Practice Areas
AI for Criminal Defence Lawyers: Disclosure Review, Evidence and Privilege
AI criminal defence tools review unused material, bodycam and transcripts and build timelines. The guardrails: privilege after Heppner, verified citations and deepfakes.
On this page
Criminal defence lawyers face a government with more investigators, more data and more time, and the National Association of Criminal Defense Lawyers concluded in August 2026 that defenders therefore have a duty to use AI: "The danger is that only one side does." The tools earn that duty in disclosure review, bodycam and transcript analysis and timeline building. The guardrails are privilege, which a client's own chatbot use does not carry, verified citations, and a working scepticism about machine-generated evidence.
"The danger is that only one side does"
The NACDL white paper, "Parity in Practice: The Defender's Duty to Ethically Use AI" (4 August 2026), starts from the prosecution side: offices already use AI to manage evidence, speed discovery and review old convictions for resentencing. Its president, Andrew Birrell: "If the defense sits this out, we do not stay neutral. We fall behind. And when the defense falls behind, the client pays for it." The paper names the working examples. The Kentucky Department of Public Advocacy uses JusticeText to review bodycam footage; the Colorado State Public Defender uses Reduct to clip, caption and build exhibits from bodycam and interrogation video; Maryland's Office of the Public Defender runs case files on eDefender; the California Innocence Project has used AI to prioritise wrongful-conviction reviews; a Wisconsin defender used JusticeText to surface police misconduct. Reported gains include up to 85% less manual data entry, and a Thomson Reuters vendor blog claims 63% less discovery-review time.
The paper's method is its best argument. It was drafted with AI assistance and then every citation was verified by hand, "exactly the method we are asking defenders to use". Birrell's formulation of the rule is the one to pin above the desk: "treat AI output like the work of a sharp new associate who has never set foot in a courtroom. You check the citations. You read the cases. The tool drafts. The lawyer decides." It recommends formal AI policies within a year, staff training, vendor review and a model policy template.
Unused material and the volume problem
A podcast on the American Arbitration Association's AAAi channel describes a public defender trying to review more than a million pages of discovery for a client facing multiple life sentences. That is the defence problem in one sentence: the volume of disclosure and unused material now exceeds what a small team can read, and what is not read cannot be used.
Generative review is the answer to the reading, with known characteristics. In Redgrave LLP's case study (June–July 2026) on a 45,000-document set, Relativity aiR reached 87–88% recall against 64% for active-learning TAR, at the cost of lower precision: the tool over-includes, so the second-pass pile is larger. For a defence review that is the right trade. Recall is what finds the exculpatory document; the second pass is lawyer time that would otherwise be spent reading everything. The mechanics, and the court's refusal in Schulte v. LinkedIn (N.D. Cal., 30 June 2026) to treat generative review as anything other than ordinary TAR, are in AI in e-discovery.
| Material | What the tool does | What the lawyer does |
|---|---|---|
| Documentary disclosure and unused material | Codes for relevance to defence themes; flags exculpatory and impeachment content; builds an index | Writes the coding protocol; reads everything flagged; samples the rest and records the elusion rate |
| Bodycam and CCTV | Transcribes, indexes by speaker and time, clips segments | Watches every clip relied on; checks the transcript against the audio |
| Interview and interrogation recordings | Transcribes; finds inconsistencies with statements | Reads the full transcript of any passage cited; checks caution and warnings |
| Phone extractions and messages | Timeline of contacts; keyword and pattern search | Interprets; instructs a forensic examiner where provenance matters |
| Expert reports | Summarises; cross-references methods and data | Reads the report; briefs the defence expert |
Bodycam, interviews and the summary that invented witnesses
Video and audio are where AI saves the most defence hours and where the worst recent failure occurred. On 9 September 2026 the New Mexico Supreme Court fined Santa Fe lawyer Stephen Aarons $5,000, held him in contempt and referred him to the disciplinary board after his brief in a murder appeal "contained false testimony from wholly fabricated witnesses", including "fictional statements that the shooter was wearing dark pants and a white shirt". He had fed a computer-generated transcript and case materials to ChatGPT at the start of the appeal expecting "a bulletproof summary". At the 21 August hearing Justice C. Shannon Bacon asked: "Counsel, do you watch the news? Do you listen to the radio? ... the problem with lawyers relying on AI hallucinations is an above-the-fold story every single day."
The failure was not the transcription; it was trusting a summary without a citation trail. A transcript workflow that holds up:
- Transcribe with a tool built for evidence (JusticeText, Reduct, or a review platform), keeping the original recording as the exhibit.
