Unfair discrimination

Original Linkedin Post

An employer can be accountable for a rejection decision it never saw, made by a tool it did not build.

MIT subdomain 1.1, unfair discrimination and misrepresentation, covers unequal treatment of individuals or groups by AI, often on the basis of race, gender or other sensitive characteristics.

 In Mobley v. Workday, a collective action in the Northern District of California, the applicant alleges he was rejected from more than 100 roles by employers using the vendor’s AI screening tools. In 2024 the court allowed the claim to proceed on the theory that a vendor screening candidates on an employer’s behalf can be its agent. A nationwide age-discrimination collective was conditionally certified in May 2025, distribution of the class action notice to potential plaintiffs was authorised in February 2026, and the core discrimination claims are still live. Workday denies the allegations and says its tools do not make hiring decisions. Nothing is proven yet.

For lawyers, risk and governance professionals, a few points stand out.

Discrimination does not need a protected characteristic in the data. A model trained on past hiring patterns can learn proxies: graduation year, employment gaps, keywords.

At provider level, the vendor configures the ranking logic and holds most of the evidence about how it behaves. At deployer level, the feature may be switched on by default, with no one looking at rejection rates by age band. And the applicant can be rejected in minutes, without ever being told that AI was involved, which is what makes this risk so hard to detect from the inside.

Practical steps:

• Map where automated screening and ranking already sit in your hiring chain, including features enabled by default.
• Test outcomes across protected groups before go-live and at intervals afterwards, not only the design intent of the tool.
• Put bias-testing obligations, audit rights and evidence access into vendor contracts.
• Keep human review meaningful and documented. For high-risk uses under the EU AI Act, this sits with the deployer, not the provider.
• Decide how bias-testing work is commissioned and consider whether privilege applies.

Buying the tool does not transfer the accountability.