Algorithms That Decide

Disparate Impact

Liability for a neutral practice that falls more harshly on a protected group without a business justification, no intent required. It is the natural theory against an algorithm, because the algorithm has no intent and its outcomes can be counted.

Disparate Treatment

Liability for treating a person differently because of a protected characteristic. It reaches an algorithm when the design uses the characteristic directly, or a proxy for it, or when the employer knows the tool discriminates and keeps using it.

Vendor as Agent

The theory the Workday court accepted at the pleading stage: a company that performs a hiring function for an employer, such as screening and rejecting applicants, can be liable as the employer’s agent under Title VII, the ADEA, and the ADA even though it never employed the applicant. It is the reason a single lawsuit against a software vendor can reach the hiring decisions of thousands of employers at once.

Algorithmic Collusion

The antitrust claim that competitors who feed nonpublic pricing data into a shared algorithm and follow its recommendations have agreed to fix prices under Section 1 of the Sherman Act, even without ever speaking to each other. The RealPage settlement defines what a pricing tool may and may not ingest.

Risk Assessment and Due Process

Algorithmic scores used at bail and sentencing to predict reoffending. In State v. Loomis the Wisconsin Supreme Court allowed a proprietary score at sentencing with warnings that it may not be determinative and that its methodology cannot be examined, a compromise that remains the leading case.

Facial Recognition and Disclosure

The use of biometric matching to identify suspects, its documented error rates, and the growing insistence by courts that the defense be told when and how it was used. A match is an investigative lead, and treating it as an identification is how wrongful arrests happen.

Bias Testing

A controlled comparison of outcomes across groups for otherwise identical inputs. Employers run it to find problems before plaintiffs do, and the Workday privilege ruling shows why they run it through counsel.

Part 1: Map the Doctrine

Read the Holland & Knight and Duane Morris summaries of Workday, the Justice Department’s RealPage release, and the Wex entry on disparate impact. Then use the Doctrine Mapper to build a one-page chart of which legal theory reaches which kind of automated decision, hiring, pricing, and criminal risk, and who the defendant is under each. Verify the statutes it cites against the Resources.

Part 2: Run the Résumé Experiment

Use the Experiment Designer to build a set of twelve résumés for a warehouse operations manager position: four base profiles with equivalent qualifications, each in three variants that change only the applicant’s name, the graduation year (implying an age near 28 or near 55), or the presence of a two-year employment gap. Write a screening prompt that asks the model to rank the top five candidates for interview and give a reason for each. Run the same prompt with the same twelve résumés in two different chatbots, twice each, and tally how often each variant lands in the top five. Apply the four-fifths rule from the Uniform Guidelines to your tallies and record the result honestly, whether or not it shows a disparity.

Part 3: Two Sides and a Fix

Use the Plaintiff’s Theory Builder to write the disparate impact case an applicant could bring on your results, naming the statute, the protected group, the practice, and the evidence, and then use Compliance Counsel to write the employer’s response: what an audit through counsel would look like, what notice Colorado’s law would require, and what changes to the screening process would reduce the disparity. Close with a paragraph on whether your experiment could be adapted to test a rent-setting algorithm or a pretrial risk score, and what would be different.

Part 4: Post Your Deliverables

Create a single new discussion thread containing your doctrine chart, your twelve résumés and screening prompt, your tally table with the four-fifths calculation, your plaintiff’s theory, your compliance fix, and the closing paragraph. State plainly whether your experiment found a disparity and how confident you are in a result from four runs.

Part 5: Share Chat Link

Include one AI chat link with a 1–2 sentence explanation of what the conversation shows and why you chose to share it.

Doctrine Mapper

Act as an employment and antitrust law professor. Build a one-page chart of the legal theories that reach automated decision-making in three settings: AI hiring screens, algorithmic pricing tools, and criminal risk assessments. For each setting, name the statute or doctrine, the theory (disparate impact, disparate treatment, agency liability of a vendor, Sherman Act Section 1, due process), who the defendant would be, what the plaintiff has to prove, and the leading case or enforcement action. Cite each statute by section and each case with enough detail to find it, and say plainly where the law is unsettled or where you are unsure of a citation.

A comparative chart with citations to check. Putting the three settings side by side shows that the same algorithmic problem draws different theories, and the citation rule turns the chart into a verification list rather than a lecture.

Experiment Designer

Act as an industrial-organizational psychologist who designs adverse impact studies. Help me build a controlled experiment to test whether an AI screening prompt treats applicants differently by race, sex, age, or employment gap. Create four base résumés for a warehouse operations manager with equivalent education, experience, and skills, then generate three variants of each that change only one thing: the applicant’s name (choose names commonly associated with different races and sexes in published audit studies), the graduation year (implying an age near 28 or near 55), or a two-year gap in employment. Keep everything else identical. Then write a screening prompt that asks a model to rank the top five for interview with reasons, and a tally sheet for recording outcomes by variant across multiple runs. Explain how to apply the four-fifths rule from 29 C.F.R. Part 1607 to the results and what a result from only four runs can and cannot show.

Controlled variation as method. Changing one attribute at a time is what lets you attribute a difference in outcomes to that attribute, and the instruction to explain the limits of four runs keeps you from overclaiming a result.

Plaintiff’s Theory Builder

Act as a plaintiff’s employment lawyer. I will paste the results of a résumé experiment on an AI screening prompt. Write the disparate impact theory an applicant could bring on these facts: the statute (Title VII, the ADEA, or the ADA), the protected group, the specific employment practice challenged, the statistical evidence and how the four-fifths rule applies to it, who the defendants are including any vendor as the employer’s agent, and what the employer’s likely defenses would be. Where the evidence is too thin to support a claim, say so and explain what more a plaintiff would need. Cite the statutes and the leading case by name so I can verify them.

Building a case from your own data. Forcing the theory onto the numbers you actually produced shows how much evidence a claim needs, and the instruction to admit thinness is what keeps the exercise honest when the experiment finds little.

Compliance Counsel

Act as outside counsel advising an employer that uses an AI résumé screening tool. I will paste the results of a bias experiment on a screening prompt and the plaintiff’s theory built on it. Advise the employer on three things: how to structure a bias audit through counsel so the results are privileged, as the court permitted in Mobley v. Workday; what notice and human review Colorado’s SB 26-189 will require when it takes effect and how to prepare; and what changes to the screening prompt, the data, or the process would reduce the disparity. Be concrete and cite the ruling and the statute so I can check them. Do not tell the employer the tool is safe if the results say otherwise.

The other side of the same facts. Advising the employer on the same experiment shows that compliance is a design problem rather than a denial, and the closing instruction keeps counsel from telling the client what it wants to hear.