Tag: employment discrimination

  • The Rules That Never Came: How a January Deadline, an Eighteen-Day Rulemaking, and a Vanishing Web Page Left Illinois Employers Guessing at AI Notice

    The Rules That Never Came: How a January Deadline, an Eighteen-Day Rulemaking, and a Vanishing Web Page Left Illinois Employers Guessing at AI Notice

    Illinois HB 3773 has made undisclosed AI use in employment decisions a civil rights violation since January 1, 2026. The state agency charged with writing the implementing rules published them on May 15, withdrew them eighteen days later, and has not refiled — so employers must work out what compliant notice looks like while the duty itself is already enforceable.


    A Statute That Went Live Without Its Manual

    On August 9, 2024, Governor J.B. Pritzker signed House Bill 3773, making Illinois the second state in the country, after Colorado, to enact a dedicated law governing employers’ use of artificial intelligence [1][2]. The bill, Public Act 103-0804, amends the Illinois Human Rights Act, and its operative provisions took effect January 1, 2026 [1]. It reaches any employer with one or more employees in Illinois — or hiring in Illinois — with no small-employer exemption, and law firm analyses read the notice duty as extending to applicants as well as current employees [3].

    The Act creates three distinct civil rights violations. An employer may not use AI in a manner that has the effect of subjecting employees to discrimination on the basis of a protected class — even if the discrimination is unintentional. An employer may not use ZIP codes as a proxy for protected classes in AI-driven processes. And an employer may not fail to notify an employee when AI is being used in a covered employment decision [1]. The covered decisions sweep the full employment lifecycle: recruitment, hiring, promotion, renewal of employment, selection for training or apprenticeship, discharge, discipline, tenure, and the terms, privileges, or conditions of employment [1].

    The definition of artificial intelligence is deliberately broad: “a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments” — a definition that expressly includes generative AI [1][2]. Unlike other state laws governing AI in employment, Illinois imposed no requirement that AI be a substantial factor or the sole basis for a decision: if AI plays any role, notice is likely required [3].

    One more feature matters for everything that followed. The statute directs the Illinois Department of Human Rights (IDHR) to adopt rules specifying “the circumstances and conditions that require notice, the time period for providing notice, and the means for providing notice” [1]. In other words, the legislature bound employers to a notice duty and delegated the details to an agency that had not yet written them.


    Eighteen Days in the Register

    On May 15, 2026, IDHR published its proposed implementing rules in the Illinois Register, amending the Department’s procedures at 56 Ill. Adm. Code 2520 [4]. The proposal added a new Subpart J — “Use of Artificial Intelligence in Employment” — with two new sections: 2520.900 (Definitions) and 2520.910 (Notice), alongside an amendment to the Department’s recordkeeping section [4]. Publication triggered a 45-day public comment period, with written comments due June 29, and a public hearing scheduled for June 10, both in person in Chicago and via WebEx [4].

    Eighteen days after publication, the process stopped. On June 2, 2026, IDHR announced it was temporarily withdrawing and postponing the proposed rules, and the June 10 public hearing was “temporarily postponed” as well [5]. The agency’s full stated reason, as quoted by Ogletree Deakins: “IDHR is currently reviewing matters related to the proposed rulemaking and will provide updated information regarding next steps as it becomes available. Postponement is necessary to allow for continued collaboration with other state agencies.” [5] No revised timeline was provided.

    Four months later, the silence holds. No refiled version of the rules has appeared in the Illinois Register since the withdrawal, and tracking services that follow the rulemaking report the same absence [6]. As of this article’s publication, the announcement page IDHR pointed stakeholders to returns a “Page Not Found” error. The rules that were supposed to define the notice duty exist only as a withdrawn draft — but the duty itself has been enforceable since January 1.


    What the Withdrawn Rules Would Have Required

    Because the draft text was published in the Register, employers can still read exactly what IDHR had in mind — and law firms have treated the withdrawn rules as the best available preview of what refiled rules will demand [7]. The operative command in proposed Section 2520.910 was blunt:

    “An employer must not use artificial intelligence in any covered employment decision unless the employer provides notice of such use of artificial intelligence to its employees, including prospective employees, and their exclusive bargaining representatives.” [4]

    The Trigger: Influence, Not Decision-Making

    Several features of the draft explain why employers are not waiting for the refiling to begin work:

    • “Use” reaches the periphery, not just the decision. The draft defined use as “any instance in which the output of an artificial intelligence system influences or facilitates a covered employment decision” — language that captures AI playing even a peripheral role, including where the output is incidental or supplementary to human decision-making [4][7]. Seyfarth Shaw’s read is that unless the rules are revised, Illinois could cover more use cases than any other state AI employment law currently on the books [7]. The core logic is a pipeline test rather than a decision test: anything that touches how a covered decision gets made counts, whether or not the machine has the final say.
    • Vendors and staffing firms are on the hook too. An “agent” — any third party acting on an employer’s behalf to exercise control over recruitment, hiring, promotion, discipline, or the like — is itself an “employer” under the draft [4]. Depending on the control exercised, staffing firms, technology vendors, external recruiters, and other service providers could share compliance obligations with the employing entity [7]. The principle: whoever effectively controls the employment action cannot escape the notice duty by contracting it out.
    • Notice would be continuous and multi-channel. Current employees would receive notice annually and within 30 days of adopting a new or substantially updated AI system; prospective employees would receive it as part of the job notice or posting itself [4]. Notice would have to run through all applicable channels at once: employee handbook, physical posting in a conspicuous location, intranet or external website (with a conspicuous homepage link), and job postings [4].
    • Six required content elements. The draft notice would name the AI system’s developer, product name, and vendor; identify which covered employment decision the system influences; explain the system’s purpose and the categories of personal information or employee data it collects; list the types of job positions affected; provide a point of contact for questions; and state the right to request a reasonable accommodation [4]. Accessibility requirements follow: plain language, availability in the languages commonly spoken by the employer’s workforce, and reasonable accessibility for employees with disabilities [4].
    • The carve-outs are narrow. Notice would not be required for AI used for ordinary business operations — designing a job posting, generating promotional text — or for software that does not meet the statutory AI definition, or for AI features in a mixed system that are not used for covered decisions [4]. Everything else with an inference function in the employment pipeline stays in scope.

    Why the Withdrawal Changes Nothing, Legally

    The single most important point, repeated across every legal analysis of the withdrawal: postponing the rulemaking did not postpone the law. The three violations — discriminatory effect, ZIP-code proxy, failure to notify — are created by the statute, not the rules, and they have been in effect since January 1, 2026 [5][7]. Seyfarth’s summary of the posture is that “the contours of compliant notice in Illinois remain unclear” even as the obligations bind [7]. One research tracker’s headline calls the situation what it has become: “an AI notice duty with no rules behind it” [6].

    Why would an agency withdraw its own rules eighteen days into the comment period? IDHR said only that postponement was necessary for “continued collaboration with other state agencies,” and Ogletree Deakins notes the agency may be considering substantive changes before taking public input [5]. The practical effect is the same either way: enforcement risk now attaches to notice failures even though no regulator has defined what a compliant notice must contain, when it must be delivered, or through which channels. The statute ordered IDHR to answer exactly those three questions; the answers are missing.

    The Recordkeeping Preview

    The withdrawal also complicates the recordkeeping picture. The same proposal would have amended the Department’s records-preservation section, and the Register’s own small-business analysis concedes that “Procedures for tracking issuance of notices will be required” [4]. Employers who wait for final rules to build those procedures will have to reconstruct their notice history after the fact.

    The Federal Backdrop Does Not Pause the State Clock

    Employers hoping for relief from Washington have reason to wait indefinitely rather than act. On December 11, 2025, President Trump signed an executive order directing the Department of Justice to challenge state laws that regulate AI, including laws like HB 3773 [3]. But as Reinhart Boerner Van Deuren noted when the order issued, an act of Congress is typically necessary to preempt state laws, and the executive order’s own legal future is uncertain [3]. The core logic for employers is that preemption is a bet on a timeline no one controls, while the notice duty is a compliance task whose deadline already passed on January 1, 2026.


