An Algorithm Is the Common Question Dukes Said Was Missing
For fifteen years the standard defence to a large employment-discrimination class was that decisions were made locally, by many people, for many reasons. Automating the decision removes that defence. The exposure in an AI screening system is not only that it might be biased — it is that a single system makes a single question answerable for everyone it touched.
What Dukes actually decided
Wal-Mart Stores, Inc. v. Dukes is remembered as a case about class size. It was not. The Supreme Court held in 2011 that the plaintiffs had failed Rule 23(a)(2) commonality because the company's actual policy was to delegate pay and promotion decisions to local managers. There was no common mode of exercising discretion, so there was no single question whose answer would resolve the claims of the class. The opinion asked for "some glue" holding the alleged reasons for the decisions together, and found that a policy of allowing discretion is the least uniform practice imaginable.
Read that as a specification and an automated screening system satisfies it line by line. The model is one artefact with a version number. The threshold is a number in a configuration file. The training data is one corpus. The decision rule does not vary by region, manager or mood. Whether it disadvantages a protected group is a question with one answer, provable by one statistical analysis of one dataset — which is the textbook description of a common question.
The four Rule 23 elements, read against a model
Commonality — 23(a)(2)
EASIER FOR PLAINTIFFSOne system, one rule, uniformly applied. The common question writes itself, and it is answerable on the defendant's own logs rather than on thousands of depositions.
This is the element Dukes made hard and automation makes easy. It is also the element a defendant can no longer influence after deployment, because the uniformity is a fact about the product.
Typicality — 23(a)(3)
EASIER FOR PLAINTIFFSA named plaintiff scored by the same model on the same features is typical almost by definition. The usual objection — that the representative's circumstances were unusual — has little purchase when the circumstance in question is a score.
Where it can still bite: a plaintiff rejected at a later human stage for a documented, unrelated reason may not be typical of a class rejected by the model itself.
Predominance — 23(b)(3)
CONTESTEDThe live battleground. Defendants argue individual questions — would this applicant have been hired anyway, what are their damages — swamp the common ones. Plaintiffs answer with Rule 23(c)(4) and certify liability as an issue class.
Comcast v. Behrend requires a damages model that matches the liability theory, which is a real constraint. It constrains the damages class, not a liability-only issue class.
Class definition and scope
WHERE DEFENCE LIVESWhich model version, which configuration, which date range, which employer. A class spanning three model versions and two threshold changes is three classes pretending to be one, and saying so is the strongest available argument.
It only works if your change log shows the versions and dates. Without records, the plaintiff's single undifferentiated class stands unrebutted.
The vendor case changes the denominator
There is a second, larger consequence that has no analogue in the pre-automation era. When decisions were made by managers, the class could not be bigger than one employer's workforce. When decisions are made by a product sold to thousands of employers, the population screened by the same model is not one employer's applicants — it is the vendor's entire customer base.
The Mobley v. Workday litigation is where that theory is being tested. The court declined to treat the vendor as immune merely because it was not the employer, reasoning that an entity performing a function traditionally performed by the employer can be reached as its agent; and in 2025 it preliminarily certified a nationwide age-discrimination collective spanning applicants across many employers. A vendor that has always modelled its legal exposure as contractual indemnity to individual customers is exposed here to one aggregated claim instead.
The design choices that change the answer
None of this is destiny. Certification turns on facts about how the system was built and run, and those facts are decided before any complaint exists:
Per-customer configuration and retraining
A model tuned on each employer's own data, with thresholds each sets independently, is genuinely several decision systems. It is also harder to govern and harder to audit centrally — a real trade, not a free defence.
Human review that demonstrably diverges
Keep the reviewer's decision, the score they saw and the reason recorded. If the record shows reviewers departing from the ranking on stated grounds at a measurable rate, individualised decision-making is a fact you can evidence rather than a claim you assert.
A versioned change log with dates
Model version, feature set, threshold, effective dates. This is what narrows a class to a period, and it is the single cheapest artefact on this list.
