Jason Rosensweig on Civil Rights, the Notice Regime, and Illinois' 2026 AI Hiring Law

Editor's note

When we recorded this episode with Jason, the IDHR rulemaking process was still proceeding. On June 2, 2026, IDHR announced it was temporarily postponing the rulemaking process, including the June 10 public hearing, stating that the postponement was necessary to allow for continued collaboration with other state agencies, and did not provide a revised timeline. Critically, the underlying statutory obligations under HB 3773 remain in effect; the law took effect on January 1, 2026.

We met Jason Rosensweig at a moment when most state AI regulation is being debated in the abstract — as omnibus bills, federal preemption fights, and standalone statutes that imagine algorithmic harm as a new kind of problem requiring a new kind of regulator. Jason is one of the few people in the country who is actually writing the rules that will govern AI inside a sixty-year-old civil rights regime.

When Illinois decided to regulate AI in hiring, it made a choice most states didn't. It didn't pass a standalone AI statute. It didn't mandate bias audits. It didn't adopt the Colorado consumer-protection frame or the New York City audit model. Instead, Illinois put AI squarely inside the Illinois Human Rights Act, treating algorithmic discrimination as a civil rights problem, enforced by the same agency that handles housing, credit, real estate, and public accommodations complaints. House Bill 3773 took effect on January 1, 2026, and the Illinois Department of Human Rights, where Jason works, is now drafting the implementing rules under what will be known as Subpart J.

Jason came to that work along an unusual path. He holds a PhD from the University of Chicago's Committee on Social Thought, where his dissertation examined what it takes for people to live together in a free and pluralist society. He earned a BA in comparative literature and an MA in French literature from Stanford and was a visiting scholar in the Faculty of Philosophy at Cambridge. Outside IDHR, he teaches Freedom, Equality, and Justice in American Government as an adjunct professor with the Northwestern Prison Education Project, and has served as a commissioner on the Illinois Commission on Discrimination and Hate Crimes since 2021. The conversation he is steering sits at the intersection of political philosophy, civil rights jurisprudence, and software engineering — and Jason is one of the few people in state government anywhere with serious credentials in the first two and an honest, working interest in the third.

The Architectural Choice: Civil Rights, Not AI Statute

The most consequential decision in HB 3773 isn't a clause. It's the chassis. Illinois's legislature chose to amend the Human Rights Act rather than pass an omnibus AI bill, and Jason was direct about why. The Senate in particular wanted to extend existing civil rights doctrine to a new technology, not invent a parallel regulatory regime. The substantive innovations were modest by design. Most of what the law does, doctrinally speaking, was already there. Disparate impact has been a tool of civil rights enforcement for decades. Liability for the use of technology in employment decisions has been settled law for years. As Jason put it, an employer has never been able to say, "I didn't fire that person — the computer did." HB 3773 puts that principle in bright letters and adds one new piece: a notice requirement.

That architectural choice has consequences most national commentary has missed. It means courts will read this law against decades of civil rights jurisprudence rather than against the much shorter and more contested record of AI-specific statutes elsewhere. It means the enforcement agency, IDHR, already has institutional muscle memory for the analytical work the law requires. And it means the doctrinal centerpiece — disparate impact — applies. Intent doesn't matter. Effect does. An employer using an AI tool that disproportionately disadvantages a protected class is liable even if no one in the company set out to discriminate.

The Zip Code Story: From a One-Issue Bill to a Notice Regime

One of the most interesting threads in the conversation was the origin story of the bill itself. The legislation we now call HB 3773 didn't start as a comprehensive AI-in-employment statute. It started as a single-issue bill introduced by Representative Jamie Andrade, narrowly focused on prohibiting the use of zip codes as a proxy for race. IDHR helped redraft it into something broader, and that's how the notice regime emerged — as, in Jason's phrasing, the most effective and reasonable way to go about it. The zip-code reference stayed in the final law as a remnant of the original draft.

