Why There's No Single "Best AI" — and Why Firms Are Stuck

Ask ten vendors which AI is "best for law firms" and you'll get ten confident answers. The data tells a messier story. Personal generative AI use at work actually rose from 27% to 31% between 2023 and 2024, while firm-level adoption fell from 24% to 21%, according to the ABA's 2025 Legal Industry Report. Lawyers are using AI anyway — their firms just aren't leading.

That gap has a governance problem underneath it. In one survey of more than 1,300 legal professionals, 43% said their firm has no formal AI policy and no plans to create one. Only 9% have a written, enforced policy. So individual attorneys are improvising with consumer tools while the firm provides no verification workflow, no confidentiality vetting, and no training — 54% reported receiving no responsible AI training, with none planned.

And the tools being marketed as enterprise-grade aren't earning blind trust. Independent research found that Lexis AI and Westlaw AI produce incorrect information 17–34% of the time, despite "hallucination-free" marketing claims. Courts have documented 246 incidents of fabricated citations as of September 2025, involving 135 attorneys. The $5,000 sanction against two attorneys who filed a ChatGPT-drafted brief citing six fictitious cases is now a standard cautionary tale.

Here's the frame that actually helps: "best" is a property of the task, not the vendor. A Princeton study of public defenders found exactly this split — defenders ranked AI as most useful for evidence investigation (transcribing, summarizing, and synthesizing large volumes of digital records) and least compatible with courtroom representation and defense strategy. That maps cleanly onto how the work breaks down:

• General-purpose LLMs (ChatGPT, Claude, Gemini) for drafting, brainstorming, and summarizing text you can verify.
• Specialized evidence-review tools for discovery dumps — bodycam footage, jail calls, 911 audio, scanned police reports — where transcription quality and citation accuracy matter.
• Research platforms for case law, used with the understanding that outputs require independent verification before anything goes near a filing.
• Nothing AI-generated touching strategy, witness judgment calls, or courtroom work — per ABA Formal Opinion 512, lawyers can't delegate professional judgment to AI and must verify all outputs.

For criminal-defense work specifically, the evidence-review category is where the leverage sits. When a case arrives as 40 hours of bodycam and a stack of scanned reports, the question isn't "which AI is best" — it's which tool answers questions about that evidence with citations you can check against the source. That's the standard we built Evidence Copilot around: every AI answer cites the exact video moment or PDF page it came from, and every output is marked as pending human review.

The vendors won't tell you this, because "it depends on the task" doesn't close deals. But it's the honest answer — and the one the ethics rules already assume.

The Ethics Floor: Four Checks Every AI Tool Must Pass

Before you fall in love with any tool's demo, put it through the ethics floor. ABA Formal Opinion 512 (July 2024) made clear that the existing Model Rules — competence, confidentiality, communication, and fees — already apply to generative AI, and a 50-state survey shows states converging on the same recurring themes: verify everything, protect client data, keep judgment with the lawyer, and bill honestly. Only a minority of states (Arizona, California, Florida, Illinois) have issued formal guidance so far, but the floor is the same everywhere.

The stakes are not hypothetical. Courts have documented 246 incidents of fabricated citations as of September 28, 2025, involving 135 attorneys and 103 pro se litigants — including the June 2023 case where two New York attorneys were fined $5,000 for filing a ChatGPT-drafted brief citing six fictitious cases. Independent research also found that Lexis AI and Westlaw AI produce incorrect information 17–34% of the time, despite "hallucination-free" marketing claims. That's the cautionary baseline for every tool you evaluate.

