The Direct Answer: When AI Involvement Triggers Disclosure Obligations in Termination Decisions
As of September 2026, employers in a growing number of states must disclose when artificial intelligence tools play a substantive role in termination decisions. California's SB 947, often described with the phrase "robots can recommend, but real people must pull the trigger," establishes the clearest framework: automated decision systems may screen, score, or rank employees, but a qualified human reviewer must exercise actual, independent judgment before an adverse employment action like termination takes effect. Where an automated decision system contributed meaningfully to that decision, affected employees are generally entitled to notice that such a system was used, a description of the data and criteria it relied upon, and in several jurisdictions, an explanation of the logic involved and an opportunity to appeal to a human decision-maker.
Also worth reading: What are the vendor contract requirements under Colorado's new ADM law for employers using automated decision tools in HR? · What are the joint pay assessment requirements under the EU Pay Transparency Directive and how must employers comply? · What does AI bias auditing in HR actually involve in 2026, and how should employers prepare for state audit requirements?
The obligations are not uniform. Colorado's revised AI Act requires multi-stage notices for consequential decisions in employment, including adverse actions such as dismissal, demotion, and discipline. Connecticut's SB 435 imposes its own notice and explanation duties for employers using AI in employment decisions affecting terms and conditions of work. Illinois, New York City's Local Law 144 (in effect for bias-audited automated employment decision tools since July 2023), and a handful of additional jurisdictions add layered requirements. An employer with remote employees in five states may be subject to five different disclosure regimes simultaneously, which is precisely why compliance programs built for a single jurisdiction keep failing audits.
It is worth being candid about the state of play: there is no single federal statute that directly governs AI termination disclosures in 2026. The patchwork means the practical answer to "what must I disclose?" is "it depends on where your employees sit, where your vendor's systems operate, and whether the AI output was a but-for contributor or merely background information." Employers who assume the answer is uniform across their workforce create avoidable legal exposure.
How and Why These Rules Emerged
The regulatory momentum traces back roughly five years. Early enforcement actions focused on algorithmic hiring bias, but by 2024 and 2025, regulators and plaintiffs' attorneys turned attention to the back end of the employment lifecycle: performance monitoring, automated productivity scoring, and termination recommendations. Surveillance-driven productivity metrics, sometimes generated by AI notetakers, keystroke analytics, and scheduling algorithms, began appearing in termination documentation, and courts started asking whether employees had any visibility into how those metrics were produced.
California's SB 947 responded directly to this concern. Its central design principle is human agency: automated systems may inform decisions, but a human must review relevant, accurate data, be trained on the system's limitations, and retain authority to override the recommendation. Colorado's revamped AI Act took a developer-and-deployer approach, requiring impact assessments and layered notices at different stages of the decision process. Connecticut's omnibus legislation folded employment AI into a broader consumer-data framework, imposing duties on employers using AI for decisions with material effects on workers. The unifying theory across all three is transparency as a precondition to accountability: if a worker does not know an algorithm influenced their dismissal, neither a regulator nor a court can evaluate whether the decision was discriminatory, accurate, or procedurally fair.
There is also an anti-discrimination rationale that should not be understated. Under existing Title VII, ADEA, and ADA doctrine, employers remain fully liable for discriminatory outcomes produced by vendor-supplied tools. Disclosure requirements function as an early-warning system; they force documentation that either supports the decision's legitimacy or exposes its weaknesses before litigation rather than after.
What Disclosures Must Actually Contain
Disclosure content requirements vary, but the substantive categories recur across jurisdictions. Employers should expect to provide most or all of the following where AI contributed to a termination decision:
| Disclosure Element | California SB 947 | Colorado AI Act | Connecticut SB 435 |
|---|---|---|---|
| Notice that an automated decision system was used | Required for consequential decisions | Required, multi-stage | Required for covered employment decisions |
| Description of data sources and criteria | Required | Required in impact assessment and notice | Required |
| Explanation of decision logic | Required upon request | Required for adverse decisions | Required |
| Human review and override record | Mandatory human decision-maker | Deployer must ensure reasonable care | Required documentation |
| Appeal or correction pathway | Employee may contest data accuracy | Notice must include appeal rights | Correction opportunity provided |
| Timing | At or before adverse action | Before the decision takes effect | Within statutory notice window |
Practical Steps for Employers Right Now
Start with an inventory. You cannot disclose what you do not know you are using. Catalog every tool in the HR stack that scores, ranks, flags, or recommends: applicant tracking systems, performance management platforms, productivity analytics, attrition-risk models, and yes, AI notetakers whose summaries feed review files. For each tool, determine whether its output reaches a termination decision-maker before the decision is made. If the answer is yes, the tool is in scope for disclosure regimes in Colorado, California, and Connecticut.
Second, rewrite your termination workflow to satisfy the human-review standard. The decision-maker must receive more than the algorithm's output; they must review the underlying data, be trained on the tool's known error rates and limitations, and document independent reasoning. A checkbox that says "reviewed AI recommendation" without evidence of actual review will not survive scrutiny. Third, build the notice templates now. Pre-drafting disclosure letters, appeal procedures, and data-correction forms takes weeks; doing it after a claim arrives takes months and costs far more.
