Who Decides When a Machine Decides, and Who Watches?

Disparate Impact

Liability for a practice that looks neutral but falls more heavily on a protected group, proved by counting results rather than by showing intent. Under Griggs v. Duke Power Co. and the 1991 amendments to Title VII, the plaintiff names the specific practice and shows the disparity it causes. The employer must then show that the practice is job related and consistent with business necessity.

Disparate Treatment

Liability for treating a person differently because of a protected characteristic, which turns on intent. It reaches a tool built to use the characteristic or a stand-in for it, such as a graduation year for age, and an employer that keeps using a tool it knows discriminates. For age, it is the only claim the ADEA gives a job applicant in Florida’s federal courts, because the Eleventh Circuit reads the Act’s disparate-impact provision to protect employees, not applicants.

Four-Fifths Rule

A rule of thumb from the federal Uniform Guidelines on Employee Selection Procedures: a group selected at less than four-fifths the rate of the most-selected group generally counts as evidence of adverse impact. Divide the lower selection rate by the higher, and a ratio under 0.80 flags the practice for a closer look. It is an agency’s yardstick, not a statute, so it binds no court, and in a small sample a gap that size can appear by chance.

Vendor as Agent

The theory that a company performing a hiring function for an employer can be liable as the employer’s agent under Title VII, the ADEA, and the ADA, though it never employed the applicant. In Mobley v. Workday, a federal court in California accepted it at the pleading stage, reasoning that an employer that lets a screening tool reject applicants has delegated a function employers traditionally perform. Under it, a suit against one vendor can reach hiring at many employers at once.

Risk Score and Due Process

A risk score predicts, from data about groups of people, how likely a person is to commit another crime, and courts meet such scores at bail and sentencing. In State v. Loomis, the Wisconsin Supreme Court held that a sentencing court may consider one without violating due process, though not to decide whether a defendant is incarcerated or for how long. The report that carries the score must warn the judge of its limits, among them that the company has kept secret how the tool weighs its factors.

Facial Recognition as a Lead

A face search compares a probe photo with a database of known photos and returns a list of candidates, each with a similarity score that measures resemblance. Florida policies such as St. Petersburg’s treat a result as a lead that needs corroboration, never by itself as probable cause to arrest. What the defense must be told is unsettled, and New Jersey’s highest court now requires, case by case, basic disclosure of the tool and how police used it.

Third-Party Doctrine

The rule that information a person hands to a business, like bank records or dialed phone numbers, carries no reasonable expectation of privacy, so the government’s collecting it is not a search. Carpenter v. United States refused to extend the rule to a phone’s cell-site location records, and in 2026 Chatrie v. United States held that obtaining a person’s Google Location History is a search too. Courts have not settled where the line falls for other records of movement, such as license-plate camera logs.

Public Records Request

Florida’s chapter 119 opens the records of state and local agencies to inspection and copying by any person, whatever the reason, unless a statute makes a record exempt. A request need not be in writing, and the custodian must acknowledge it promptly, cite the statute behind anything withheld, and charge only the fees the statute allows. Exemptions cover plate-camera data that identifies a person and information revealing surveillance techniques, so a response may come back with parts blacked out.

Part 1: Read, Watch, and Choose

Start with the background reading. Then watch the video under the title from 27:20 to about 42:40, where two law professors and a university’s human resources officer argue over who answers when a hiring tool discriminates. Choose your path and read its sources in Resources. For the screener test, read Griggs at pages 429 to 436, the four-fifths rule in 29 C.F.R. § 1607.4(D), the Mobley order at pages 5 to 16, and the first two pages of Villarreal. For the records request, read the paragraphs of the Dillon complaint that Resources lists, the syllabus of Chatrie, St. Petersburg’s policy, and section 119.07, Florida Statutes. Do Part 2 or Part 3, not both.

Part 2: The Screener Test

Beacon’s screener. Beacon Pet Supply, the Daytona Beach retailer with about 60 employees, is hiring a warehouse operations manager for its Daytona Beach warehouse. It licenses a résumé screener that reads every application and ranks the applicants for interviews, and the vendor will not say how the ranking works. Beacon’s owner wants a test before the next round of hiring: a chatbot given the same job, the same instruction to rank, and résumés that differ in one detail at a time.

