AI Ethics Watch — 2026-09-25

Top Stories

OMB M-26-04 Makes "Unbiased AI" a Federal Contract Requirement

OMB M-26-04 operationalizes Executive Order 14319's Unbiased AI Principles by embedding them as mandatory clauses in federal LLM solicitations. Every federal procurement of large language models must now carry these anti-bias contract terms, with a key compliance deadline of 11 March 2026 having been laid out in the guidance. This is a major shift: AI ethics requirements for government vendors are no longer aspirational policy but enforceable contract law, and vendors selling LLMs to the U.S. government must now demonstrate bias mitigation as a condition of business.

Italy and the EU AI Act Enforcement Machine Come Online

Under the EU AI Act, from 2 August 2026 the AI Office and member state authorities are formally responsible for implementing, supervising and enforcing the regulation, with the AI Office holding direct enforcement powers over general-purpose AI models. Additionally, each member state must establish at least one AI regulatory sandbox at the national level per Article 57. Fresh analysis from the Center for Democracy and Technology's September 2026 Europe Bulletin highlights a busy Brussels autumn: the State of the Union address, significant developments in the Digital Omnibus process, and crucial implementation milestones under both the AI Act and the Digital Services Act.

AI Hiring Tools Face Mounting Discrimination Litigation

AI hiring tools are drawing class-action lawsuits and a patchwork of city and state regulations targeting discrimination, disability exclusion, and lack of candidate notice, with employment discrimination suits raising the core question of who is liable — the employer, the software vendor, or both — when algorithms screen out applicants by age, race, gender, or disability. One vendor faces claims it scraped data on over one billion workers to assign a 0-to-5 score without required disclosures. A new analysis published this week examines how employers adopting AI throughout the employee lifecycle must navigate evolving local, state, federal, and supranational requirements governing bias, privacy, transparency, and human oversight.

Regulation & Policy Tracker

• United States (OMB/White House): OMB M-26-04 converts EO 14319's Unbiased AI Principles into federal contract terms for LLM procurement, with an 11 March 2026 compliance deadline requiring every federal LLM solicitation to carry unbiased-AI clauses
• European Union: As of 2 August 2026, the AI Office and member state authorities hold enforcement responsibility under the AI Act, including enforcement powers over GPAI models, and each member state must stand up at least one AI regulatory sandbox under Article 57
• European Commission (Digital Omnibus): CDT's September 2026 bulletin flags significant developments in the Digital Omnibus process and implementation milestones under the AI Act and Digital Services Act following the State of the Union address
• U.S. (White House, March 2026 follow-through): A national policy framework issued in March 2026 recommended Congress legislate broad preemption of state AI laws under a light-touch standard, though preemption is not settled law and states largely retain enforcement authority today

Bias & Accountability

• AI hiring systems (multiple vendors/employers): Lawsuits allege AI hiring tools screened out applicants on the basis of age, race, gender, or disability; one vendor is accused of scraping data on over one billion workers to assign a 0-to-5 "employability" score without required candidate disclosures — testing whether vendors or employers bear legal liability
• Workday (employment discrimination suit): A federal employment discrimination lawsuit against Workday puts AI-powered hiring software on trial and could determine who is liable when algorithmic screening produces discriminatory outcomes
• Stanford Law Review (scholarly accountability debate): An essay by Hoang Pham, Hannah Cha, and Rashon Poole argues "AI bias" disputes are best understood as normative disagreements about representation — whether AI should mirror the world as it is or promote a more inclusive future — a framing that will shape how antidiscrimination law applies to AI

Analysis: What This Means

The pattern this week is convergence: bias governance is moving from principles to mechanisms. OMB M-26-04 turns unstated "unbiased AI" ideals into enforceable federal contract clauses while AI hiring discrimination suits are testing precisely who bears liability when algorithms discriminate. Companies building or selling AI products should note the emerging dual exposure — vendor-side suits like the Workday case targeting technology providers, and employer-side compliance duties around notice, human oversight, and audit requirements. Meanwhile, the EU AI Act's enforcement machinery is now activated across member states, meaning organizations selling into Europe need concrete sandbox-ready compliance operations rather than policy documents.

What to Watch Next

• EU AI Act regulatory sandboxes deadline: Each EU member state must establish at least one national-level AI regulatory sandbox by 2 August 2026 under Article 57; compliance status is now being tracked post-deadline
• Brussels Digital Omnibus process: Significant developments are underway following the State of the Union address, with pending changes to AI rules that would simplify and potentially water down some requirements
• AI hiring discrimination class actions: Continued litigation in the growing class-action wave over automated resume screening and scoring tools — including claims of a vendor scraping data on over one billion workers — will likely produce further rulings clarifying liability allocation between employers and AI vendors