That detail tells you the doctrinal lineage. This law is downstream of redlining, not downstream of Silicon Valley product safety. The reference point is Chicago's South Side and the long history of credit, housing, and employment decisions made through ostensibly neutral geographic variables that turned out to track race. Jason walked through the distinction between disparate treatment and disparate impact using exactly that framing: disparate treatment is a man and a woman applying for the same job and the man getting it because he's a man. Disparate impact is the bank that doesn't lend on the South Side and tells you it isn't about race — it's just about zip codes. AI bias, in his view, is almost always going to be a disparate-impact problem, not a disparate-treatment one.

The Notice as Data Infrastructure

The most counterintuitive feature of HB 3773 — for anyone whose mental model of AI hiring law is set by New York City's Local Law 144 or Colorado's SB 24-205 — is that Illinois requires neither a bias audit nor a formal impact assessment. The compliance mechanism is disclosure. If an AI system influences or facilitates a covered employment decision, the employer has to tell people. IDHR is currently drafting the rules for what that notice must say, in what form, and on what timeline.

Why notice instead of audit? Jason explained the chicken-and-egg problem at the core of civil rights enforcement in a world of opaque algorithms. People who experience algorithmic harm rarely know to complain about it. If you don't know an AI tool was involved in your rejection, your discipline, or your missed promotion, you don't know to file a charge. Without complaints, there's no evidentiary record. Without an evidentiary record, neither IDHR nor the courts can evaluate disparate impact in actual cases. The notice solves that. It surfaces the existence of the tool, which is the precondition for everything else.

This reframes how to think about the law's purpose. The notice isn't only a compliance hook for individual employers. It's data infrastructure. As Jason put it, one way to think about the notice requirement is that it's about surfacing more information about what's actually happening in order to make better policy and laws going forward. Illinois is not pretending to know in advance which AI uses are discriminatory. It is building the visibility infrastructure to find out.

Influence or Facilitate: Why the Reach Is Wider Than People Think

The draft Subpart J rules define a covered use of AI as any instance in which the output of an AI system influences or facilitates a covered employment decision. Jason walked through what that means with a concrete example. Imagine a tool that scores a resume or a video interview. The score isn't the decision — a human looks at it and makes the call. But if the tool's scoring mechanism systematically misreads, say, the facial expressions of candidates with disabilities, the discrimination is in the input to the human decision, not the output of the machine. The law covers that.

The example that stayed with us most came from the stakeholder process. Nurses' unions surfaced that some hospitals are using AI to analyze electronic health records — pulse-ox drops, response times, time-to-bedside — to generate performance assessments of nurses. The nurses weren't told. They didn't consent. Their evaluations were being shaped by a system they couldn't see, couldn't contest, and didn't know existed. Jason described this as a perfect example of the kind of sector-specific use case that emerged only because IDHR talked to people who would never have come up in a tech-policy stakeholder process: nurses, staffing firms, restaurant associations, hospital administrators.

For Illinois employers, the doctrinal frame is now this: if you use AI anywhere in the employment lifecycle — sourcing, scoring, scheduling, monitoring, performance synthesis, promotion modeling, retention — the question is no longer whether the AI made the decision. It's whether the AI shaped what a human decision-maker saw. That's a much wider net, and most compliance teams we work with are scoped for the narrower one.

Vendors, Employers, and the Politics of Liability

Jason described a multi-stakeholder engagement process that involved roughly fifteen bilateral meetings and several larger group convenings. The conversations spanned unions, healthcare provider associations, retail manufacturers, the restaurant association, the chamber of commerce, staffing firms, and tech industry groups. One of the most common employer pushbacks was a recurring question: couldn't you put this burden on the developers instead of on us?

Jason's answer was doctrinally clean. That isn't what the legislature did, and there are good reasons. Employers make the employment decision. Employers contract for the tool. Employers control whether to deploy it on a specific role, against a specific candidate pool. The statute puts liability where decision-making authority sits.