Validation evidence contemporaneous with use
A job-analysis and validation study supports the business-necessity defence at the merits stage, where a certified class is defeated rather than avoided. Dated after the complaint, it is worth much less.
Adverse-impact testing you actually act on
The awkward one. Testing creates the document that quantifies impact, and a document showing a disparity you kept using is the plaintiff's best exhibit. The answer is not to stop testing — it is to attach a dated record of what changed after each test.
The uncomfortable symmetry
Every artefact that proves compliance also supplies common proof. Bias-audit summaries published under New York City's Local Law 144 are public documents quantifying impact ratios. Scoring logs required for records-retention are the dataset a plaintiff's statistician would otherwise have to fight for in discovery. This is genuinely a cost of the compliance regime and it should be stated rather than finessed. It is still not a reason to keep fewer records: the alternative posture — a uniform automated decision system with no validation evidence, no change log and no impact testing — loses on the merits with nothing to argue from, and in several jurisdictions the absence of the audit is itself the violation.
Can every candidate actually reach your application flow?
Screening exposure starts before the model. An application form with unlabelled fields, a keyboard trap, or an assessment step a screen-reader user cannot complete excludes a protected group at the door — uniformly, across every applicant, which is the same commonality problem in a different statute. Scan the page free.
Scan Your Application Page for Free →This is general background on how class and collective mechanics interact with automated decision systems, not legal advice, and certification practice varies by circuit.
Frequently Asked Questions
Why does an algorithm help a plaintiff satisfy commonality?
Wal-Mart Stores v. Dukes (2011) refused certification because pay and promotion decisions were made by thousands of local managers exercising discretion, so there was no common answer to the question of why any individual was treated as they were. A model inverts every part of that: one artefact, one training set, one scoring function and one threshold, applied the same way to everyone it touched. The question 'does this system disadvantage a protected group' has a single answer, which is exactly what Rule 23(a)(2) asks for.
Is that a settled holding or an inference?
It is an inference from Dukes' reasoning that courts have begun to accept in practice rather than a Supreme Court holding about algorithms. Dukes itself said the plaintiffs needed 'some glue' holding the decisions together and found none in a policy of discretion. The argument that a scoring model is that glue is the natural reading, and the Mobley litigation against Workday is the most visible instance of a court letting a claim proceed on a vendor-wide basis because the tool was common across employers.
What happened in Mobley v. Workday?
An applicant alleged that a vendor's screening product rejected him across many employers on the basis of race, age and disability. In 2024 the Northern District of California allowed the case to proceed against the vendor on an agency theory — the vendor performing a traditional employer function was not shielded merely by not being the employer. In 2025 the court granted preliminary certification of a nationwide age-discrimination collective under the ADEA. The collective mechanism there is the Fair Labor Standards Act opt-in procedure rather than Rule 23, but the certifying logic is the same: the screening tool is the common element.
Does human review in the loop defeat certification?
Only if it is real and documented. A reviewer who looks at a ranked list and hires from the top has not introduced individualised decision-making; the model decided who was visible. Human review breaks commonality when the record shows reviewers actually reaching different outcomes from the same scores on reasoned grounds, at a rate you can evidence. If your own logs show the top-scored candidate advanced in almost every case, the review is documentation of the model's decision, not an alternative to it.
Can a class be certified on liability alone?
Yes, and it is the likely shape. Rule 23(c)(4) allows an issue class on particular questions, so a court can certify 'did this system produce a disparate impact' for the class while leaving damages to individual proceedings. That structure removes the predominance objection that individual damages otherwise supply, which is why defeating certification by pointing at damages variation is weaker here than in a wage case.
Does the federal retreat from disparate impact reduce this risk?
It redistributes it. Federal enforcement priorities shifted away from disparate-impact theories in 2025, but enforcement priorities are not the whole of the law: Title VII's private right of action remains, and state statutes — the Illinois amendments, the California civil-rights regulations on automated decision systems, the Colorado and New York City regimes — supply independent claims and in some cases better remedies. Less agency activity with the same private exposure means the class mechanics matter more, not less.