The Defender's Duty: Why a Law Professor Says Using AI Is Now an Ethical Obligation

A law professor just argued that refusing to use AI is the ethical risk

Most writing about artificial intelligence in law starts from the assumption that the cautious lawyer is the one who stays away. A new article makes the opposite case. In Parity in Practice: The Defender’s Duty to Ethically Use AI, Mason R. Clark, Assistant Professor of Law at St. Mary’s University School of Law, argues that criminal defense lawyers now have an affirmative professional obligation to engage with these tools — and that the greater danger is a defense bar that stays on the sidelines while the other side does not.

The argument deserves attention beyond criminal defense, because the mechanism it describes applies to any firm competing against a better-resourced opponent.

The gap is not enthusiasm. It is resources.

Clark’s central observation is uncomfortable and specific: adoption is uneven between prosecutors and defenders. Prosecutors’ offices are increasingly experimenting with enterprise-grade AI systems. Meanwhile, many public defender offices and solo practitioners lack not only access to those systems, but the resources and expertise to challenge software outputs as unreliable in court.

That second half is the part firms tend to miss. The disadvantage is not only that one side drafts faster. It is that one side can interrogate the technology and the other cannot. Clark notes that courts have largely permitted police and prosecutors to use “black box” systems whose internal decision-making is difficult to explain. A defender who does not understand how probabilistic genotyping or facial recognition actually works — including documented error rates and disparities — cannot meaningfully contest evidence produced by it.

Framed that way, AI competence stops being a productivity question and becomes an advocacy question.

The duty runs in two directions

The article treats AI fluency as serving two distinct professional duties.

The first is the familiar one: competence and diligence in your own work. Overwhelming caseloads are the normal condition of defense practice, and Clark treats these tools as a genuine equalizer for research, organization, and drafting under that pressure.

The second is adversarial. Defenders who understand these systems can hold the prosecution’s technology to a standard, and can articulate to courts and legislatures where practical AI use collides with ethical obligations. Clark goes further and suggests the defense bar has an opportunity to set the standards — pointing to work like the National Association of Criminal Defense Lawyers’ Task Force on Artificial Intelligence. Documented internal standards become the benchmark against which the other side’s practices can be measured.

The bottleneck is verification, not access

Here is where our own experience matches the paper, and where we would sharpen it.

Access to a capable model is no longer the constraint. Anyone can subscribe to a frontier model this afternoon. The constraint is that a lawyer cannot sign their name to work they cannot check. Clark is blunt about the failure mode: defenders risk presenting fabricated content if they do not verify outputs, and he warns that the ubiquity of these tools encourages “skill flattening” — competence quietly eroding because the machine handled it.

So the useful question is not which model is smartest. It is how long does it take to verify what it produced. A summary you must re-read the transcript to trust has saved you nothing. A draft citation you must look up by hand has moved the work, not reduced it.

That is why the tools worth adopting are the ones built to be audited:

• Citations that point at a location, not a vibe. When a transcript search returns an exact page and line, verification takes seconds instead of an afternoon. This is the design principle behind Lawnova PDF, which indexes court transcripts with precise page and line references.
• Drafts that show their sources. An appellate brief section is only useful if you can see which record documents produced it. COAPP attributes generated sections to the underlying record and enforces court formatting and word limits rather than leaving them to a final panicked check.
• Confidentiality you control. Clark treats client confidentiality as a first-order constraint, and it is the reason a locally hosted, airgapped option matters for privileged material — not everything belongs in someone else’s API log.

None of that removes the lawyer’s judgment. It shortens the distance between an AI output and a human who has confirmed it, which is the only place the time saving legitimately comes from.

The disclosure rules are already here

One detail in the article is worth every practitioner’s attention: nearly two dozen jurisdictions now require attorneys to provide disclosure and/or verification statements for documents prepared with the assistance of generative AI. Clark takes his own medicine and opens the article with exactly such a statement, certifying that every citation was manually verified to confirm it exists and stands for what he says it does.

That is a reasonable template for a firm policy. Not a ban, and not a free-for-all: a written record of where these tools were used, and a human certification that the output was checked.

What to do this quarter

If you take one thing from Clark’s article, make it this: decide, in writing, where your firm allows these tools, what verification each use requires, and who signs off. Firms that do this get the efficiency and keep the defensibility. Firms that leave it informal end up with neither, and eventually with an explanation to make to a judge.

The parity argument is the one that should move managing partners. Your opponent’s capability is not waiting for your firm’s comfort level.

Source: Mason R. Clark, Parity in Practice: The Defender’s Duty to Ethically Use AI, St. Mary’s University School of Law. Available on SSRN, abstract 6337279: ssrn.com/abstract=6337279. This post is our summary and commentary; the article itself is the authority, and it is worth reading in full.