    What Employers Should Do Now

    Legal analyses converge on a short list of steps that make sense even without final rules:

    • Audit the AI footprint first. Employers should inventory the tools used across the employment lifecycle — recruiting, screening, monitoring, discipline — including AI embedded in ordinary HR software that may trigger notice obligations without anyone deciding to “adopt AI” [3]. The breadth of the statutory definition makes the inventory the hard part; every downstream judgment depends on knowing where inference systems touch covered decisions.
    • Draft notice against the six-element preview. Seyfarth recommends using the withdrawal window to expand notice and recordkeeping processes so that when the rulemaking resumes, employers “are not starting from scratch” [7]. A notice that names the system, the decision, the data, the positions, a contact, and accommodation rights satisfies the withdrawn draft — and cannot be less than what the statute requires.
    • Update nondiscrimination policies, especially on ZIP codes. The proxy prohibition operates independently of the notice duty, and ZIP-code inputs are common in ad targeting and screening tools; Reinhart flags policy review on exactly this point [3].
    • Build the notice log before the rules demand it. The Register’s small-business analysis already anticipates tracking issuance of notices as a compliance procedure [4]. A dated record of who was noticed, when, and through which channel is cheap insurance under a law that treats violations as civil rights violations.
    • Watch the Register, not just the press releases. IDHR pointed stakeholders to its legislative updates page, but the binding event will be the republication of proposed rules in the weekly Illinois Register, where the last version appeared on May 15 [4][5].

    The Vacuum Is the Policy Story

    Illinois chose a disclosure-based model for AI in employment: the law’s enforcement theory is that workers cannot contest decisions they cannot see. That choice makes the notice regime the load-bearing wall of the whole statute — and for nearly a year that wall has been held up by the statute alone, with the agency’s blueprint published, withdrawn, and unpublished again. A rule withdrawal is not deregulation; it is a delay of clarity, not of duty. The interpretive risk does not disappear while IDHR collaborates with other agencies. It shifts onto every employer making a good-faith guess at what “notice” means.

    For other states watching the second mover in this space, the lesson cuts both ways. Statutory text can bind before agencies speak, which argues for drafting laws that are self-executing in their essentials. But Illinois also shows the cost of delegating the details: eighteen days of proposed rules, a canceled hearing, a 404 where the announcement used to be, and thousands of employers writing notices to a standard that does not officially exist.


    References

    [1] Illinois General Assembly — [PRIMARY] Public Act 103-0804 (House Bill 3773), amending the Illinois Human Rights Act, 775 ILCS 5. Signed August 9, 2024; effective January 1, 2026.

    [2] DLA Piper — [SECONDARY] Mobley, Todd, and Eigen, Zev. “Illinois becomes second state to enact AI law for employers.” August 29, 2024.

    [3] Reinhart Boerner Van Deuren — [SECONDARY] Gentry, Michael J., and DeLange, Matt. “Illinois Employers Face AI Transparency Deadline Despite New Executive Order.” December 18, 2025.

    [4] Illinois Register — [PRIMARY] Volume 50, Issue 20. “Department of Human Rights: Notice of Proposed Amendments,” 56 Ill. Adm. Code 2520 (Subpart J — Use of Artificial Intelligence in Employment). May 15, 2026.

    [5] Ogletree Deakins — [SECONDARY] Colvin, Jennifer L., Francis, Simone R.D., Ochs, Danielle, and Zagger, Zachary V. “Illinois Postpones Proposed Regulations on AI in Employment.” June 4, 2026.

    [6] CASRAI — [SECONDARY] “Illinois HB 3773: An AI Notice Duty With No Rules Behind It.” Updated September 25, 2026.

    [7] Seyfarth Shaw — [SECONDARY] “Illinois Department of Human Rights Temporarily Withdraws Proposed Rules on Use of Artificial Intelligence in Employment.” June 8, 2026.


    AI automated and generated content. Human reviewed.

  • The Corroboration Rule: How a First-in-the-Nation No-Robo-Boss Law, a Connecticut Statute Gone Live, and a 1970s Consumer Law Are Extending AI Oversight From the Hire to the Fire

    The Corroboration Rule: How a First-in-the-Nation No-Robo-Boss Law, a Connecticut Statute Gone Live, and a 1970s Consumer Law Are Extending AI Oversight From the Hire to the Fire

    California now requires a human to verify an algorithm before it can discipline or fire a worker. Connecticut’s AI act took effect October 1 with a clause that makes automation no defense to a discrimination claim. And a January class action is testing whether the Fair Credit Reporting Act already regulates AI hiring scores. The perimeter of AI-at-work law has moved past the front door.


    One Week, Two Coasts

    On September 30, 2026, California Governor Gavin Newsom signed Senate Bill 947, the “No Robo Bosses Act of 2026,” authored by Senator Jerry McNerney of Pleasanton and sponsored by the California Federation of Labor Unions, AFL-CIO [1]. The law bars California employers from relying solely on automated decision systems (ADS) to fire or discipline workers, and it mandates human oversight and verification when an ADS assists those decisions. Its sponsor’s office calls it the first such law in the nation [1].

    The next day, a continent away, a different statute crossed its own threshold. Connecticut’s Artificial Intelligence Responsibility and Transparency Act — Public Act 26-15, formally titled “An Act Concerning Online Safety” and approved May 27, 2026 — reached the October 1, 2026 effective date that switches on its employment provisions [2]. One newly live section amends the state’s fair employment law so that using an automated employment decision tool “shall not be a defense against a complaint alleging a discriminatory practice” [2]. Another requires employers filing federal WARN layoff notices to tell the state Labor Department whether the layoffs relate to artificial intelligence or another technological change [2].

    Why does this pair matter? Because nearly every AI-at-work rule written since 2019 — New York City’s bias-audit law, Illinois’s video-interview act, Maryland’s facial-recognition consent rule — regulated the front door: how applicants are screened in. What changed in the last week of September 2026 is the perimeter. California’s law is aimed squarely at the back door, discipline and termination, and Connecticut’s duties follow the worker through the whole employment relationship: hiring, promotion, discipline, discharge, and terms of employment [2]. The governance question is no longer “may an algorithm score a candidate” but “may an algorithm be the last word on a person’s job.”

    “No worker should ever be fired or disciplined by a robo boss. Artificial intelligence systems have the potential to increase productivity, but they’ve also been prone to errors, bias, and misjudgments,” McNerney said at the signing [1]. “AI must remain a tool controlled by humans, not the other way around” [1].

    Behind the rhetoric sits a real market. The senator’s release points to more than 550 “bossware” products now sold to employers for managing workplaces [1]. The law’s premise is that some of those products are already making consequential calls about people — and that until last week, nothing in any state’s law required a named human to stand behind them.


    What a Robo Boss May No Longer Do

    SB 947 becomes operative on July 1, 2027, and it draws its line with two words: “solely” and “primarily” [3]. An employer may not rely solely on an ADS to make a disciplinary or termination decision. Where an employer primarily relies on ADS output, a human being must corroborate the proposed decision before it becomes final. The statute defines an ADS broadly — any computational process or software derived from machine learning, statistical modeling, data analytics, or artificial intelligence that generates output, such as a score, ranking, classification, or recommendation, used to assist or replace human discretionary decision-making in employment — while carving out ordinary infrastructure like spam filters, firewalls, antivirus software, calculators, and datasets [3].

    The corroboration step is where the law does its real work. The human reviewer must check the ADS output against the data that produced it or against other relevant material — supervisory or managerial evaluations, personnel files, employee work product, peer reviews, witness interviews, even relevant online customer reviews [3]. And the review must be substantive. In the words of one firm’s analysis: “A human reviewer must do more than rubber-stamp the ADS result,” because the reviewer “must exercise independent judgment to verify the output before the firing or disciplinary decision is final” [3]. If the output cannot be corroborated, or the reviewer determines it is inaccurate, incomplete, or misleading, the employer may not use it to discipline or terminate the employee at all [4].

    The design logic deserves spelling out, because it is a template other legislatures are watching. California did not outlaw the productivity tracker or the algorithmic flag; it relocated accountability. The employer keeps the efficiency of machine screening but must attach a named human judgment — and a record of it — to every consequence. The statute also bars three darker uses: deploying an ADS to violate labor, employment, safety, or civil-rights law; using one to infer a worker’s protected characteristics under the state’s Fair Employment and Housing Act; and using one to predict whether a worker will exercise legal rights and then taking adverse action on that prediction [4].


    The Notice and the Burden Shift

    The law’s enforcement design is quieter than its name. Violations carry a $500 civil penalty, enforced by the Labor Commissioner and public prosecutors — a figure too small to frighten a large employer on its own [3]. The sharper provision is procedural: once it is shown that an ADS was used in a disciplinary or termination decision, “the burden of proof shifts to the employer to prove either that it did not primarily rely on the ADS or that it complied with the law’s human-review and notice requirements,” and this shift applies broadly to “any civil action or administrative proceeding” [3]. The final law dropped an express private right of action, but that shift gives plaintiffs’ counsel something arguably more useful: in any later wrongful-termination or discrimination suit, the employer starts on the defensive on exactly the fact pattern it controls least [3].

