military status, sexual orientation, pregnancy, unfavorable discharge from military service, citizenship status, and work-authorization status. Section 2-102(A) separately addresses unlawful discrimination and refers to citizenship and work-authorization status. Because the AI clause points to protected classes under the Article, the harassment-specific list is context, not automatically the complete AI inventory.
That distinction also affects evidence. A system may not collect a particular protected characteristic, and an employer may not see it in the model's input fields. The absence of a field does not establish that the employment process cannot affect people on a protected basis. At the same time, the source does not authorize a conclusion about any particular group or dataset without facts. The correct response is to keep the legal basis and the factual evidence separate: identify which protected-class question is being asked, then identify what the process actually shows.
Zip code is an input-path question, not merely an outcome statistic
The same sentence in Section 2-102(L)(1) separately prohibits an employer from using zip codes as a proxy for protected classes under the Article. A proxy is, in ordinary language, a variable used as a stand-in for something else. The statute names zip codes, but it does not define the term “proxy,” explain the required relationship between a zip code and a protected class, or identify particular geographic areas that are forbidden. Those missing details make the factual path especially important.
Suppose the hiring service described above receives an applicant's zip code. That fact alone does not tell us what the employer is doing with it. The zip code might be stored for administrative purposes and never enter the ranking. It might be used to estimate commute distance. It might be used as a filter, a weighting factor, a lookup, or a direct feature in the model. A recruiter might also use it outside the software to decide which applicants to prioritize. Each possibility creates a different factual record.
The proxy inquiry asks what role the zip code plays and whether it is being used as a stand-in for a protected class. It therefore follows the input and decision logic, not only the final distribution of outcomes. If an employer cannot explain where the geographic value enters, how the system transforms it, or whether a person uses it in a downstream decision, the missing explanation is itself relevant to understanding the process. It is not proof of a violation, but it prevents a reviewer from claiming that the input is harmless without examining its function.
The statute's choice to name zip codes matters. It does not say “geography” in general, nor does it provide a list of all variables that might correlate with protected status. A mailing address, travel distance, county, neighborhood label, or office location may raise related questions in a particular process, but the selected text does not allow those variables to be declared equivalent to zip code or automatically treated as prohibited proxies. Conversely, a reviewer should not assume that a zip code is irrelevant simply because it is a common business field. The relevant question is its use as a proxy in the covered employment setting.
Why the two substantive pathways cannot be collapsed
The effects and proxy pathways examine different parts of a process.
- The effects pathway starts with a decision or employment condition and asks what consequence the AI-assisted practice had for people on a protected basis.
- The proxy pathway starts with a named input—zip code—and asks whether the employer used it as a stand-in for a protected class.
The first is primarily an outcome question. The second is primarily an input-and-use question. They can overlap, but neither is reducible to the other.
Imagine that a zip code is used in a hiring model and that the resulting selections differ across protected groups. The same facts may support both lines of inquiry, but the records do different work. Outcome evidence can show how the selection process operated across groups. Input and model-use evidence can show whether zip code was part of the route by which the process operated. Removing the zip code might change the outcome, but that does not tell a reviewer whether the original use was a proxy. Conversely, a reviewer might identify a proxy concern without having enough data to measure an outcome difference, especially in a small or newly deployed process. The absence of a measured disparity is not the same thing as an answer to every input-path question.
The reverse situation is also possible. An AI-assisted process may show an effects concern even when no zip code was used. Other inputs, training data, design choices, or human decisions may be relevant facts. Because the Act separately names zip-code proxy use, a reviewer should not respond to an effects concern by searching only for geography. The proxy provision is important precisely because it is a distinct route to accountability, not a universal explanation for every unequal result.
The diagram does not rank the two pathways or supply a legal decision rule. It keeps their denominators and evidence sources from being mixed. The effects branch needs a clear employment consequence and a defensible comparison. The proxy branch needs a clear account of the zip code's role and its relationship to a protected class. A single “fairness score” would not, by itself, answer both questions.
Neutral purpose is relevant, but not conclusive
Employers often describe automated systems through their objective: efficiency, consistency, safety, retention, or better matching. Those objectives can be relevant facts. They may explain why a system was adopted and what decision it was intended to support. But the phrase “has the effect” means that purpose is not the end of the inquiry.
The same is true of a claim that the system treats everyone identically. Applying one formula to every person can still produce different consequences, and applying one formula does not reveal whether a named input was selected because it stood in for another characteristic. Uniform procedure and neutral language are not the same as a complete account of effect or proxy use. The source text does not say that a vendor's neutrality statement, an employer's good faith, or a human manager's review automatically resolves either pathway.
This is not an argument that every difference is unlawful or that every use of geographic information is a proxy. It is an argument about the order of reasoning required by the statutory language. First identify the covered employment use. Then ask what consequence it produced and, separately, how zip code was used. Only after those facts are clear can an informed legal analysis address the standards that apply.
There is also an interpretive question about the grammar of the proxy clause. It appears inside the provision titled “Use of artificial intelligence,” alongside the prohibition on AI that has a discriminatory effect. Yet the sentence says “or to use zip codes as a proxy” without repeating the same words about using artificial intelligence immediately before that phrase. The selected text therefore does not fully settle whether the proxy prohibition is limited to zip-code use through an AI system or reaches a broader employer practice within the AI subsection. Nor does it establish whether a measurable adverse outcome or an intent to discriminate must be shown before the proxy language applies. Rules or later adjudication may clarify those questions.
The safest reading for a present process review is neither maximal nor minimal. Do not declare that every geographic field is forbidden. Do not declare that zip code is harmless unless the final outcome looks equal. Instead, document the variable's purpose, point of entry, transformation, decision role, and relation to the protected-class question being examined. Label any conclusion that depends on an undefined term or absent proof standard as provisional.
Public Act 103-0804 thus makes the substantive analysis more demanding in a precise way. A reviewer cannot stop at intent, and cannot stop at outcome. The effects pathway asks how an AI-assisted employment practice treated people in a covered decision or condition. The proxy pathway asks whether a specifically named geographic input was used as a stand-in for a protected class. Those inquiries may converge in a real case, but they require different evidence and may remain at different levels of certainty. Once those two paths are kept apart, the remaining question is equally distinct: whether the employer told an employee that covered artificial intelligence was being used, and what the Department's rules will clarify about the notice's content and timing.