- Ask the model for an index and a chronology in which every entry carries a timestamp and, for written statements, a page.
- Ask for inconsistencies between a witness's recorded account and their written statement, quoted verbatim with both references.
- Open every timestamp before the passage is used in a motion, a cross-examination plan or a brief.
- Never ask a general chatbot for a "summary of the trial"; ask for quotations with references and build the summary yourself.
You are assisting a criminal defence lawyer. Work only from the uploaded
transcript and the uploaded witness statement. Do not add facts.
1. List every factual assertion in the witness statement about the
events of [DATE], numbered, quoting the statement with its paragraph.
2. For each assertion, find the passage in the transcript where the same
witness addresses the point. Quote it verbatim with page and line
(or timestamp). If the transcript does not address it, write
"NOT ADDRESSED".
3. Flag every difference in time, place, sequence, description of
persons, clothing, vehicles or weapons. Quote both versions.
4. Do not characterise credibility or suggest which version is true.
The transcript method in general is in AI deposition and transcript summaries, and for a closed case file Gemini Notebook is a reasonable container, provided it runs in a Workspace or enterprise account: see Gemini Notebook for lawyers.
Privilege after Heppner
Defendants research their own cases. Those on bail have months to do it, and the tool they reach for is a consumer chatbot. In United States v. Heppner (S.D.N.Y., bench ruling 10 February 2026, opinion 17 February 2026) a defendant had used consumer Claude to draft research memos and strategy about his own defence; the FBI seized 31 such documents. Judge Rakoff held that none was privileged or work product. Privilege never attached, so nothing was "waived": "Claude is not an attorney", and Anthropic's consumer terms permit logging and disclosure. Rakoff signalled that the outcome might differ where counsel directs the AI use, along the lines of the Kovel doctrine for accountants. A reported ruling in Warner v. Gilbarco (E.D. Mich., February 2026) went the other way for lawyer-directed work product, calling AI programs "tools, not persons"; that ruling rests on a single vendor report and should be cited with care.
Three consequences for a defence practice. First, the client conversation at the first meeting now includes a warning: do not put your case into a chatbot; if you already have, tell your lawyer what and where. Second, any AI research the defence team wants protected is done by the team, under instruction, in a business-tier tool that does not train on inputs (the confidentiality checklist compares tiers). Third, chatbot logs are discoverable material: a federal court ordered 20 million ChatGPT logs produced in the OpenAI copyright litigation (affirmed 5 January 2026), so a co-defendant's or a witness's AI history is something to ask for as well as something to fear. The full analysis is in privilege and AI and the short version in does using ChatGPT waive privilege.
Machine-generated evidence and deepfakes
Proposed Federal Rule of Evidence 707 would have required machine-generated evidence offered without an expert to meet the reliability standards of Rule 702. The Advisory Committee voted 8–1 to publish it (the Department of Justice dissenting), comments closed on 16 February 2026, and, according to Proskauer's account of the committee papers, the Standing Committee decided on 3–4 June 2026 not to recommend action and sent the rule back for revision and further study alongside the separate deepfakes question. The August 2026 to February 2027 comment cycle covers Evidence Rules 104 and 902 only. As of September 2026, then, challenges to AI-generated evidence run through the existing rules: authentication, Rule 702 and the expert.
The NACDL paper tells defenders to be alert to altered or fabricated material entering a case "from any source", to scrutinise digital evidence closely and to use forensic experts where authenticity is in doubt. The irony of the leading deepfake case is instructive: in Kohls v. Ellison (D. Minn., January 2025) a Stanford professor's expert declaration defending Minnesota's deepfake law was struck because it cited AI-invented studies. England's judicial AI guidance (31 October 2025) tells judges to be aware of "potential challenges posed by deepfake technology". A defence checklist for any AI-touched exhibit from the prosecution:
- What was the original source, and does it still exist in unprocessed form?
- What tool processed it (enhancement, transcription, translation, facial or voice matching), which version, with what settings?
- Who operated it, and can they be cross-examined?
- What is the tool's documented error rate, and on what population was it measured?
- Has the defence expert reviewed the processing chain, not only the output?
Sentencing research and the citation rule
Sentencing submissions and appellate grounds are drafted under the worst time pressure in the practice, and the disciplinary record reflects it. In the Sixth Circuit's 2026 review of Farris, a lawyer with a 40-year clean record was removed from the Criminal Justice Act appointed-counsel panel and referred for discipline over fabricated citations, according to Norton Rose Fulbright's June 2026 survey of sanctions. Austria's Oberster Gerichtshof (14 Os 95/25i, 7 October 2025) dismissed a Graz defence lawyer's Nichtigkeitsbeschwerde in a drug-trafficking case without reaching the merits, under § 285d Abs 1 StPO, because the AI-drafted appeal contained invented OGH decisions and misquotations, and pointed to the lawyer's personal responsibility under § 9 Abs 1 RAO. The client's appeal was lost on the form of the document.