    Workers also get information rights. When an employer primarily relies on an ADS to discipline or terminate, the affected employee must receive a separate, stand-alone written notice — in plain language, at the time of the decision — stating that the employer primarily relied on the ADS, confirming that a human reviewed and verified the output, naming a human contact for questions, and stating that retaliation for exercising protected rights is prohibited [3]. The employee may also request a meaningful, objective description of the data the system used, with personal information about customers, co-workers, and other individuals anonymized first [3]. Unionized workforces have one escape hatch: a qualifying collective bargaining agreement may waive the law’s provisions if it satisfies specified requirements, including protections from algorithmic management [4].


    Connecticut: The Statute That Went Live

    Connecticut’s Public Act 26-15 took a different architectural route: phased effective dates. As of October 1, 2026, two provisions are live. First, the amended fair employment statute: the use of an automated employment-related decision technology “shall not be a defense against a complaint alleging a discriminatory practice,” while a commission or court “may consider evidence of anti-bias testing or similar proactive efforts” — including the quality, efficacy, recency, scope, and results of such testing — when weighing the claim [2]. That structure creates a direct incentive: automation is no shield, but documented bias testing is now admissible mitigation. An employer who tested well has something to show; an employer who never tested has nothing to show.

    Second, a data point no other state collects. Every employer that serves a federal WARN layoff notice on Connecticut’s Labor Department must now disclose “whether the layoffs that are the subject of such written notice are related to the employer’s use of artificial intelligence or another technological change” [2]. For the first time, a state will accumulate a running ledger of AI-related mass layoffs — a number that until now has existed only in vendor estimates and consultancy surveys.

    The heavier employer duties arrive on October 1, 2027. Deployers of “automated employment-related decision technology” — defined as any technology that processes personal data and generates a prediction, recommendation, classification, ranking, or score that is a “substantial factor” in an employment-related decision — must disclose when an individual is interacting with the tool, and must provide written notice before the decision disclosing the tool’s purpose, its trade name, the categories of personal data it will analyze and how, and the sources of that data [2]. Enforcement is solely by the Attorney General as an unfair trade practice, with a sixty-day cure window for violations through the end of 2027 and no private right of action [2].

    One design choice deserves emphasis for the vendor market: the duty lands on the deployer, not the developer. As one Connecticut firm’s client alert put it, “using a vendor’s tool does not insulate you from liability” — the compliance burden falls on the employer who licensed the software [5]. The act does allow a developer to contractually assume the deployer’s notice duties, which converts a statutory obligation into a procurement negotiation [2]. Either way, the buyer now has to ask the question.


    The Old-Law Track: A 1970s Statute Meets a Billion Profiles

    While legislatures write new rules, litigators are testing whether the old ones already cover this ground. On January 20, 2026, Erin Kistler and Sruti Bhaumik filed a proposed class action against Eightfold AI in Contra Costa County Superior Court — described by the filing firms as the first U.S. lawsuit to accuse an AI hiring company of violating the Fair Credit Reporting Act, the federal statute that has governed who may assemble reports about people for employment decisions since the 1970s [6].

    The complaint’s theory is simple: if a company assembles and evaluates information about job applicants and sells the result to employers, it is a consumer reporting agency, whatever software it uses. Eightfold’s platform — built, according to the company’s own materials cited in the filing, on more than 1.5 billion global data points including profiles of more than one billion people — allegedly draws on public sources such as LinkedIn, Hoovers, Crunchbase, and GitHub to produce reports that rank applicants’ “likelihood of success” on a 0-to-5 scale, with lower-ranked candidates “often discarded before a human being ever looks at their application” [7].

    “There is no AI-exemption to these laws, which have for decades been an essential tool in protecting job applicants from abuses by third parties — like background check companies — that profit by collecting information about and evaluating job applicants,” the complaint reads [7].

    Eightfold disputes the framing. Spokesperson Kurt Foeller told Reuters the platform operates on data shared by candidates or provided by customers: “We do not scrape social media and the like. We are deeply committed to responsible AI, transparency, and compliance with applicable data protection and employment laws” [6]. The stakes are wide either way: one-third of Eightfold’s customers are Fortune 500 companies, and its tools sat behind applications by both named plaintiffs — Kistler applied to companies including PayPal, Bhaumik to companies including Microsoft, and neither was hired [6].

    The core logic of the FCRA route — and the reason it matters even if the case settles — is that it requires no new legislation. The statute’s duties of notice, consent, and the right to dispute predate the internet, and applying them means the burden of a “likelihood of success” score falls on the scorer rather than the scored. A hiring platform that must treat its 0-to-5 output as a regulated consumer report must also let the applicant see it and contest it. That is a different business model than an invisible ranking.


    From the Front Door to the Whole House

    Zoom out and the week’s news becomes a legible trend line. As of early October 2026, thirteen enacted rules govern AI in hiring, discipline, and firing across six states and New York City, and the pattern among them is generational [8]. The early laws — New York City’s 2023 bias-audit rule, Illinois’s 2020 video-interview act, Maryland’s facial-recognition consent requirement — targeted the hiring screen. The newest ones reach into the job itself: Illinois now covers discipline and discharge, California’s SB 947 is aimed squarely at them, and Connecticut’s notice duties attach to every covered decision from October 2027 [8].

    California also shows how layered this is becoming. SB 947, operative July 1, 2027, stacks on top of the Civil Rights Council’s automated-decision regulations that took effect October 1, 2025, and the Privacy Protection Agency’s automated-decision rules with a January 1, 2027 compliance date — three regimes, three clocks, one employer [3]. Newsom signed three other workplace AI bills the same day, including one requiring mass-layoff notices to say when AI or automation caused the layoff [8]. Colorado, meanwhile, has run the full arc: its 2024 AI Act was delayed, then repealed and re-enacted as a narrower disclosure law that starts January 1, 2027 [8].

    For employers, the compliance questions are converging on four, whichever state they operate in:

    • Which tools score people? Both new statutes define their reach by function — a score, ranking, or recommendation that materially influences an employment decision — not by brand name or vendor category [2].
    • What decisions do those tools feed? The newer laws follow the output past hiring into promotion, discipline, discharge, and layoffs [8].
    • Who checks the output? California now requires a corroborating human exercising independent judgment; Connecticut’s commissions and courts will weigh anti-bias testing in discrimination claims [3].
    • What is the worker told? Both states converge on disclosure — before the decision in Connecticut, at the decision in California — plus a route for the individual to see what data was used [2].

    The deeper principle underneath the two statutes and one lawsuit is the same one the Fair Credit Reporting Act encoded in 1970: consequential decisions about people require a process a person can see, question, and hold someone accountable for. What changed between September 30 and October 1, 2026, is not that principle — it is the number of employment decisions it now reaches.


    References

    [1] [PRIMARY] McNerney, Jerry. “Newsom Signs McNerney’s No Robo Bosses Act of 2026 Requiring Human Oversight of AI in the Workplace.” California State Senate, 5th District press release (September 30, 2026)

    [2] [PRIMARY] Connecticut General Assembly. Substitute Senate Bill No. 5, Public Act No. 26-15, “An Act Concerning Online Safety.” Approved May 27, 2026; employment provisions effective October 1, 2026 and October 1, 2027

    [3] [SECONDARY] Chapman, Kendra C., Susan W. Pangborn, James Smith, and Mariah Elizabeth Pavich Weaver. “No Robo Bosses: California Requires Human Review Before AI-Driven Workplace Discipline.” Kilpatrick Townsend & Stockton LLP (October 2, 2026)

    [4] [SECONDARY] “California Passes No Robo Bosses Act.” California Workplace Law Blog, Jackson Lewis P.C. (October 2026)

    [5] [SECONDARY] “Connecticut’s AI Act Takes Effect October 1: What Employers Need to Do Now.” Employment Law Letter, Shipman & Goodwin LLP (September 2026)

    [6] [SECONDARY] Godoy, Jody. “AI company Eightfold sued for helping companies secretly score job seekers.” Reuters (January 21, 2026)

    [7] [PRIMARY] Kistler, Erin, and Sruti Bhaumik v. Eightfold AI Inc., No. C26-00214. Class Action Complaint. Superior Court of California, County of Contra Costa (January 20, 2026)

    [8] [SECONDARY] Digital Applied Team. “AI Hiring and Firing Laws by US State: The 2026 Table.” Digital Applied (October 3, 2026)


    AI automated and generated content. Human reviewed.

