Why Defenders Are Asking This Question Now

The question isn't curiosity — it's caseload. Utah public defenders are assigned up to 525 misdemeanor cases per person, and some Louisiana defenders carry as many as 1,000 clients, according to an academic study of public defenders. Nobody watches 19 hours of bodycam on 21 videos the old way under that load.

The discovery volume is the real driver. Bodycam footage, interrogation video, jail calls, 911 audio, scanned police reports — every case arrives as a media dump, and the same study found that evidence investigation — transcribing, summarizing, and synthesizing large volumes of digital records — is the task defenders see as most promising for AI. NACDL's president put the advantage plainly: catching the contradiction buried in 10 hours of bodycam footage.

There's also a parity argument gaining traction. A report cited by Courthouse News argues that because prosecutors are already adopting AI, defenders have an ethical duty of competence to do the same — responsibly. "AI is already inside the courtroom," as NACDL's president put it. The defenders who learn to use it well are the ones who out-lawyer the other side.

But the pressure to adopt collides with three hard problems:

• Hallucination risk is documented, not hypothetical. Stanford researchers found Lexis+ AI and Ask Practical Law AI produced incorrect information over 17% of the time, and Westlaw with integrated CoCounsel hallucinated more than 34% of the time.
• Courts have logged 246 fabricated-citation incidents as of September 2025, involving 135 attorneys — and ABA Formal Opinion 512 requires lawyers to verify every generative AI output.
• Confidentiality makes consumer ChatGPT dangerous for case work. OpenAI discloses that records may be used for training, and third-party data can fall outside privilege and become discoverable. NACDL guidance flags "significant confidentiality concerns" with commercial chatbots.

That's why "just use ChatGPT" is not a safe answer. Pasting a police report into a consumer chatbot risks privilege, produces unverifiable output, and gives you no way to trace an answer back to the exact second of video or page of the PDF it came from. For defenders, the useful version of this technology isn't a chatbot — it's a grounded evidence-review workspace where every AI answer carries a citation into the original source and stays marked pending human review, which is the approach Evidence Copilot is built around. The question has shifted from whether to adopt AI to which tools can survive ethical scrutiny.

Two Categories of Legal AI — and the Hallucination Problem in Both

Yes, there's a ChatGPT for lawyers — several, actually — but they fall into two very different camps, and both have a hallucination problem you need to understand before you trust either one.

The first category is general-purpose chatbots: ChatGPT, Claude, and Gemini. These are the tools most lawyers reach for first, and they carry real confidentiality risks. OpenAI discloses that ChatGPT records may be used for training, and third-party data can fall outside attorney–client privilege and become discoverable via subpoena — a risk Sam Altman himself has acknowledged. NACDL guidance flags significant confidentiality concerns with commercial chatbots and recommends enterprise or legal-specific tools instead.

The second category is legal-specific AI: CoCounsel (Thomson Reuters), Harvey, Lexis+ AI, Westlaw AI-Assisted Research, and JusticeText. These are built for legal workflows — research, drafting, evidence review — and lawyers prefer them: in a survey of 2,800+ legal professionals, 29% expressed greater trust in legal-specific tool output and 43% prioritized integration with trusted software.

Here's the uncomfortable part: legal-specific tools still hallucinate. Stanford RegLab/HAI's independent testing found:

• Lexis+ AI and Ask Practical Law AI produced incorrect information over 17% of the time
• Westlaw with integrated CoCounsel hallucinated more than 34% of the time
• Courts had documented 246 fabricated-citation incidents as of September 2025, involving 135 attorneys and 103 pro se litigants

Those numbers matter because vendor marketing tells a different story. Thomson Reuters reports CoCounsel cuts discovery review by up to 63%, speeds timelines 79%, and answers client questions 83% faster — but those figures are vendor-reported with no published methodology, and they sit awkwardly next to Stanford's independent findings. Treat efficiency claims with the skepticism you'd apply to any prosecutor's summary.

The lesson isn't to avoid AI — it's to demand grounded, citation-linked answers you can verify. That's the design principle behind Evidence Copilot: every AI answer in the evidence agent cites the exact video second or PDF page it came from, and every output is marked AI-generated and pending human review. ABA Formal Opinion 512 already requires lawyers to verify all generative AI output, so tools that make verification easy aren't a luxury — they're how you use AI without ending up in the next fabricated-citation count.

Evidence Review Is the Use Case That Actually Matters for Defense

Ask any public defender which AI task actually matters, and the answer is rarely "write me a brief." An academic study of public defenders found that evidence investigation — transcribing, summarizing, and synthesizing large volumes of digital records — is the task they see as most promising for AI. Courtroom representation and defense strategy ranked as least compatible, and that distinction should shape how you evaluate any tool.

