What Counts as AI, and Who Falls Within the Rule?
Before asking whether an automated employment practice creates a discriminatory effect, uses a zip code as a proxy, or requires notice, a reader has to identify the system and the employment setting with some care. This chapter examines how Public Act 103-0804 defines artificial intelligence, which employer and worker relationships the Illinois Human Rights Act recognizes, and where the amendment places AI across the employment lifecycle. The practical result is a scope test: classify the system by what it does, locate the decision it touches, and preserve the text's uncertainty where recruitment, applicants, or borderline tools do not fit neatly into one label.
The law defines AI by what it infers and produces
Section 2-101(M) does not define artificial intelligence by a product name, a programming language, or a claim that a vendor makes about its software. It defines AI as a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs. Those outputs can include predictions, content, recommendations, or decisions that can influence physical or virtual environments. The definition expressly includes generative artificial intelligence.
The sequence matters. A system receives something—such as an application, a work history, a performance record, a prompt, or another set of data. It has an objective, which may be stated openly or built into the system's design. It then infers how to produce an output. The output is not limited to a final yes-or-no decision. It may be a prediction about an applicant, a recommendation to a manager, a ranking of candidates, content that becomes part of an evaluation, or a decision generated by the system itself.
That functional approach is significant for employment practices because many systems divide the decision into stages. One tool may screen résumés and return a shortlist. Another may recommend which employees should receive a training opportunity. A third may summarize performance records for a supervisor. A fourth may generate interview questions or draft an evaluation that a human later adopts. In each case, the system's position in the chain can differ, but the statutory definition directs attention to the machine-based inference and its output rather than asking only who clicked the final button.
This does not mean that every automated operation is conclusively AI under the amendment. A fixed rule that produces the same result from the same input without inference could raise a different classification question. A spreadsheet can calculate a total; a database can retrieve a record; a scheduling program can apply a stated sequence of rules. The supplied text does not provide a technical threshold for deciding when those tools become machine-based inference. It also does not create a model-size exemption, a vendor exemption, or a distinction between software an employer builds and software it buys. Borderline cases therefore require facts about how the system actually works, not a conclusion based on its marketing label.
The definition's reference to outputs that can influence physical or virtual environments also helps explain why workplace software can matter even when no machine directly acts on a person. A recommendation can influence a hiring decision. A generated performance summary can influence discipline. A predicted likelihood of success can influence who is invited to an interview. “Influence” is part of the definition's description of the output; it does not by itself prove that a particular employment use violates the law. It does, however, make it difficult to dismiss an AI system merely because a human remains somewhere in the process.
Generative AI is included, not substituted for AI
Section 2-101(N) adds a separate definition of generative artificial intelligence. It describes an automated computing system that, when prompted with human prompts, descriptions, or queries, can produce outputs that simulate human-produced content. The examples include text, images, multimedia such as audio or video, and other content that people would otherwise produce by human means.
The relationship between the two definitions is important. Generative AI is expressly included within the broader definition of AI, so the law does not treat “generative” as the only form worth considering. A system that predicts whether an applicant should advance and a system that drafts an evaluation may both be relevant, but for different functional reasons. The first produces a prediction or recommendation; the second produces content that may enter a decision process.
Consider a hypothetical employer that asks a generative system to summarize interview notes and suggest follow-up questions. The tool may never issue a formal hiring decision. Yet the content it produces can shape what a recruiter sees, what the recruiter asks next, and how the candidate is described to a decision-maker. The statutory definition does not settle whether that particular use violates the Act. It does establish why a content-generating tool cannot be excluded from the initial scope analysis simply because it does not calculate a score.
The same principle applies to systems used after hiring. A generative assistant might draft a performance review, a prediction tool might flag employees for possible promotion, and a recommendation engine might prioritize access to training. The relevant question is not whether the software looks like a conversational assistant or a traditional analytics product. It is whether the system's input, inference, and output connect to one of the employment contexts named by the AI provision.
The definition therefore supplies a recognition rule, not a conclusion of liability. It tells a reader what to inventory. It does not specify how accurate a system must be, how much human involvement is enough to change the analysis, or whether a particular vendor's technical architecture meets the definition. Those questions remain factual or interpretive.
The diagram is a recognition model. It shows why an output that is preliminary, advisory, or generative can still deserve attention. It does not say that every item reaching the last box is automatically a violation. The next scope question is whether the relevant employer, person, and employment decision fall within the Act's terms.
Scope follows the employer relationship and the decision stage
The AI provision operates inside the Illinois Human Rights Act, so its use of “employer” and “employee” must be read against the definitions in Section 2-101. That reading makes coverage broad, but not limitless.
For the ordinary employer category, the Act includes a person employing one or more employees within Illinois during 20 or more calendar weeks within the calendar year of, or the year preceding, the alleged violation. The definition also separately includes the State and its political subdivisions, municipal corporations, and other governmental units or agencies without regard to the number of employees. A party to a public contract and a joint apprenticeship or training committee are also included without regard to employee count.
The practical point is not that every organization in Illinois is automatically covered. It is that employee count alone is not a safe shortcut. The Illinois Department of Human Rights states on its legislative page that, unlike some federal laws, the Illinois Human Rights Act applies to most employers regardless of size. That is agency explanation rather than the full legal test, but it is a useful warning to small employers and to organizations that outsource their technology. Outsourcing the screening does not, by itself, answer whether the employer's use falls within the provision; the cited definitions do not create a vendor-based exemption.