  • The Verification Surge: How a Fake-Candidate Flood, a Multi-Nation Alert, and a $3.75 Million Biometric Settlement Are Rebuilding the Front Door of Hiring

    The Verification Surge: How a Fake-Candidate Flood, a Multi-Nation Alert, and a $3.75 Million Biometric Settlement Are Rebuilding the Front Door of Hiring

    Employers facing AI-generated applicants are deploying identity scans, camera pans, and VPN checks at the application stage — and discovering that every anti-fraud control they add creates discrimination and privacy exposure of its own.


    The Flood Comes First

    Three developments in 2026 describe a hiring system under strain from the outside in. The first is volume. More than a third of roughly 1,230 headhunters and hiring managers surveyed by the hiring software firm Greenhouse now spend half their workweek filtering for what the survey calls “spam and junk” submissions, many of them AI-generated, and the research firm Gartner projects that by 2028, one in four candidate profiles worldwide will be fake.[1] In one survey, 6% of candidates admitted to participating in interview fraud — either posing as someone else or having someone pose as them.[2]

    The second development is intent. This is not just resume inflation. In July 2026, the U.S. State Department and FBI, joined by agencies from Japan, South Korea, Australia, Canada, France, Germany, Italy, the Netherlands, New Zealand, and the United Kingdom, issued a joint alert warning that North Korean IT workers “pose an insider threat to companies and are involved in data exfiltration, cryptocurrency theft, and theft of sensitive information,” and that they employ “increasingly sophisticated methods, including the integration of AI, to obfuscate their identities and expand their activities globally.”[3] The Justice Department separately reported that North Korean IT workers had fraudulently obtained employment with more than 100 U.S. companies using stolen and fake identities, and that a nationwide enforcement action included searches of 29 “laptop farms” across 16 states.[4]

    The third development is cost. About 23% of U.S. companies surveyed by the background-check provider Checkr said hiring fraud cost them more than $50,000 in the past year — delayed projects, compliance issues, and rehiring costs — and for 10% of respondents, losses exceeded $100,000.[1]

    Together these describe the same structural condition the AI hiring lawsuits have: the screening layer is now the attack surface. When anyone can generate hundreds of polished applications, or walk into a video interview wearing someone else’s face, the problem is no longer who gets filtered out — it is whether anyone on either side of the transaction is real.


    The Countermeasures Arrive Early in the Process

    Employers are not waiting for regulation. The response has been to move identity verification from the end of hiring — the traditional background-check stage — to the front door: the application itself.

    • Biometric entry checks. The staffing industry’s first survey of its kind, covering more than 120 companies, found that more than 40% of employers hiring temporary workers now use some form of identity-verification software to screen applicants. The real estate firm PLACE began requiring applicants to verify identity through CLEAR using a government-issued ID and a selfie, and the platform flags hundreds of potentially suspicious applications each day.[1]
    • Device forensics without consent. Persona’s screening platform detects whether a candidate writes from a newly created email address, whether their computer clock is set to a different time zone than claimed, whether they are logging in through a VPN, or whether their IP address “is jumping around the globe faster than a person can travel.” A Persona client recently found a prospect impersonating a Coinbase employee using details copied from the employee’s verified LinkedIn profile.[1]
    • Live-action liveness tests. Interviewers ask candidates to remove virtual backgrounds, pan their cameras around the room, and wave a hand in front of their face — maneuvers meant to expose deepfake glitches, hidden coaches, and proxy interviewers.[5]
    • Outlet checks and other improv. The American Staffing Association is training recruiters to ask interviewees to turn their webcam toward a nearby electrical outlet to prove they are in the United States.[1]

    The logic driving these moves is not paranoia; it is the collapse of a decades-old sequencing assumption. Businesses have historically shared information about their staff and workflows with candidates before requesting identification or running background checks. Fraud at the application stage breaks that sequence: by the time a compromised hire is discovered — in one widely cited case, a security firm’s new software engineer had passed four video interviews and a background check before his malware attempt was flagged — the exposure has already happened.[6] Verification is therefore moving to the earliest possible moment. Persona’s CEO frames the shift bluntly: “Background checks are fundamentally about the past. The real risk is that you aren’t even the person that you claim to be.”[1]

    There is a trade-off embedded in this migration, and it lands on honest candidates. The signals driving the new verification — device metadata, IP trails, email age — “often [are] done without their input, marking a change from traditional background checks, which often require applicants to submit basic personal information and sign consent forms.”[1] An applicant who is exactly who they say they are now submits a selfie, opens their room to a camera pan, and accepts that their metadata is being read before anyone has decided whether to interview them.


    The Collision: Fraud Controls Are Discrimination and Privacy Risks

    Here is where the story turns from an HR problem into a policy problem. The tools being deployed to catch fakes are the same categories of tool — facial recognition, liveness detection, automated flagging — that have generated the largest discrimination and privacy enforcement actions of the past decade.

    The Federal Trade Commission’s Rite Aid action remains the clearest blueprint for what regulators expect. The FTC found that Rite Aid deployed AI-based facial recognition in hundreds of stores without reasonable safeguards, that the technology “falsely flagged consumers, particularly women and people of color, as shoplifters,” and that the company deployed the technology despite its vendor’s express disclaimer of “any warranty as to the accuracy or reliability of the product.” The resulting five-year ban and order required pre-deployment testing, ongoing monitoring, employee training, and vendor oversight.[7] An identity-verification stack at the application stage is the same technology under a different job title. Deploy it without bias testing and documentation, and an employer may trade fraud liability for discrimination liability.

    The Illinois Biometric Information Privacy Act (BIPA) shows the privacy half of the collision. HireVue — whose video-interview platform is used at the application stage — agreed in August 2026 to a $3.75 million settlement of claims that it collected voice and facial biometrics from job applicants without the consent and disclosures Illinois law requires.[8] The lesson generalizes beyond one vendor: any employer that routes candidates through a biometric scan as a condition of applying inherits BIPA’s consent, disclosure, and retention obligations, and BIPA carries statutory damages of $1,000 per negligent violation and $5,000 per intentional or reckless violation.[8]

    The core logic of the collision is that every fraud control is also a selection procedure. A liveness test that some faces fail more than others, or a device-forensics flag that correlates with poverty (shared devices, prepaid data plans, VPN use in regions with repressive networks), does not just reject fraud — it sorts people. Under Title VII, a selection procedure with an unjustified adverse impact is unlawful regardless of intent, and the four-fifths standard courts and the EEOC have long used to measure adverse impact applies to automated tools the same way it applied to paper tests.[9]


    The Governance Gap

    No state legislature has stepped into this space yet. Forty-six states have some form of deepfake legislation, but the existing statutes address elections, pornography, and voice cloning — not employment fraud.[10] A candidate who is deepfaked into a job is covered, if at all, by general fraud statutes never designed for the scenario. The patchwork of state AI hiring laws that does exist — the bias-audit and notice mandates discussed in earlier pieces in this series — regulates how employers screen candidates, not how they verify them.

    The verification layer is therefore being built almost entirely by market practice and vendor product design, with two exceptions at the federal level, and they point in opposite directions:

    • Government is demanding verification, not regulating it. The July 2026 joint alert does not just warn companies; it instructs them to “implement measures to counter the tactics” it lists — enhanced identity verification, strict review of identification documents, and in-person interviews where feasible.[3] The DOJ’s enforcement actions against facilitators — including U.S. nationals sentenced for running laptop farms — create “should have known” negligence exposure for employers who skip controls.[4] The compliance pressure runs toward more screening, not less.
    • The discrimination and privacy frameworks still bind. Title VII’s disparate-impact provisions are codified by statute and enforceable by private plaintiffs regardless of federal enforcement posture.[9] Illinois BIPA and its analogues impose consent duties on biometric collection with private rights of action.[8] Nothing about a fraud-control purpose suspends these obligations.