Vendor accountability isn't absent. It's just mediated through the contract. Jason was direct about what he wants AI vendors selling into Illinois to understand: as they're building products and selling them and negotiating contracts, there is accountability for safety and responsibility down the road, and they need to think about it. The operational consequence is that 2026 is the year vendor contracts in Illinois start to look meaningfully different. Indemnification provisions will tighten. Audit rights will be clarified. Disclosure obligations will be passed through. And there will be a real divergence between vendors who can produce defensible documentation of their training data, their evaluation methodology, and their adverse-impact testing — and vendors who cannot. The market is about to start sorting them.

The Pace Problem and the "Bad Faith" Argument

The most provocative moment in the conversation came when we asked Jason directly about the pace problem. The standard critique of state AI regulation is that government simply cannot move at AI's speed. Compute thresholds in the EU AI Act and elsewhere are anchored to training-compute numbers that algorithmic efficiency improvements and Chinese open-source releases have already begun to render obsolete. By the time rules are promulgated, the technology has moved.

Jason was unsentimental about it. The argument that this is moving so fast that we can't do anything about it, he said, is not always made in good faith.

The structural answer is in the design. Illinois regulates uses, not mechanisms. It regulates hiring and firing and promotion decisions, not compute thresholds or specific architectures. The technology can change underneath; the prohibited outcomes don't. That is the design principle worth carrying forward. The most durable AI regulation is not going to be the one that names the latest model architecture. It is going to be the one that names the decisions people actually care about.

What Civil Rights Brings to the AI Conversation

Throughout the conversation, Jason returned to a theme we think the AI policy community has not fully absorbed: civil rights enforcement is a perspective, not just a venue. It is a way of thinking about technology that asks different questions than the questions tech policy typically asks. It asks who bears the cost of error. It asks whether the rationality of the system as a whole obscures harm to specific people. It asks what visibility the harmed party has into the decision that affected them. And it brings to those questions a doctrinal toolkit that has been refined across decades of housing, credit, and employment litigation.

Jason was clear that civil rights people and philosophers should not be the only voices in the AI conversation. But he was equally clear about what gets lost when they aren't part of it. If you focus only on how to make the technology faster or better, you miss some of the broader effects and other consequences. That's the underlying philosophical commitment that runs through HB 3773 and the work IDHR is doing under it. It is also the perspective that will increasingly shape AI policy at the state level as federal civil rights enforcement scales back and states step in to fill the vacuum — a shift Jason described as probably here to stay.

Why This Episode Matters for the AI Ecosystem

Most conversations about AI in employment happen at the level of policy abstraction: ethics frameworks, model audits, responsible AI principles. This episode is different because Jason brings the perspective of someone who is actually doing the doctrinal and operational work of translating a civil rights statute into a regulatory regime that will shape how thousands of Illinois employers, and the vendors selling into them, operate.

For founders, operators, and executives in the Chicago AI ecosystem, this episode offers three critical takeaways:

1. The architecture of an AI law shapes everything that follows. Whether your state regulates AI as a technology problem or as a civil rights problem will determine which doctrines apply, which agency enforces, and what evidence is required. Illinois made a deliberate choice, and other states will follow this pattern, particularly as federal pullback pushes more civil rights enforcement to state agencies.

2. The reach of "influence or facilitate" is the compliance frontier. If your operations rely on AI anywhere in the employment lifecycle, the question is no longer whether the AI made the decision. It is whether the AI shaped what a human decision-maker saw. That is a much wider net than most compliance programs are scoped for, and 2026 is the year to widen the scope.

3. Vendor contracts are about to do a lot of work. Because Illinois places liability on employers, not developers, vendor contracts will be where the real liability transfer happens. Indemnification, audit rights, disclosure obligations, and documentation of adverse-impact testing are about to become competitive differentiators. The vendors who can produce defensible documentation will win accounts. The ones who cannot will start losing them.

Listen to the full episode on AI in Chicago on Spotify or wherever you get your podcasts. For more insights from the operators, builders, and thinkers scaling applied AI from the heart of the Midwest, subscribe and follow at chicagopodcast.ai.

AI in Chicago is hosted and produced by Khullani Abdullahi, Founder of Techné AI.