The employee definition is also specific. It includes an individual performing services for remuneration within Illinois for an employer, an apprentice, and an applicant for an apprenticeship. The section separately says that, for purposes of the sexual-harassment subsection, “employee” also includes an unpaid intern who meets stated conditions concerning education, compensation, supervision, displacement of regular employees, and the benefit of the work. That special unpaid-intern rule is expressly tied to the sexual-harassment provision. It should not be carried into the AI-use provision as though the supplied text had said that every unpaid intern is an employee for every purpose.
The relationship between applicants and the AI provision deserves the same care. Section 2-102(L)(1) lists recruitment and hiring among the contexts in which the employer's AI use is addressed. That makes candidate-facing processes plainly relevant to the scope discussion. But the operative language also refers to AI that has the effect of subjecting “employees” to discrimination, and paragraph (2) speaks of notice to “an employee.” The excerpt does not separately explain every question about when an applicant is treated as an employee for these paragraphs. A careful account should not declare that all applicants are covered in exactly the same way, nor should it conclude that recruitment tools are outside the provision. The wording creates an interpretive boundary that facts, rules, and later legal interpretation may have to resolve.
The employment lifecycle is wider than the hiring moment
Section 2-102(L)(1) applies “with respect to” a long chain of employment decisions and conditions. The list begins with recruitment and hiring, then reaches promotion, renewal of employment, selection for training or apprenticeship, discharge, discipline, tenure, and the terms, privileges, or conditions of employment. The broader discrimination provision in Section 2-102(A) uses a similar set of employment actions.
This sequence prevents a narrow mental model in which AI law means only résumé screening. Recruitment and hiring are important because they are familiar entry points for scoring and ranking tools. But a system used to decide who is considered for promotion can alter a career after hiring. A system that selects workers for training can shape access to future opportunities. A tool that recommends discipline can affect a person's current conditions of employment. A renewal or tenure recommendation can determine whether an existing relationship continues.
The lifecycle also changes the facts that matter. In recruitment, the relevant output may be a ranking or recommendation. In promotion, it may be a performance prediction or a summary of past evaluations. In training selection, it may be a forecast of who will benefit. In discipline, it may be a flag or generated narrative that makes one worker appear more problematic than another. The statutory list does not tell us that each output is unlawful; it tells us where the employer's AI use can connect to a legally protected employment action.
The phrase “terms, privileges, or conditions of employment” is particularly important because it reaches beyond a single formal decision. A tool that allocates desirable shifts, access to opportunities, or workplace benefits may affect the conditions under which people work even if no one receives a document labeled “hire,” “promote,” or “discharge.” The supplied text does not define the outer limit of that phrase for AI systems, so the factual connection still matters. But the structure cautions against treating only final status changes as consequential.
Nor does the provision say that the AI system must be the final decision-maker. The definition names predictions and recommendations, and the employment list is framed around the employer's use of AI in relation to a decision or condition. A human manager may accept, modify, or reject an output. That human role can matter to the facts and to later legal analysis, but the supplied amendment does not state a blanket rule that human review removes the AI use from consideration. A system can influence a decision without formally owning it.
A useful scope review can therefore proceed in four connected questions. First, what system received what input, pursued what explicit or implicit objective, and generated what kind of output? Second, which employer category and worker relationship are involved? Third, which listed employment decision or condition did the output touch? Fourth, which aspects remain uncertain because the text does not define the technical boundary or the applicant relationship in more detail? These questions are narrower than a general inquiry into whether an organization “uses AI,” and they produce a record that can support the separate effects, proxy, and notice inquiries.
A scope example without a legal conclusion
Imagine a small Illinois employer uses a commercial application-ranking service during recruitment. The service reads résumés, infers a ranking, and sends a shortlist to a recruiter. The employer does not need to have trained the system for the initial scope questions to arise. The system appears to match the statutory description of a machine-based process that receives input and produces a recommendation or prediction. The employer's size does not, by itself, resolve coverage because the statutory employer definition includes the 20-week condition and other categories, while the IDHR describes the Act's coverage as broad.
The use is connected to recruitment and hiring, two contexts named in Section 2-102(L)(1). The shortlist is not necessarily the final decision, but it can influence who receives further consideration. Those facts are enough to identify a potentially covered AI-assisted employment use for further analysis. They are not enough to decide whether the practice has a discriminatory effect, whether a zip code operates as a proxy, or what notice rules require.
Change one fact. Suppose the employer uses a generative tool to turn interview notes into candidate summaries, while a recruiter makes the final recommendation. The content-producing system still deserves an initial scope review because generative AI is included in the statutory AI definition and content is one of the outputs named in the broader definition. Whether the particular use falls within all parts of Section 2-102(L) would depend on the employment context and the facts, not on the label “assistant.”
Change another fact. Suppose the organization uses an automated scheduling tool that simply applies a fixed rule supplied by a manager and never infers from input how to generate a prediction, recommendation, content, or decision. The supplied Act does not resolve whether that tool falls within the AI definition. That uncertainty should be recorded rather than filled by either extreme: “all software is AI” or “only advanced models count.” The definition is broad in function, but it is still a definition with elements.
The scope analysis is valuable because it prevents the three substantive questions from being asked about the wrong object. A fairness comparison of the wrong output cannot explain the employer's actual process. A search for a geographic proxy is incomplete if the reviewer does not know which inputs were used. A notice record is meaningless if the organization has not identified which AI use the notice was supposed to describe. Scope is not a technical preface to the legal issue; it determines the facts to which the legal issue attaches.
Public Act 103-0804 thus does two things at once. It uses language broad enough to reach predictions, recommendations, content, and decisions across a continuing employment relationship. At the same time, it leaves important questions to the actual system, the actual employer-worker relationship, and the actual decision stage. Once those boundaries are mapped, the next problem becomes more precise: whether the covered use produced a discriminatory effect and whether its inputs—especially a zip code—created a separate proxy concern.