The Problem: Claude Is Useful, Not Purpose-Built — and That Gap Gets Defenders Sanctioned

Ask Claude to summarize a suppression hearing transcript you paste in, and it will do a genuinely good job. Ask it to research the law on consent searches, and you're rolling dice: general-purpose chatbots like Claude hallucinate on legal research queries between 58% and 82% of the time, according to Stanford HAI research cited in a recent industry report. Specialized legal AI trained on verified legal corpora gets it wrong roughly one time in six. That gap is the whole story.

The consequences aren't theoretical. There have been more than 120 documented AI hallucination cases in U.S. legal filings since mid-2023, with at least 58 in 2025 alone. Federal judges have sanctioned attorneys who submitted briefs citing cases that never existed. For a defender, one fabricated citation in a motion to suppress can cost you credibility with a judge you appear before every week.

The honest verdict: Claude drafts and reasons well, but it is useful, not purpose-built. Legal capability gets layered onto a general-purpose model through plugins and connectors rather than built around a legal workflow from the ground up. Even Anthropic's own repository disclaimer describes its output as "a draft for attorney review," not legal advice.

Where that gap actually bites in criminal defense:

• Unsourced legal research. Without a connector to a current database, Claude answers from training data — a citation-accuracy risk on exactly the questions your motions turn on.
• Keyword-search failure modes. A Quinn Emanuel attorney documented Claude confidently stating a contract clause didn't exist because the keyword wasn't found — it had never actually read the document.
• No grounding in your evidence. Claude has no native way to cite the exact second of a bodycam video or page of a police report, so any answer about your discovery dump is unverifiable by construction.

That last point matters most for defenders. The failure mode isn't Claude being dumb — it's Claude being fluent without a source. A workspace like Evidence Copilot exists to close that gap: every AI answer carries a clickable citation into the exact video moment or PDF page it came from, and outputs are marked pending human review. Claude can be pointed at that grounded record through an MCP connection; on its own, it can't create one.

So the framing for the rest of this article is simple. Claude is a strong reasoning and drafting engine. Whether it's safe for your practice depends entirely on what it's connected to — and whether every output is verifiable before it touches a filing.

The Privilege Trap: What United States v. Heppner Means for Anyone Using Claude on a Criminal Case

If you're a criminal-defense attorney and your client types "what should I tell my lawyer" into Claude on their personal account, you now have a federal judge's answer to whether those messages are protected: they're not.

In United States v. Heppner, Judge Jed S. Rakoff of the Southern District of New York issued a first-impression ruling on February 17, 2026, holding that a criminal defendant's self-directed Claude exchanges were neither attorney-client privileged nor work product. The defendant was convicted on all counts on May 7, 2026, and those Claude exchanges were introduced as evidence at trial. As detailed analysis of the ruling explains, this is the first decision of its kind, and it should change how every defense team approaches AI.

Rakoff's reasoning breaks into three parts:

• AI is not an attorney, so communications with a chatbot are not privileged on their face.
• The confidentiality prong failed under the terms of service then in effect.
• Heppner's use was self-directed, not counsel-directed — no lawyer involved the AI as an agent of the representation.

The court did leave a path open. Under the Kovel doctrine, counsel-directed AI use on a platform with contractual confidentiality might preserve privilege. That distinction matters: a defender who directs a client or investigator to use a specific tool, on an enterprise tier with contractual data protections, is in a categorically different position than a defendant browsing on a consumer account.

That consumer tier is where the trap lives. Anthropic's August 28, 2025 policy shift allows opt-in model training with five-year data retention on consumer plans — opting in fails the Heppner confidentiality prong on its face. Only Claude for Work, Government, Education, and API tiers remain no-train with enterprise data handling. This tracks the broader ethics picture: under Rule 1.6, inputting client information into consumer platforms may breach confidentiality, and as the BCG Attorney Search report notes, many platforms train on submitted data by default unless enterprise agreements are in place.