Fourth, pressure-test your vendors. Require contractual commitments that they will provide impact assessments, bias audit results, decision logic documentation, and notice of material model changes. Under Colorado's framework, a deployer's good-faith reliance on a developer's documentation can mitigate liability, but only if that documentation actually exists and is current. Finally, train HR business partners and managers. Most disclosure failures in early enforcement actions stem not from absent policies but from line managers who never followed them.
Comparison: Compliance Approaches and Their Trade-offs
Employers generally choose among three implementation models, each with real drawbacks worth acknowledging honestly.
| Feature | Manual Compliance | Vendor-Managed | Platform-Based Compliance Management |
|---|---|---|---|
| Upfront cost | Low ($0–$15K internal time) | Included in vendor fees | $30K–$150K+ annually for mid-size employers |
| Jurisdiction coverage | Prone to gaps | Limited to vendor's tool | Multi-state, tool-agnostic |
| Audit trail quality | Inconsistent | Vendor-controlled | Centralized and timestamped |
| Speed of updates to new laws | Slow | Dependent on vendor roadmap | Near-real-time regulatory monitoring |
| Risk of over-reliance | Low but error-prone | High if vendor documentation is thin | Moderate; still requires internal ownership |
Common Mistakes That Create Liability
The most frequent error is the "rubber-stamp" human review. Employers designate a manager, have them glance at an AI-generated performance score, and call it human oversight. Regulators and courts increasingly look for evidence of genuine independent reasoning: contemporaneous notes, consideration of mitigating context, documented questions about the data. A review completed in under two minutes for a decision affecting someone's livelihood invites skepticism.
Second, employers routinely misjudge when AI "contributes" to a decision. If an attrition-risk model puts an employee on a list and a manager terminates them citing "performance concerns" drawn from an AI-generated dashboard, that is AI involvement even if no one consciously treated the model as decisive. Third, companies forget remote and multistate employees. A Colorado resident working remotely for a Texas-based employer is covered by Colorado's notice requirements; jurisdiction follows the worker, not the headquarters.
Fourth, disclosures get buried. A clause on page nine of an onboarding acknowledgment signed two years before termination does not satisfy contemporaneous notice obligations in most covered jurisdictions. Fifth, employers over-collect in the name of transparency, publishing model internals or trade secrets they are not legally required to reveal. Disclosure obligations concern the decision's logic and data categories, not your proprietary source code. Finally, organizations neglect AI notetakers and meeting-summarization tools, which quietly generate the written records that later substantiate termination decisions, and which Mayer Brown and other commentators have flagged as an emerging, undermanaged risk area.
Timing: When to Act and Key 2026 Dates
Act now, regardless of where you operate. California SB 947's operational requirements and human-review mandates are in force, and Colorado's revamped AI Act provisions with multi-stage notices have taken effect with enforcement following a phase-in. Connecticut SB 435's employment AI provisions are active for covered employers. Additional state legislatures have introduced similar bills in the 2026 session cycle, and several attorneys general, including contested races in Texas where workplace AI rules have become campaign issues, have signaled enforcement interest. The realistic planning window for building compliant workflows is measured in months; the realistic window after an enforcement inquiry opens is measured in days.
A sensible schedule for a mid-size employer: complete the AI inventory within 30 days, finish workflow redesign and template drafting within 90 days, complete manager training within 120 days, and conduct the first internal audit of AI-assisted adverse actions within 180 days. Employers who wait for federal preemption should note that Congress has repeatedly discussed but not passed comprehensive AI employment legislation as of September 2026. Waiting is a bet on legislative gridlock, which may pay off, but a single state AG inquiry in the interim can cost more than the entire compliance program.
Costs and What Drives Them
Budget expectations vary by employer size. Small employers (under 100 employees) operating in one or two states may achieve baseline compliance with $10,000 to $40,000 in combined legal review, template drafting, and training costs. Mid-size multistate employers typically spend $50,000 to $250,000 in year one, covering vendor contract renegotiation, impact assessments (which outside counsel or consultants often price at $15,000 to $50,000 per tool), and compliance platform licensing. Large enterprises routinely exceed $500,000 when bias audits, litigation-hold-ready documentation systems, and dedicated compliance headcount are included.
Against this, weigh the downside. AI-related employment litigation, including class claims over algorithmic termination decisions, has produced settlements in the seven-figure range, and statutory penalties under state AI acts can accrue per violation, meaning per affected employee. The economic argument for compliance is not abstract; it is actuarial.
The Bottom Line
AI termination disclosure requirements in 2026 follow a consistent logic across California, Colorado, Connecticut, and other covered jurisdictions: tell workers when automated systems influenced consequential adverse decisions, explain the data and logic involved, guarantee meaningful human review, and preserve records proving you did all of it. The requirements are manageable but unforgiving of half-measures. A human signature on an algorithmic recommendation without genuine review satisfies no one, and silence satisfies fewer. Employers who inventory their tools, rebuild workflows around authentic human judgment, and automate the record-keeping, rather than the deciding, will be positioned to answer both regulators and employees with confidence.