The reason is simple: evidence review is verifiable work. When the question is "what did the officer say at minute 43 of the bodycam," there is a correct answer sitting in the file. NACDL's president framed the advantage as catching the contradiction buried in 10 hours of bodycam footage — not generating arguments from thin air.

The proof is already in the record. JusticeText, now used by 70+ public defense agencies, reported one case with 157 court videos where review dropped from roughly 140 hours to 12. A Nevada investigator reviewed 21 bodycam videos — about 19 hours — in 2 hours. Those are transcription-and-synthesis outcomes, not strategy outcomes.

But the same research shows why free-form chat fails here. Stanford RegLab/HAI found Lexis+ AI and Ask Practical Law AI produced incorrect information over 17% of the time, and Westlaw with integrated CoCounsel hallucinated more than 34% of the time. Courts had documented 246 fabricated-citation incidents as of September 2025. When the stakes are a wrongful conviction, an answer you cannot trace back to the source is worse than no answer.

That is the case for a grounded-citation standard — and it is the standard we built Evidence Copilot around:

• Every AI answer links to the exact second of video or page of the PDF it came from, so verification is one click, not one re-watch.
• Every output is marked AI-generated and pending human review, mapping directly to ABA Formal Opinion 512's requirement that lawyers verify generative AI work.
• Originals stay immutable; redactions and derived files are new, audited copies.
• Inconsistencies — report vs. bodycam, statement vs. jail call — surface as unreviewed signals before anyone thinks to ask.

The workflow that matters for defense looks like this: upload the discovery dump, get timestamped diarized transcripts and OCR'd document text, then ask questions whose answers carry citations into the source itself. Ask "when did the Miranda warning happen" and get the timeline with the exact video moment attached, not a confident paragraph you have to fact-check line by line.

That is the honest answer to "is there a ChatGPT for lawyers" for defense work: not a chatbot that talks like a lawyer, but an evidence agent that shows its sources on every claim.

How to Evaluate and Adopt a Defense-Specific AI Tool

Before you sign up for anything, run the vendor through the same scrutiny you'd give a prosecution expert. The stakes justify it: independent research found that legal AI tools billed as reliable still produce incorrect information 17–34% of the time, and courts had documented 246 fabricated-citation incidents by late 2025. A tool that can't show its work is a malpractice risk, not an assistant.

Here's the vetting checklist we'd use, grounded in what the research and NACDL's risk framework actually demand.

1. Does every answer cite the exact source? Ask the vendor to demo a question against real evidence and show where the answer came from. If the output is free-form chat with no link back to the specific video second or report page, you're doing the verification work anyway — just slower. Stanford RegLab/HAI research found Westlaw with integrated CoCounsel hallucinated more than 34% of the time when answers aren't grounded in verifiable sources. Tools like Evidence Copilot are built around this problem: every AI answer carries a clickable citation into the exact moment of bodycam footage or page of a scanned police report, and every output is marked AI-generated and pending human review.

2. Is output gated for human review? ABA Formal Opinion 512 requires lawyers to verify all generative AI outputs. The tool should make that easy, not fight it — look for systems where AI suggestions (redaction regions, timeline events, cross-exam questions) file as review cards rather than auto-applied changes.

3. Are originals immutable? You're handling evidence. Redactions and edits should produce new derived files while the original stays untouched, with an audit trail. If a tool modifies source media, walk away.

4. Is pricing published? A sales wall is a red flag for budget-strapped defense work. Compare: CoCounsel runs roughly $400/user/month as a research and memo tool; JusticeText runs about $1,200/attorney/year for transcription-focused review. Evidence Copilot publishes its pricing outright — Solo at $39/user/month, Firm at $99/user/month, and a PD Office tier at $750/attorney/year, with every plan including the full feature set rather than paywaling capabilities.

5. Does it handle the actual evidence? Discovery isn't text documents. Your tool needs to work on the files you actually get:

• Bodycam and interrogation video with speaker-diarized, timestamped transcripts
• Jail calls and 911 audio, searchable and citable
• Scanned police reports and photo exhibits, OCR'd into searchable text
• Redaction for discovery turnover — frame-accurate video, word-level audio

Finally, apply NACDL's risk framework: AI supports evidence review — transcription, summarization, surfacing inconsistencies — while case strategy and client counseling stay with the attorney. As the NACDL guidance frames it, AI should support, not supplant, human judgment in high-risk uses. The right tool makes your review faster and more verifiable. The judgment stays yours.

If a vendor clears all five checks, take the 20-minute walkthrough with a real case's evidence types in hand — that demo tells you more than any marketing page.