Here's the four-check framework, drawn from Opinion 512 and the state guidance:

• Verify every output. The lawyer, not the tool, is responsible for accuracy. Any AI answer you can't check against the underlying source — the actual bodycam minute, the actual page of the police report — fails this test. Citation-grounded output is the minimum bar, not a nice-to-have.
• Protect confidentiality. 41% of surveyed lawyers cite data privacy as their top AI concern, and for good reason: uploading a client's discovery dump to a consumer chatbot can waive privilege. Confirm what the vendor does with your data before anything sensitive touches it.
• Keep judgment with the lawyer. Florida Bar Opinion 24-1 puts it plainly: a lawyer must not delegate to generative AI any act that could constitute the practice of law. AI can transcribe, summarize, and flag inconsistencies; it cannot decide your suppression theory or cross-exam strategy.
• Handle fees correctly. On hourly billing, you can charge for time spent inputting information and reviewing AI output — but not for time the AI saved you, and not for learning the tool. Arizona, California, and Florida all take this position.

This framework matters because most firms haven't built it themselves: 43% of legal professionals report their firm has no formal AI policy and no plans to create one, and only 9% have a written, enforced policy. If your firm is in that majority, these four checks are your de facto policy until you write a real one.

For evidence review specifically — the use case defenders rank as AI's highest value — the verification check is where most tools break. That's the design principle behind Evidence Copilot: every AI answer cites the exact video second or PDF page it came from, and every output is marked AI-generated and pending human review before it goes anywhere near a filing. A tool that can show its work passes the ethics floor; one that can't doesn't belong in your case file.

Match the Tool to the Task: Drafting, Research, and Evidence Review

The Princeton public-defender study makes one thing clear: defenders don't want a single AI that does everything. They ranked evidence investigation — transcribing, summarizing, and synthesizing large volumes of digital records — as the highest-value AI use case, while rating courtroom representation and defense strategy as least compatible with AI. The practical answer is to match the tool to the task.

Drafting and exploration is where general-purpose LLMs earn their keep. ChatGPT, Claude, and Gemini can produce the bare bones of a motion outline, summarize a document, or help you reframe an argument — but as Cayce Lynch of Tyson & Mendes puts it, AI output "is absolutely no substitute for our professional judgment" and requires substantial revision before it goes out the door. Florida Bar Opinion 24-1 goes further: a lawyer must not delegate to generative AI any act that could constitute the practice of law.

Legal research is a different animal. Purpose-built research tools generate memos and case citations, but quality is uneven — independent research found that Lexis AI and Westlaw AI produce incorrect information 17–34% of the time despite "hallucination-free" marketing claims. Courts have documented 246 fabricated-citation incidents as of September 2025. Verify every case cite before it touches a filing.

Then there's the task defenders consistently rank highest: evidence review. A discovery dump — bodycam video, jail calls, 911 audio, scanned police reports — is exactly where AI helps most, because the work is mechanical transcription and synthesis, not judgment. Caseload reality makes this urgent: Utah public defenders have been assigned up to 525 misdemeanor cases per person, and some Louisiana defenders carry caseloads as high as 1,000 clients. Nobody has hours to scrub raw footage by hand.

When evaluating tools for this task, the Princeton study's practitioner safeguards and ABA Formal Opinion 512 point to a common minimum bar:

• Every AI output is verified by a human before it's relied on
• Answers cite back to the exact source — the video moment or PDF page — so verification is fast, not a re-review
• Confidentiality protections cover client data end to end
• The tool never makes legal conclusions or replaces attorney judgment

That's the standard we built Evidence Copilot around: upload the discovery dump, get timestamped diarized transcripts and OCR'd document text, and ask questions of an AI evidence agent that answers with clickable citations into the exact second of video or page of the PDF — with every output marked AI-generated and pending human review, and originals kept immutable. Evidence review is where AI pays for itself first; drafting and research come second, and strategy stays with you.

What Grounded Evidence Review Looks Like in Practice

The discovery dump lands on Monday: 40 hours of bodycam, three jail calls, a scanned police report, and a preliminary hearing on Thursday. Here is what a grounded evidence-review workflow actually looks like, step by step, using Evidence Copilot as the example.

First, upload everything — video, audio, photos, PDFs up to 2 GB — by web upload or by forwarding the dump to the case's inbound email address. The Princeton study of public defenders found that defenders rank evidence investigation — transcribing, summarizing, and synthesizing large volumes of digital records — as the AI use case with the most value, so this is where the work should happen (a 2025 IRB-approved study).