    The result is that employers are being pushed to verify harder by the same government that has pulled back from telling them how to verify lawfully. Jones Walker’s analysis calls the standard what it is: “The goal isn’t perfect detection; it’s documented reasonableness.”[10]


    What a Defensible Front Door Looks Like

    The emerging consensus among employment lawyers, Gartner, and the enforcement record converges on a handful of practices that reduce both fraud and legal exposure at once:

    • Test the tool before and after deployment. Rite Aid’s core failure was deploying facial recognition without assessing accuracy or demographic performance, then not monitoring it. Pre-deployment bias testing and documented four-fifths analysis are the floor.[7]
    • Layer verification, do not rely on one gate. Gartner recommends identity verification embedded at multiple stages plus anomaly alerts, with in-person interviews where the role permits — a control the candidates themselves prefer, with 62% of surveyed candidates saying they are more likely to apply to a position requiring in-person interviews.[2]
    • Tell candidates what is happening. Communicating how AI and verification are used, and defining acceptable AI use in applications, is both a Gartner recommendation and the cheapest available defense against BIPA and notice-and-consent claims.[2]
    • Vet the vendor, in writing. Rite Aid deployed its system despite the vendor’s disclaimer of accuracy — and the FTC held Rite Aid, not the vendor, responsible. Verification-vendor contracts now carry the compliance weight that hiring-AI vendor contracts did before Mobley.[7]
    • Escalate carefully on flags. A fraud flag is an adverse action in waiting. Employment counsel involvement before withdrawing an offer based on biometric or identity-proofing results is now standard advice.[6]

    The deeper principle behind all five is symmetrical accountability. Fraud tools and hiring tools are converging on the same pipeline, which means the duty-of-care standards built for one — vendor responsibility, bias testing, human checkpoints, documentation — are migrating to the other. An employer cannot credibly claim that its identity-verification stack is “just security” any more than a hiring vendor could claim it was “just software.” Both are selection systems now.


    The Unsettled Question

    What no one has resolved is whether verification can scale honestly. The fraud economics are asymmetric: building a fake persona costs almost nothing, while every honest applicant absorbs friction — selfies, camera pans, metadata exposure — that the fraudulent ones can route around. If the defense succeeds, the front door of hiring becomes a checkpoint, and the burden of proof of personhood shifts onto every job seeker before any human reads their resume. If it fails, the alternative is the world the DOJ and FBI describe: sanctioned regimes operating inside corporate networks on payroll. The policy question for 2027 is whether that burden gets allocated by law — with consent, contest, and appeal rights — or continues to be allocated by whatever vendors ship next quarter.


    References

    [1] Los Angeles Times (Bloomberg) — “Why 1 in 4 job applications could be fake by 2028” by Taylor Nicole Rogers (July 30, 2026) [SECONDARY]

    [2] Gartner, Inc. — “Gartner Survey Shows Just 26% of Job Applicants Trust AI Will Fairly Evaluate Them” (July 31, 2025) [SECONDARY]

    [3] U.S. Department of State and FBI — “Alert to Countries, Companies, and Other Entities Regarding North Korean IT Workers” (July 31, 2026) [PRIMARY]

    [4] U.S. Department of Justice — “Justice Department Announces Coordinated, Nationwide Actions to Combat North Korean Remote Information Technology Workers’ Illicit Revenue Generation Schemes” (June 30, 2025; updated) [PRIMARY]

    [5] WhatJobs News — “Employers Add Camera Panning and Hand-Wave Tests to Remote Job Interviews to Catch AI Fraud” by Clare Magness (August 31, 2026) [SECONDARY]

    [6] Whiteford, Taylor & Preston — “Are You For Real? An Increase in AI Deepfake Job Candidates and Employees Raises ‘Red Flags’ for U.S. Employers” by Lisa M. Brauner (September 17, 2026) [SECONDARY]

    [7] Federal Trade Commission — “Rite Aid Banned from Using AI Facial Recognition After FTC Says Retailer Deployed Technology without Reasonable Safeguards” (December 19, 2023) [PRIMARY]

    [8] Top Class Actions — “$3.75M HireVue Illinois BIPA class action settlement” (August 27, 2026) [SECONDARY]

    [9] Warden AI — “EEOC AI Hiring Guidance: Key Employer Compliance Rules” by Martyn Redstone (updated August 10, 2026) [SECONDARY]

    [10] Jones Walker LLP — “Your Next Data Breach May Start With a Job Interview: The Deepfake Candidate Problem” by Andrew R. Lee and Jeffery L. Sanches, Jr. (January 29, 2026) [SECONDARY]


    AI automated and generated content. Human reviewed.

  • The Veto Rematch: How a Narrower Bill, a Burden-Shifting Rule, and a September 30 Deadline Are Deciding Whether an Algorithm Can Fire a Californian

    The Veto Rematch: How a Narrower Bill, a Burden-Shifting Rule, and a September 30 Deadline Are Deciding Whether an Algorithm Can Fire a Californian

    One year after Governor Gavin Newsom vetoed the first “No Robo Bosses Act” as overly broad, a narrower successor has passed the California Legislature by wide margins and now sits on his desk. Senate Bill 947 would make California the first state in the nation to bar employers from relying solely on an automated decision system to fire or discipline a worker — and it would shift the burden of proof onto employers to show that a human actually checked.


    The Veto, and the Year Between

    On Sunday, August 30, 2026, the California Assembly approved Senator Jerry McNerney’s SB 947 on a 53-14 vote; the Senate followed 28-10 the next day, sending the bill to the governor’s desk [1]. If enacted, it would be, in its author’s words, “the first such law in the nation” — a statute written directly into the Labor Code rather than privacy or civil rights law [1]. Governor Newsom has until September 30, 2026 to sign or veto it [2].

    The backstory is a veto, and the bill is best read as a point-by-point answer to it. In 2025, both houses passed the original No Robo Bosses Act, SB 7, before Newsom returned it without his signature on October 13, 2025 [3]. His veto message conceded the core concern — “I share the author’s concern that in certain cases unregulated use of ADS by employers can be harmful to workers” — but rejected the method: “rather than addressing the specific ways employers misuse this technology, the bill imposes unfocused notification requirements on any business using even the most innocuous tools” [3]. He called its restrictions on automated decision systems (ADS) “overly broad,” singled out its limit on using customer ratings as input data, and pointed to forthcoming California Privacy Protection Agency regulations, writing that “before enacting new legislation in this space, we should assess the efficacy of these regulations to address these concerns” [3].

    The year between produced a bill engineered around each objection. McNerney reintroduced the measure on February 2, 2026, and the revisions tracked the veto letter with unusual fidelity [4]. The pre-notification requirement — the provision Newsom called “unfocused” — was removed in its entirety, replaced by notice only after an adverse decision [5]. A private right of action, included in earlier drafts, “was removed in the Assembly Privacy Committee” [5]. The final enrolled text narrows the protected class from “workers” to “employees,” dropping independent contractors and app-platform “deactivation” decisions altogether, and confines the bill’s central duties to two decisions: discipline and termination [6]. Even the definition of ADS now explicitly excludes spam filters, firewalls, antivirus software, and calculators — a direct rebuttal to the “most innocuous tools” line [6].

    What survived is the principle. McNerney frames it simply: “AI must remain a tool controlled by humans, not the other way around” [1]. Whether that principle, delivered in this narrower package, clears the governor’s desk is the question the next four weeks will answer.


    What the Bill Actually Requires

    SB 947 adds Part 5.5.5 to Division 2 of the Labor Code — sections 1520 through 1526.7 — and becomes operative on July 1, 2027 [6]. Its architecture rests on a distinction that does most of the work: the difference between relying solely and relying primarily on an automated system.

    • Three absolute prohibitions. An employer may not use an ADS to violate labor, occupational safety, employment, or civil rights law; to infer an employee’s protected status under California’s Fair Employment and Housing Act; or to predict and then take adverse action against an employee for exercising legal rights [6]. The logic of listing these separately is that they identify uses where the problem is not a biased score but the purpose of the system itself — profiling protected traits or punishing legal activity is unlawful no matter how accurate the model is.
    • The solely/primarily rule. An employer “shall not rely solely on an ADS when making a disciplinary or termination decision” — an outright ban on fully automated firings [6]. But if an employer primarily relies on ADS output, the bill requires something subtler: the employer must “direct a human to corroborate the decision using data that was collected or used to produce the ADS output or other relevant corroborating or supporting information,” with an explicit list of what counts — supervisory evaluations, personnel files, work product, peer reviews, and witness interviews “that may include relevant online customer reviews” [6]. Note what the list does: customer ratings, the exact input Newsom defended in his veto, remain legal — but as evidence a human weighs, not as an automated verdict.
    • A corroboration kill-switch. If the employer cannot corroborate the ADS output, or the human reviewer concludes the output is “inaccurate, incomplete, or misleading,” the employer “shall not use the ADS output to make a disciplinary or termination decision” [6]. This is the provision that converts human review from theater into a gate: a failed check does not merely create liability, it voids the basis for the action.
    • Data access. An employee subjected to an ADS-driven decision has the right to request, and the employer must provide, “a meaningful, objective description of the employee’s own data” the system relied on — anonymized as to customers and coworkers [6].
    • Post-use notice. At the moment an employer informs an employee of a discipline or termination decision that primarily relied on an ADS, the employer must deliver a separate, plain-language, stand-alone written notice stating that an ADS was primarily relied upon, that a human reviewed and corroborated the output, who to contact for more information, and that retaliation for exercising these rights is prohibited [6].