Three practical takeaways for the defense bar:

1. Treat any self-directed, consumer-tier AI exchange on a criminal case as discoverable. Outside S.D.N.Y., Heppner is persuasive rather than binding — but persuasive is enough to lose with.
2. Route AI use through counsel. The Kovel path requires counsel-directed use on a platform with contractual confidentiality, not ad hoc client experimentation.
3. Put the confidentiality question in writing before anyone uploads anything: which tier, whose account, what retention terms.

This is also why workflow matters as much as the model itself. Tools built for defense work — where every AI output is marked pending human review and grounded in the actual evidence with citations — are designed around the confidentiality and verification duties Heppner and ABA Formal Opinion 512 impose. The chat is only as safe as the platform it runs on.

The Ethics Baseline: ABA Opinion 512 Duties You Can't Outsource to a Chatbot

Before you ask Claude anything about a case, the baseline is simple: the ethics rules don't care which tool you used. ABA Formal Opinion 512 and state bar guidance treat generative AI the same whether it's Claude, ChatGPT, or a purpose-built platform — and Arizona's guidance puts it bluntly: using AI "does not alter or lessen the ethical obligations of attorneys."

The duties that apply to every Claude session:

• Competence (Rule 1.1). You need a working understanding of the tool's failure modes — including hallucination — before you rely on it on a client matter.
• Confidentiality (Rule 1.6). Pasting client names or case specifics into a consumer platform may breach Rule 1.6, since many platforms train on submitted data by default unless you're on an enterprise tier.
• Candor and verification (Rule 3.3). Every citation must be checked against the actual source before anything is filed. Fabricated case citations have produced sanctions and disciplinary action.
• Supervision (Rules 5.1/5.3). Partners and supervisors own the AI outputs of everyone they supervise — law clerks included.

The verification duty deserves emphasis for defenders. General-purpose chatbots like Claude hallucinate on legal research queries 58–82% of the time, according to Stanford HAI research cited by BCG Attorney Search — versus roughly one in six for specialized legal AI. More than 120 documented hallucination cases have appeared in U.S. filings since mid-2023, with at least 58 in 2025 alone. Federal judges have sanctioned attorneys who submitted AI-generated briefs without reading the cases.

There's also a subtler failure mode worth knowing: Claude can run keyword searches instead of actually reading a document. A Quinn Emanuel lawyer reported that Claude confidently declared a termination-at-will clause didn't exist — because the keyword wasn't found, not because the model had read the contract. Imagine the same failure on a police report where the exculpatory detail is phrased in unexpected words. That's a Brady problem, not just a quality problem.

The practical rule is one Anthropic itself states: treat every Claude output as a draft pending attorney review, not legal advice. Nothing goes into a motion, a cross-exam outline, or a client communication without you confirming it against the source. That's the same posture we build into Evidence Copilot — every AI answer is marked AI-generated and pending human review, and the product never makes legal conclusions.

State guidance is still uneven — some bars have formal opinions, others only preliminary recommendations — so check your own jurisdiction. But the core duties are settled: Claude doesn't lower the bar. You do the reading, you do the verifying, you sign the filing.

The Fix: Ground Claude in Your Actual Evidence via MCP

The problems in the previous section all trace back to one cause: Claude answering from training data instead of your case. The documented failure mode is instructive — a Quinn Emanuel attorney watched Claude confidently declare a clause didn't exist because a keyword search missed it; the model had never actually read the contract. On unsourced legal research, that risk isn't theoretical: general-purpose chatbots like Claude hallucinate on legal queries 58–82% of the time, per Stanford HAI research cited in an industry report on generative AI in law.

The fix is grounding. When Claude queries your actual evidence instead of its training data, every answer becomes verifiable — and this is exactly what MCP integration enables. Anthropic's connector ecosystem is the core of its legal push, with 20+ MCP connectors in Claude for Legal, but most connectors target transactional and research platforms like Westlaw, iManage, and DocuSign. None of them touch the actual work of criminal defense: the bodycam, the interrogation footage, the jail calls, the police report.