Run the Experiment Designer in a fresh chat. It builds four base résumés with equal qualifications, each for an applicant near 28. Then it makes three variants of each that change one thing: the name, the graduation year (to one that puts the applicant near 55), or a two-year gap in employment. Check each variant against its base before you go on, because a variant that changes two things spoils the comparison. Paste the screening prompt and the twelve variants into a new chat of one chatbot and record the five it picks. Run it again in another new chat with the résumés in reverse order, because the order of a list can sway a model’s choices.

Build a tally with a row for each of the twelve résumés and a column for each run, marking each time a résumé made the top five. Then compare groups, one change at a time. For the name, the four renamed résumés are one group and the other eight are the comparison; age and the gap work the same way. A group’s selection rate is its picks divided by its chances, and four résumés over two runs give eight chances. Divide the lower rate by the higher and mark every ratio under 0.80.

Write a memo of about 250 words to Beacon’s owner. It has four parts. The numbers give each group’s selection rate and each ratio, with the chances counted. The law names, for any ratio under 0.80, the statute and theory that would reach it for a job applicant in Florida, and whether the vendor could be sued along with Beacon. What is missing says what a plaintiff would still need that your test cannot supply. The change recommends one thing Beacon should do and says plainly what your test shows and what it cannot. If no ratio falls under 0.80, the memo says so and explains what that result does and does not clear. Score the memo one point for each part it gets right. Then run Compliance Counsel in a fresh chat with your tally and memo, revise once, and score the memo again.

Part 3: The Records Request

Choose the Volusia Sheriff’s Office or the Daytona Beach Police Department, whichever fewer classmates have written to so far. Write a request under chapter 119, Florida Statutes, for four kinds of records. Ask for the agency’s written policies on face searches, its written policies on license-plate cameras, and its contracts, agreements, or memoranda of understanding with the companies or agencies that supply either one. Add one more record you choose that would show how the tools are used, such as training materials or the reports the Pinellas County Sheriff’s Office sends agencies on their face searches. An agency owes you only records it already has, so ask for documents, not answers to questions. Describe each record so a clerk could find it without guessing. Ask for electronic copies, for an estimate before any charge, and for the statute behind anything withheld or blacked out. You need not say why you want the records.

Then write a note of about 250 words on what Chatrie and Dillon mean for what comes back. Score your request and note one point for each of four parts. The records part asks whether the request names each record so a clerk could find it and cites chapter 119. The terms part asks whether it requests electronic copies, an estimate, and the statute behind anything withheld. The Chatrie part asks whether the note says what the plate-camera policy should say about searching stored camera records, and why. The Dillon part asks whether the note says what the face-search policy should say about treating a result as a lead and passing on its limits. It also asks whether the note states plainly what the records can and cannot show. Run the Records Request Reviewer in a fresh chat with the request and the note, revise once, and score them again. Then send the request through the agency’s portal in Resources.

Part 4: Post Your Work

Create a single new discussion thread. For the screener test, post your tally with its rates and ratios, the revised memo, both scores, and a sentence on what Compliance Counsel led you to change. For the records request, post the request as you sent it, with the agency and the date and without your email address or phone number. Add the revised note, both scores, and a sentence on what the Records Request Reviewer led you to change. If the agency has answered by the time you post, add what came back. Before you post, check your work against its four parts.

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.

Experiment Designer

Changing one detail at a time is controlled variation, the method of the audit studies that test hiring for bias: if only the graduation year differs, a gap in outcomes points at age. Asking the model to say what two runs cannot show builds the limit into the design before a result tempts you to claim too much.

Compliance Counsel

Reading the memo from both sides is a form of red-teaming: the applicant’s lawyer finds where the case is thin, and the employer’s counsel finds where the advice reaches too far. Checking the arithmetic first, and admitting when it is unsure of the law, keeps the model’s critique on ground you can verify.

Records Request Reviewer

Casting the model as the clerk who has to fill the request is role assignment, and it shows you how your words read from the other side of the counter. Holding it to the statute text you pasted is source grounding, which keeps a confident guess about an exemption from passing as law.