Next, the system returns diarized transcripts with word-level timestamps and OCR'd text for scanned reports and photo evidence, so a receipt or a screenshot becomes searchable, citable text. You rename speakers once and it propagates across every segment.

Then you ask questions. "What did the officer say about consent to search?" Every answer comes back with citations to the exact video second or PDF page — click and the moment plays. This matters because independent research found Lexis AI and Westlaw AI produce incorrect information 17–34% of the time despite "hallucination-free" marketing, and courts have identified 246 incidents of fabricated citations (per the same research). Citation-to-the-moment is the minimum bar for evidence work.

The workflow then moves to the tasks that win motions:

• Inconsistency signals: every new file is compared against earlier evidence, surfacing report-vs-bodycam discrepancies before anyone asks.
• Cross-exam drafts: suggested questions only, each tied to a quoted source, saved as attorney-review notes.
• Suppression prep: a consent/search-scope/stop-duration/Miranda sweep that drafts a motion outline with both sides quoted and cited.
• Redaction for discovery turnover: frame-accurate video, audio, and PDF redaction — originals never touched, every action audited.

Every output is marked AI-generated and pending human review, consistent with ABA Formal Opinion 512's requirement that lawyers verify all AI output and never delegate professional judgment (ABA ethics guidance).

Compared with the alternatives — JusticeText at roughly $1,200/attorney/year with no published pricing or accuracy data on its site (per its homepage), or a free Whisper/NotebookLM stack that breaks at case scale — Evidence Copilot publishes everything: Solo $39/user/mo, Firm $99/user/mo, and a PD Office tier at $750/attorney/year. Every plan includes the full workspace. No sales wall, no feature tiers.

If that workflow fits your practice, book a 20-minute walkthrough at evidencecopilot.com/book, or start on the Solo plan at evidencecopilot.com/signup.

Adopting AI Without Getting Burned: A 30-Day Rollout

Most lawyers using AI at work are doing it without a net. A survey of legal professionals found that only 9% of firms have a written, enforced AI policy — while 43% have none and no plans to write one. For a solo defender or small firm, that gap is where sanctions risk lives.

The fix is smaller than you think. You do not need a 40-page governance framework; you need a one-page policy and one disciplined month.

Week 1: Write a one-page policy. Cover three things: which tools are approved, what client material may never go into them, and who verifies AI output before it touches a filing. This mirrors the recurring themes in state bar guidance — verification before submission, confidentiality safeguards, and written firm policies. ABA Formal Opinion 512 already applies competence and confidentiality rules to AI, so your policy is just operationalizing duties you already owe.

Week 2: Vet confidentiality. Read the vendor's data-handling terms before uploading a single bodycam or jail call. Remember that 41% of surveyed lawyers cited data privacy concerns — a legitimate worry, but one you can resolve contractually the same way you vetted your case-management platform.

Week 3: Run one verification workflow on one case type. Pick a single motion to suppress or preliminary-hearing prep. Upload the police report, bodycam, and 911 audio to a grounded tool like Evidence Copilot, then check every AI answer against the cited source moment — the exact second of video or page of the PDF. The hallucination data justifies this discipline: independent research found major legal research AI produces incorrect information 17–34% of the time.

Week 4: Measure. Time your old workflow against the new one on a comparable case. Only 18% of professional services organizations collect any AI ROI metrics, so this alone puts you ahead — and tells you whether the tool earns its keep.

Your one-pager should answer:

• Approved tools and permitted use cases (e.g., evidence review yes, strategy no)
• Confidentiality rules: what client data may be uploaded, and to which platforms
• Verification standard: no AI output reaches a filing without source-checking
• Billing treatment, per ABA Opinion 512 and your state's rules
• Who owns the policy and when it gets reviewed

Thirty days, one page, one case type, one measurement. If you want to see the verification workflow in action before committing, book a 20-minute walkthrough at evidencecopilot.com/book — or just start on the Solo plan at $39/user/month and run it on a real case. No urgency, no sales wall; the published pricing is on the site either way.