    Two scope details matter as much as the duties. “Employer” is defined to include every branch of state and local government — cities, counties, the University of California, California State University, community college districts, and school districts — plus labor contractors and farm labor contractors [6]. And a collective bargaining agreement can waive the entire part, but only if it “explicitly waives this part in clear and unambiguous terms” and itself “provides protection from algorithmic management” [6] — a design that lets unions trade the statute for negotiated protections, which is precisely what the San Francisco and Los Angeles teacher contracts did this year without waiting for Sacramento.


    The Enforcement Design: A Burden Shift and a $500 Penalty

    The final bill contains no private right of action — enforcement belongs to the Labor Commissioner and to public prosecutors, who may seek injunctive relief, punitive damages, attorney’s fees, and a civil penalty of five hundred dollars per violation [1][6]. For employers, the more consequential provision is procedural. Once it has been demonstrated that an ADS was used in a disciplinary or termination decision, “the employer must demonstrate that the employer did not primarily rely upon an ADS or that the employer complied” with the corroboration and notice sections [6].

    The core logic of that shift is a discovery problem. An employee cannot subpoena a vendor’s model, weights, or logs; the party that holds the documentation of whether a human actually reviewed the output is the employer. The bill responds by making compliance the employer’s burden to prove once ADS use is established — the same instinct behind the burden-shifting frameworks in anti-discrimination law, applied to the machine. The Assembly Judiciary Committee’s analysis was blunt about the cost-benefit: an employer will need “some other piece of evidence that supports the conclusion in the ADS output and not act on the ADS output alone,” and in the committee’s view, “this does not seem like a great burden on the employer given that the employee’s livelihood is potentially at stake” [5].

    The bill also insulates itself from overlap where it can. An employer who complies with the bill’s notice requirements is excused from “any substantially similar notice provisions” under other state law — but with a carve-out preserving the California Privacy Protection Agency’s automated decisionmaking technology regulations for businesses subject to the CCPA [6]. And the whole part is displaced where federal statute, regulation, or binding contract requires ADS use for aircraft development or national security, military, space, and defense work [6].


    The Coalition Math: 41 For, 55 Against

    As of the Assembly Judiciary Committee’s analysis, 41 organizations had registered support and 55 had registered opposition, with six more opposed unless amended [5]. The support column is labor: the sponsor, the California Federation of Labor Unions, AFL-CIO — 1,300 unions and 2.3 million members, in its president’s description — alongside teachers, nurses, firefighters, and, unusually for a labor bill, the Electronic Frontier Foundation and privacy groups like Oakland Privacy and TechEquity [1][5]. The opposition column is nearly everyone who manages people: the California Chamber of Commerce, Uber, SHRM California, the hospital, restaurant, grocer, and farm bureau associations — and, tellingly given the bill’s definition of employer, the League of California Cities and the county associations [5].

    The substantive fight is over what the bill regulates. Opponents argue it “will drive up costs for consumers and employers because it would impose significant compliance burdens and any misstep would lead to costly litigation for even the smallest of employers,” and that it “will have an undesired chilling effect on technology and make it that much harder to develop the very tools that can help combat bias in decision making” [5]. A second objection is practical: in workplaces where supervisors cannot observe every worker, productivity-tracking tools “could never be relied upon unless every single output is independently corroborated,” effectively barring their use in personnel decisions [5].

    Supporters answer with evidence that the problem is not hypothetical. A 2025 Resume Builder survey of employers found that 60 percent of managers use AI systems to make critical decisions about workers — raises, terminations, layoffs — and that of those, 20 percent “let AI make final decisions without any human input or oversight” [5]. The bill’s sponsor points to algorithmic speed quotas in Amazon warehouses that forced workers to skip bathroom breaks, and to vendors selling “predictive” fraud flags built on a worker’s debt history and spending habits [5]. There are now more than 550 “bossware” products on the market [1], and the National Institute for Workers’ Rights (an advocacy organization) estimates that roughly four in ten U.S. workers already receive schedules or assignments from automated systems [7]. Its director, Jason Solomon, frames the stakes in due-process terms: “When an algorithm ends someone’s livelihood and no one can say why, we’re living in an unfair, sci-fi version of ‘at will’ employment” [7].


    The Regulatory Stack It Lands On

    SB 947 does not arrive on empty ground, which was the heart of Newsom’s veto argument. California’s Fair Employment and Housing Act regulations, effective October 1, 2025, already make it an unlawful employment practice to discriminate through automated-decision systems [5]. The Privacy Protection Agency’s ADMT regulations take effect January 1, 2027 — the same window as this bill’s July 1, 2027 operative date — and will require pre-use notice, opt-out rights, and access to information when automated technology replaces human decisionmaking in significant employment decisions [2]. The bill’s drafters accepted that overlap rather than fight it: section 1526.4 keeps the CPPA regime fully intact for covered businesses [6]. A signed SB 947 therefore means California employers run two parallel ADMT regimes — a labor-law framework looking backward at discipline and termination, and a privacy-law framework looking forward at deployment.

    The same stacking is happening nationally, with California moving from hiring regulation into the employment relationship itself. Illinois now requires notice when AI is used in employment decisions and applies its Human Rights Act to algorithmic discrimination [2]. Connecticut’s SB 5 requires disclosure of AI’s role in qualifying mass layoffs starting October 1, 2026, and plain-language disclosure when applicants interact with automated systems from October 1, 2027 [2]. Colorado’s replacement AI Act, SB 26-189, imposes notice and disclosure duties for consequential decisions from January 1, 2027 [2], and New York City has required bias audits of automated hiring tools since 2023 [2]. Meanwhile, the most comprehensive federal bill in discussion would preempt state regulation of AI model development but preserve state authority over employment-specific uses — meaning this state layer, whatever Newsom decides, is likely to remain the operative law [2].


    What to Watch

    • The September 30 decision. Newsom must act by month’s end [2]. The bill answers each specific objection in his veto letter — pre-notification gone, contractors out, innocuous tools excluded, customer ratings preserved as corroborating evidence [3][6]. The open question is whether his broader instruction — assess the CPPA regulations first — reads as a condition or a preference. A second veto would be the third consecutive AI workplace bill rejected on generality grounds.
    • The gaps, if it is signed. The National Institute for Workers’ Rights’ framework asks for five elements: human ownership of the decision, notice and transparency, data access, appeal, and anti-retaliation [7]. SB 947 delivers a version of the first three plus anti-retaliation, but has no appeal right and no private right of action — both stripped on the way to passage [5][1]. The brief notes the analogy is not exotic: federal fair-credit-reporting law has long required notice, a copy of the report, and a chance to respond before an employer acts on a background check, and Seattle and Minnesota already require notice, records access, and human review before app-based workers can be deactivated [7].
    • The union waiver in practice. Because a collective bargaining agreement can waive the part in exchange for algorithmic-management protections [6], the statute functions as a floor for non-union workplaces and a bargaining chip for organized ones. Watch whether large unionized employers trade it away — and whether the 2027 California teacher-contract round, where AI language must be re-won, cites it as the baseline.
    • Copycats and litigation pressure. If signed, California becomes the first state with a solely-reliance ban on algorithmic firings, and the corroboration-plus-burden-shift design is portable legislation [1]. At the same time, courts are already extending liability sideways — in Mobley v. Workday, discrimination claims were allowed to proceed against both an AI vendor and the employers using its screening tools [2]. A statute that tells employers exactly what documentation survives scrutiny will find a ready audience among compliance departments writing the first generation of algorithmic-firing procedures.

    References

    [1] California State Senate, Office of Senator Jerry McNerney — [PRIMARY] “Legislature Approves No Robo Bosses Act of 2026 to Ensure Human Oversight of AI in the Workplace.” Press release, August 31, 2026. (Final votes, sponsor, first-in-nation claim, enforcement design, and bossware estimate.)

    [2] Epstein Becker Green — [SECONDARY] “Workplace AI Regulation in 2026: How Employers Can Navigate the Changing Legal Landscape.” September 1, 2026. (Governor’s deadline, FEHA and CPPA ADMT regulations, Illinois, Connecticut, Colorado, New York City frameworks, federal preemption carve-out, and pending litigation.)