Evidence Copilot ships an MCP server that connects Claude Desktop or Claude Code directly to your case files. It's built for defense work specifically:

• list_cases and list_evidence — see what's in the file without opening the workspace
• search_evidence — full-text search across diarized transcripts, OCR'd police reports, and scanned exhibits
• read_window — pull the exact transcript segment around a moment, not a keyword snippet
• get_citation — retrieve a citation to the precise video second or PDF page

The server is read-only, org-scoped, and fully audited, with revocable tokens — Claude can answer questions from your evidence, but it can't modify anything, and every query leaves a record. Ask "did the officer's report match what he said on bodycam?" and the answer comes back with a citation you can click through to the exact second of footage.

Two things matter for practitioners weighing this. First, the grounding changes the verification posture required under ABA Formal Opinion 512 and state bar guidance: instead of hunting down whether a citation exists, you check whether the cited moment says what Claude claims — a task that takes seconds, not an afternoon. Second, no defense-side competitor currently offers an MCP integration, so this workflow isn't available anywhere else in the category.

Every answer still carries the same guardrail as the rest of the workspace: outputs are marked AI-generated and pending human review, and the tool never makes legal conclusions. Claude does what it's genuinely good at — reasoning across documents and drafting — while the evidence workspace does what a general-purpose model can't: guarantee the answer traces back to a real file at a real timestamp you can verify before it goes anywhere near a motion or a cross-exam outline.

Putting It to Work: A Counsel-Directed Claude Workflow for Defense Teams

So Claude can draft — but the research is blunt about what happens when it drafts from memory instead of from your file. General-purpose chatbots hallucinate on legal research queries 58–82% of the time, per Stanford HAI findings cited by BCG, and one documented failure mode has Claude running keyword searches instead of actually reading the document in front of it (GC AI's analysis). The fix isn't a better prompt. It's a grounded workspace underneath Claude that hands it the evidence first.

Step 1: Review the discovery inside a grounded workspace. Upload the dump — bodycam, interrogation footage, jail calls, 911 audio, police reports, scanned exhibits — to Evidence Copilot. Every file gets a timestamped, diarized transcript or OCR'd text. On ingest, the system compares each new file against earlier evidence and surfaces candidate inconsistencies — report versus bodycam, statement versus call audio — as unreviewed signals before anyone asks. You're not asking Claude to remember your case; you're working from a citable record where every answer points to the exact second of video or page of the PDF.

Step 2: Connect Claude through MCP, read-only. From Claude Desktop or Claude Code, the MCP server exposes five tools:

• list_cases and list_evidence — see what's in the workspace without touching originals
• search_evidence — full-text search across transcripts and OCR'd documents
• read_window — pull a specific transcript segment with its timestamps intact
• get_citation — retrieve the exact source moment behind any answer

The connection is read-only, org-scoped, token-revocable, and fully audited — no defense-side competitor offers an MCP integration.

Step 3: Draft with every claim tied to a source moment. Now Claude earns its keep on the tasks defenders actually bill: a suppression motion outline where both sides of the consent or Miranda dispute are quoted and cited, cross-exam question drafts where every suggested question traces to a quoted source, and one-click case briefs covering evidence inventory, key moments, and follow-ups. This matters ethically, not just practically. ABA Formal Opinion 512 and state bar guidance require verification of AI output, and Anthropic's own disclaimer calls its output a draft for attorney review, not legal advice. Grounded citations make that verification a two-minute check instead of a re-read of forty hours of footage.

Pricing, published and flat. Solo runs $39/user/month with 15 media-processing hours included; public-defender offices pay $750/attorney/year with 100 hours. Overage is metered and hard-capped at 3x included hours, so a runaway discovery dump can't overbill you. Every plan includes the full workspace — no feature tiers, no sales wall.

If you want to see the workflow against one of your own cases, book a 20-minute walkthrough at evidencecopilot.com/book. Or start on the Solo plan at evidencecopilot.com/signup and point it at your next bodycam dump.