The rule for research is the same as for litigators generally, described in AI for litigators: general models frame issues and draft search strings; authorities come from a grounded tool; every case is opened and read by the signing lawyer; the citation-checking protocol is completed on the file. Anthropic's open-source litigation plugin tags any citation not checked against a connected database as [verify], which is the right default for every tool.
Guardrails for a defence practice
- A written policy within the year, as the NACDL recommends, naming the permitted tools and tiers and excluding consumer chatbots from client material.
- Review protocols for disclosure and for video, with the second-pass and elusion-sample rules written down.
- The transcript rule: every quotation carries a timestamp or page, and no general-model "summary" is ever the basis of a filing.
- The client warning about chatbot privilege, delivered at the first meeting and recorded.
- The evidence checklist above applied to every AI-touched prosecution exhibit.
- Training that covers audio and video handling as well as text; Module 3 of the AI Lab for Lawyers course works through multimodal file handling of recordings and scanned exhibits in the browser.
Birrell's closing line in the NACDL paper is the right summary of where this leaves a defender: "What AI takes off your plate is the grind. The document review. The first-draft timeline. Do that faster and you have more hours for the human work, not fewer."
Frequently asked questions
Can criminal defence lawyers use AI to review disclosure?
Yes, and defender organisations already do. Generative review tools code large productions for relevance and privilege at higher recall than older technology-assisted review, and video tools transcribe and index bodycam and interview footage. The controls are a written review protocol, a human second pass on everything the tool flags as significant, and a rule that nothing goes into a filing without a page or timestamp reference.
Is a client's conversation with ChatGPT or Claude privileged?
Not on the current US authority. In United States v. Heppner (S.D.N.Y., February 2026) Judge Rakoff held that memos a defendant created with consumer Claude were neither privileged nor work product, because Claude is not a lawyer and the terms permit logging and disclosure, though privilege might attach where counsel directs the use. Warn clients, especially those on bail with time to research.
How do I challenge AI-generated or deepfake evidence?
Through the existing rules, since proposed Federal Rule of Evidence 707 was sent back for further study in June 2026. Demand the provenance and processing history of the material, put the proponent to proof on authentication and reliability under Rule 702 or its state equivalent, and retain a forensic expert where the stakes justify it. Treat AI-enhanced or AI-summarised material from the prosecution as machine output, not as a witness.
Can I use AI for sentencing research?
For framing, yes; for authorities, only through a tool grounded in a verified database and then only after opening each case. Criminal courts have removed lawyers from appointed-counsel panels and referred them to the bar for fabricated citations, and an Austrian appeal drafted with AI was dismissed unread. The citation rule does not soften because the client cannot pay for a second lawyer to check.
Put this into practice
AI Lab for Lawyers: a live, hands-on course for practising lawyers
4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.
Related reading
Use Cases
AI in E-Discovery: From TAR to Generative Review
Generative AI review beats TAR on recall (87–88% vs 64% in Redgrave's 2026 study) but loses precision. Tools, privilege, defensibility and a sampling plan.
11 min read
Use Cases
AI Deposition and Transcript Summaries: A Litigator's Workflow
AI deposition summaries work when every quotation is checked against the transcript, the tool runs on a business tier and output is anchored to page and line.
9 min read
Regulation & Ethics
Attorney–Client Privilege and AI: Heppner, Warner, the NYT Logs and Public Chatbots
Consumer chatbot memos are not privileged. US v. Heppner, Warner v. Gilbarco and the 20-million-log OpenAI order, plus rules that keep privilege intact.
10 min read
Questions
Does Using ChatGPT Waive Attorney–Client Privilege?
Using ChatGPT rarely waives privilege in the technical sense; the courts' answer is worse. In Heppner, consumer-chatbot memos were never privileged at all.
6 min read
Playbooks
How to Verify AI Legal Research: A Citation-Checking Protocol
Verify AI legal research with a five-check protocol: open every authority, match the quote, confirm holding and status, log it. Template, time budgets, tools.
11 min read
Practice Areas
AI for Litigators: Research, Drafting, Discovery and the Sanctions Trap
AI for litigators pays off in chronologies, discovery and deposition prep; it costs $110,204 when a brief cites invented cases. Where to gain and the one rule.
8 min read