    [3] Office of Governor Gavin Newsom — [PRIMARY] Newsom, Gavin. “Veto Message, Senate Bill 7.” October 13, 2025. (The governor’s objections: unfocused notification requirements, overly broad restrictions, and the directive to assess CPPA regulations first.)

    [4] Crowell & Moring LLP — [SECONDARY] Meade, Kris D., Katie Aber, Jillian Ambrose, Jacob Canter, and Adina Nelson. “California SB 947 (‘No Robo Bosses Act’): New Proposed Guardrails on Use of Automated Decision Systems in Employer Discipline and Termination Decisions.” February 11, 2026. (Reintroduction date and the bill’s relationship to the veto letter.)

    [5] California State Assembly, Committee on Judiciary — [PRIMARY] “Bill Analysis, SB 947 (McNerney) — Employment: Automated Decision Systems.” 2026. (Amendment history including removal of pre-notification and the private right of action, the corroboration analysis, support and opposition registers, and both sides’ arguments.)

    [6] California Legislature — [PRIMARY] “SB-947 Employment: Automated Decision Systems.” Enrolled September 4, 2026. (Full enacted text: Labor Code sections 1520–1526.7, definitions, prohibitions, corroboration, notice, enforcement, waivers, exemptions, and July 1, 2027 operative date.)

    [7] National Institute for Workers’ Rights — [ADVOCACY] “AI Is Not the Boss of Me: A Meaningful Human Oversight Requirement for Workplace Decisions.” Press release, August 11, 2026. (Advocacy organization; the five-element oversight framework, workforce exposure estimates, and the Seattle, Minnesota, FCRA, IBM, and FedEx precedents.)


    AI automated and generated content. Human reviewed.

  • The 2026 State AI Hiring Law Wave: A Patchwork Takes Shape

    The 2026 State AI Hiring Law Wave: A Patchwork Takes Shape

    As federal AI guidance retreats, California, Illinois, and Colorado are building the first enforceable framework for AI in employment decisions — and a billion-application class action is testing its limits.


    The Federal Floor Holds, Even as Guidance Falls Away

    In early 2025, the Equal Employment Opportunity Commission and the Department of Labor effectively rescinded their AI guidance on workplace discrimination, following an executive order aimed at reducing government oversight of AI and promoting U.S. leadership in the field. The rollback removed certain government-endorsed best practices but did not alter the underlying legal architecture. Title VII of the Civil Rights Act, the Americans with Disabilities Act, and the Age Discrimination in Employment Act remain fully in force, and employees and job applicants continue to sue over AI-related discrimination.[1]

    The core principle is straightforward: employers are liable for discriminatory outcomes produced by AI tools even when the bias is unintentional, and even when the tool was built by a third-party vendor. As the EEOC has consistently maintained, “employers remain responsible for ensuring compliance with federal EEO laws when they use software, algorithms, or AI to make employment decisions, even when those tools are designed or administered by a vendor.”[2] You cannot outsource compliance. The federal floor is sticky — it persists regardless of which administration holds office or which guidance documents are withdrawn.

    What the federal rollback did create was a vacuum. With Washington stepping back, state legislatures and state civil rights agencies moved in. By mid-February 2025, 27 bills regulating AI in employment had already been introduced in state legislatures.[1] Three jurisdictions — California, Illinois, and Colorado — have now enacted the first enforceable AI-specific employment rules in the nation, each taking a distinct regulatory philosophy.


    Three States, Three Approaches

    California: Regulations as Litigation Leverage (Effective October 1, 2025)

    California’s approach is the most litigation-driven. On June 27, 2025, the Office of Administrative Law approved final regulations from the California Civil Rights Council clarifying how the Fair Employment and Housing Act applies to automated-decision systems (ADS) in employment. The regulations took effect October 1, 2025.[3]

    The regulations define an ADS broadly as “a computational process that makes a decision or facilitates human decision making regarding an employment benefit” that may use “artificial intelligence, machine-learning, algorithms, statistics, and/or other data processing techniques.”[4] This covers resume screening tools, interview analysis software, job ad targeting systems, and gamified skills assessments.

    What makes California’s approach distinctive is that the regulations do not mandate specific compliance steps — they do not require bias audits or impact assessments. Instead, they create litigation leverage. The regulations explicitly state that evidence of anti-bias testing (or lack thereof) is “relevant to any such claim or available defense,” and that courts and agencies should consider “the quality, efficacy, recency, and scope of such effort, the results of such testing or other effort, and the response to the results.”[4] The practical effect: an employer who cannot demonstrate proactive bias testing faces a significant disadvantage in discrimination litigation. The regulations also extend recordkeeping requirements from two to four years and expand the definition of “agent” to include third-party AI vendors acting on behalf of employers.[5]

    Illinois: Disclosure and Strict Liability (Effective January 1, 2026)

    Illinois HB 3773, signed by Governor J.B. Pritzker on August 9, 2024, amends the Illinois Human Rights Act to regulate AI in employment. The bill passed both chambers with near-unanimous support — 106-0 in the House and 57-0 in the Senate — suggesting that disclosure and anti-discrimination requirements enjoy bipartisan consensus even in a polarized era.[6]

    The law has two core pillars:

    • Mandatory disclosure: Employers must notify applicants and employees whenever AI is used in “recruitment, hiring, promotion, renewal of employment, selection for training or apprenticeship, discharge, discipline, tenure, or the terms, privileges, or conditions of employment.” The notice must use plain, readable language, be available in languages commonly spoken by the workforce, and be accessible to employees with disabilities.[7]
    • Strict liability for discriminatory effects: The law prohibits AI use that has the effect of discriminating against protected classes, regardless of whether the discrimination was intentional. Intent is not a defense. A discriminatory outcome alone can establish a violation.[8]

    The law also specifically prohibits using ZIP codes as proxies for protected classes when AI tools process employment data. ZIP codes strongly correlate with race and national origin, and AI systems trained on historical hiring data can learn these proxy patterns and replicate them at scale — even when the data scientist never intended to use a protected variable.[6]

    Notably, unlike New York City’s Local Law 144 or Colorado’s AI Act, Illinois does not mandate formal bias audits or impact assessments. But because the law imposes strict, outcome-based liability, independent bias audits become the most practical way to detect disparate impact before it becomes a violation.[7]

    Colorado: The Comprehensive Model (Effective June 30, 2026)

    Colorado’s Senate Bill 24-205, the Consumer Protections for Artificial Intelligence, was signed into law on May 17, 2024, and was originally set to take effect February 1, 2026. After compromise negotiations collapsed in August 2025 — technology companies objected to liability provisions — Governor Polis signed SB 25B-004, delaying implementation to June 30, 2026.[5]

    The Colorado AI Act is the most comprehensive state-level AI employment law enacted to date. It regulates “high-risk” AI systems — those that make or are a substantial factor in making “consequential decisions” affecting consumers, including employment, housing, credit, education, healthcare, and insurance. The law imposes obligations on both developers and deployers of these systems.[9]

    For deployers (employers), the law requires:

    • Risk management policy and program: Employers must implement a structured approach to identifying and mitigating algorithmic discrimination risks.
    • Annual impact assessments: Deployers must complete and annually review impact assessments for each high-risk AI system they use.[9]
    • Consumer notice and appeal rights: Employers must notify individuals when an AI system makes or substantially influences a consequential decision about them, and provide an opportunity to correct incorrect data and appeal adverse decisions through human review.[9]
    • Attorney General reporting: Deployers must disclose any discovered algorithmic discrimination to the Attorney General within 90 days.[9]

    The law grants enforcement authority exclusively to the Attorney General, with no private right of action. Violations are treated as unfair trade practices under the Colorado Consumer Protection Act. An affirmative defense is available for entities that comply with a nationally or internationally recognized AI risk management framework — such as the NIST AI Risk Management Framework — and take corrective measures.[9]

    The delay reveals the political tension at the heart of AI regulation. The Trump Administration is actively discouraging state AI regulation, and the AI Action Plan directs OMB to consider a state’s AI regulatory climate when making federal funding decisions. House Republicans initially included a 10-year moratorium on state and local AI regulation in the One Big Beautiful Bill Act — removed from final legislation, but a clear signal of continued federal opposition to state AI laws.[5]


    The Courtroom Frontier: Mobley v. Workday

    While state legislatures write the rules, a federal courtroom in San Francisco is drawing the battle lines for enforcement. Mobley v. Workday, Inc., filed in 2023 in the U.S. District Court for the Northern District of California, is the first major class action to test whether an AI vendor can be held directly liable for employment discrimination.

