The Problem Has a Name (Sort Of)
At least 10 Maryland jurisdictions use AI tools to set bail and recommend sentences — with no meaningful oversight. Job applicants are filtered out before a human ever glances at their résumé.
The trouble is that these tools are largely opaque. They produce scores, not explanations. They encode historical patterns — patterns that often reflect systemic inequity — and then present those patterns as objective risk assessments.
Spoiler: they’re not objective. They’re just confident.
How Algorithmic Bias Actually Works in Practice
Dartigue’s composite character, Shanna, is a useful lens. Ten years of work experience. Forty-eight job applications. Zero human reviews. Six months of rejection, depleted savings, missed rent — and then a $73 shoplifting charge that landed her in front of an AI risk-scoring tool.
The tool flagged her as “high-risk.” Why? Unstable housing. Unemployment. The very conditions her algorithmic job rejections had created.
This is the feedback loop that makes AI bias so insidious in criminal justice: the system punishes people for outcomes the system helped produce.
The Legal Toolkit: What “Creative Lawyering” Actually Looks Like
Dartigue didn’t just diagnose the problem — she handed lawyers a playbook. Here’s what she recommended:
Challenge the Risk Assessment Directly
Demand to know what data fed the score. What variables were weighted? Was the model validated on a population that looks like your client? These are not rhetorical questions — they’re discovery requests waiting to happen.
Use the Daubert Standard Strategically
The Daubert standard governs the admissibility of expert testimony and scientific evidence in federal courts (and many state courts). If an AI risk score is being used as evidence, it should meet the same bar as any other expert opinion: testable methodology, peer review, known error rates.
Most of these tools can’t clear that bar. Make them try.
File Motions in Limine
Before trial, move to exclude AI-generated risk scores as prejudicial evidence. Frame it as a reliability argument. Frame it as a due process argument. Frame it as both.
Invoke the Maryland Fair Employment Practices Act
For clients facing algorithmic discrimination in hiring, the MFEPA offers a statutory hook. Disparate impact claims don’t require proof of intent — just proof of effect. That’s a meaningful distinction when the discriminating party is a black-box model.
Demand Transparency as a Right
Clients have a right to understand the basis of their detention. That right doesn’t evaporate because the decision-maker is software. Push for disclosure of the model’s logic, training data, and validation studies.
The Disparate Impact Frame Is the Right Frame
Dartigue explicitly asked lawyers to think about these cases through a disparate-impact lens — and that framing matters strategically.
You don’t need to prove the algorithm was designed to discriminate. You need to show it does discriminate, in effect, against a protected class. That’s a lower bar, and it’s the right bar when the harm is structural rather than intentional.
AI tools trained on biased historical data will reproduce that bias at scale. That’s not a bug in the legal argument — it’s the legal argument.
What This Means for AI Accountability More Broadly
Dartigue’s presentation is a signal that the legal profession is starting to catch up with the technology. Slowly, but meaningfully.
The tools being used in courtrooms and HR departments are not neutral infrastructure. They’re decision-making systems with real consequences — and they’re currently operating with far less scrutiny than a human expert witness would face.
The Daubert challenge. The motion in limine. The disparate-impact claim. These aren’t exotic legal theories. They’re standard tools being applied to a non-standard problem.
The Takeaway
“We have to push the envelope. We have to file the motions. We have to ask the questions.”
Dartigue’s call to action is essentially a product brief for the next generation of civil rights litigation. The AI tools are already deployed. The harm is already happening. The legal frameworks — imperfect, but usable — already exist.
The gap isn’t in the law. It’s in the willingness to use it.
If you’re a lawyer, that’s your cue. If you’re building or deploying AI tools in high-stakes contexts, that’s your warning.
Comments (0) No comments yet
Want to join this discussion? Login or Register.
No comments yet. Be the first to share your thoughts!