    The plaintiff, Derek Mobley, alleges that Workday’s AI-powered applicant screening tools discriminate on the basis of race, age, and disability, violating Title VII, the ADA, the ADEA, and state anti-discrimination laws. The novel legal question: whether Workday — which is not itself an employer but sells hiring software to employers — can be held liable as an “agent” of those employers under federal civil rights law.[10]

    In July 2024, Judge Rita Lin denied Workday’s motion to dismiss the “agent” theory claims, allowing them to proceed to discovery. The court found that Workday’s software “is not simply implementing in a rote way the criteria that employers set forth, but is instead participating in the decision-making process by recommending some candidates to move forward and rejecting others.”[10] The court distinguished AI hiring tools from ordinary software like spreadsheets or email systems, which have not been “delegated responsibility” for hiring decisions.

    The court’s policy rationale is striking. Judge Lin used a hypothetical to illustrate the enforcement gap: if a software vendor intentionally creates a tool that screens out applicants from historically Black colleges and universities, unbeknownst to employers, no party could be held liable without the agency theory. The court construed anti-discrimination laws broadly to close this loophole.[10]

    In May 2025, the court granted conditional certification of the ADEA claims, allowing notice to be issued to what could be one of the largest collectives ever certified. Workday has represented that 1.1 billion applications were rejected using its software tools during the relevant period, meaning the collective could potentially include hundreds of millions of members.[11]

    The case reached a new procedural milestone in May 2026, when Magistrate Judge Laurel Beeler ruled on discovery disputes. The court denied plaintiffs’ request for Workday’s bias-testing data, finding it protected by attorney-client privilege because Workday’s attorneys curated the data and the overall purpose was to provide legal advice. But the court ordered production of Workday’s EEO-1 and OFCCP documents, reasoning that because Workday uses the same AI tools internally as it markets to customers, those documents are relevant to its knowledge of potential demographic disparities.[12]

    The discovery ruling carries a practical lesson: companies using AI in hiring should structure bias testing under the direction of legal counsel to preserve privilege — but should also recognize that their own internal use of marketed AI tools creates broad discovery exposure.[12]


    The Transparency Challenge and Proxy Discrimination

    Underlying all three state laws and the Mobley litigation is a common technical problem: AI hiring systems can be so complex that even their developers struggle to explain how specific decisions are made. This “black box” problem makes it difficult for employers to audit outcomes or defend them in court.[13]

    The deeper issue is proxy discrimination. AI systems trained on historical hiring data learn the discriminatory patterns embedded in that data. Even when protected characteristics like race or gender are explicitly removed from the model, AI identifies proxy variables that correlate with those characteristics:[13]

    • Graduation years serve as a proxy for age, allowing the model to penalize older candidates.
    • ZIP codes correlate with race and national origin — which is why Illinois specifically banned their use as proxies.[6]
    • Employment gaps disproportionately affect women who took parental leave, recreating gender bias without the model ever seeing a gender field.

    The system is not explicitly programmed to discriminate — it is learning discriminatory patterns embedded in historical data and replicating them at scale. This is why outcome-based liability, as Illinois has adopted, matters: it shifts the focus from intent to effect, recognizing that algorithmic discrimination can occur without any human deliberately designing it.[7]


    Compliance Strategies for the 2026 Patchwork

    For multi-state employers, the emerging patchwork creates a compliance floor problem: meeting the requirements of one state does not guarantee compliance with another. The practical response is to build to the highest standard and apply it everywhere.

    • Inventory AI usage: Identify every AI system currently in use or anticipated for use across the employment lifecycle — recruitment, screening, interviewing, hiring, promotion, discipline, and termination. Colorado’s definition of “consequential decisions” reaches beyond traditional hiring and firing.[8]
    • Conduct privileged bias audits: Test AI tools for disparate impact across protected classes using the Four-Fifths Rule from the Uniform Guidelines on Employee Selection Procedures — if the selection rate for a protected group falls below 80% of the rate for the highest-selected group, adverse impact is indicated.[2] Structure these audits under legal counsel to preserve attorney-client privilege, as the Mobley discovery ruling demonstrated.[12]
    • Implement disclosure protocols: Develop plain-language notices for applicants and employees when AI is used in employment decisions, available in multiple languages and accessible formats, as Illinois requires.[7]
    • Build human oversight into decisions: California’s regulations require “meaningful human oversight” — someone trained and empowered to override AI recommendations. AI should function as a tool that augments human judgment, not a sole or substantial decision-maker.[13]
    • Scrutinize vendor agreements: Require AI vendors to provide algorithmic transparency, confirm anti-discrimination compliance, and carefully review warranty, disclaimer, and indemnity provisions. The Mobley case has established that vendors can face direct liability — meaning vendor contracts now carry litigation risk in both directions.[1]
    • Adopt a recognized risk management framework: Colorado offers an affirmative defense for entities complying with a nationally or internationally recognized AI risk management framework, such as the NIST AI RMF. Adopting this framework proactively builds a legal shield.[9]

    The Road Ahead

    The 2026 state law wave represents a fundamental shift in AI employment regulation: from voluntary guidance to enforceable mandates, from federal leadership to state experimentation, and from employer-only liability to a framework that reaches AI vendors themselves. The federal government’s deregulatory posture has not stopped this trend — it has accelerated it by creating a vacuum that states are filling.

    The Mobley v. Workday case will likely define the outer boundary of vendor liability for years to come. If the agent theory survives summary judgment and reaches trial, every AI hiring vendor in the country will face a new calculus: their algorithms are no longer shielded by the claim that they merely implement employer criteria. They are participating in employment decisions, and they can be held accountable for the outcomes.

    For employers, the message is clear: the legal landscape has changed faster than most compliance programs have adapted. The companies that treat AI hiring tools with the same rigor as any other employment decision — documenting, testing, disclosing, and maintaining human oversight — will be positioned to navigate the patchwork. Those that do not will find that the federal floor, the state mandates, and the courtroom have all moved without them.


    References

    [1] Husch Blackwell LLP — “AI and Workplace Discrimination: What Employers Need to Know after the EEOC and DOL Rollbacks” by Owen Davis (February 7, 2025) [SECONDARY]

    [13] HR Defense Blog — “AI in Hiring: Emerging Legal Developments and Compliance Guidance for 2026” (November 2025) [SECONDARY]

    [6] Seyfarth Shaw LLP — “New Illinois AI Law Requires Employee Notice, Affirms Existing Employer Nondiscrimination Duties” (August 30, 2024) [SECONDARY]

    [8] Ogletree Deakins — “Illinois Steps Up AI Regulation in Employment: Key Takeaways for Employers” by Simone R.D. Francis, Benjamin W. Perry, and Lauren N. Watson (August 29, 2024) [SECONDARY]

    [7] Warden AI — “Illinois Human Rights Act Amendment (HB 3773) Guide” (advocacy organization) [ADVOCACY]

    [9] Colorado General Assembly — “SB24-205: Consumer Protections for Artificial Intelligence” (2024) [PRIMARY]

    [5] Seyfarth Shaw LLP — “Colorado Postpones Implementation of AI Law as California Finalizes New Employment Discrimination Regulations and Illinois Disclosure Law Set to Take Effect” (September 12, 2025) [SECONDARY]

    [3] California Civil Rights Department — “Civil Rights Council Secures Approval for Regulations to Protect Against Employment Discrimination Related to Artificial Intelligence” (June 30, 2025) [PRIMARY]

    [4] Paul Hastings LLP — “New California Regulations on Employers’ Use of AI to Make Decisions Go Into Effect Oct. 1, 2025” by Felicia A. Davis, Sara B. Tomezsko, Ankush Dhupar, Dan Richards, and Christopher G. Cho (August 12, 2025) [SECONDARY]

    [10] Seyfarth Shaw LLP — “Mobley v. Workday: Court Holds AI Service Providers Could Be Directly Liable for Employment Discrimination Under ‘Agent’ Theory” (July 19, 2024) [SECONDARY]

    [11] Law and the Workplace — “AI Bias Lawsuit Against Workday Reaches Next Stage as Court Grants Conditional Certification of ADEA Claim” (June 2025) [SECONDARY]

    [12] Duane Morris LLP — “California Federal Court Clarifies Limits on AI Bias Testing and Applicant Data Disclosure in Mobley v. Workday” by Gerald L. Maatman, Jr., Adam D. Brown, and Elizabeth G. Underwood (June 2, 2026) [SECONDARY]

    [2] EmployArmor — “EEOC AI Hiring Guidance 2026 & Federal AI Laws” (advocacy organization) [ADVOCACY]


    AI automated and generated content. Human reviewed.