DescrybeLM: The Free Legal AI That Outscored ChatGPT on the Bar Exam

Published: March 14, 2026 | By: The Crossing Report | 9 min read

---

Summary

On March 12, 2026, a legal AI startup called Descrybe launched DescrybeLM — and made a claim that's worth taking seriously: in internal benchmarking, it answered all 200 multistate bar exam questions correctly. ChatGPT, Claude, and Gemini each missed double digits.

More relevant for small law firm owners than the benchmark: it's free.

This article explains what DescrybeLM actually does, how it's different from using a general-purpose AI tool for legal work, how to use it for a specific research task, and where it fits in the legal AI landscape — alongside CoCounsel, August, and the general-purpose tools most lawyers are already experimenting with.

---

The Problem DescrybeLM Is Trying to Solve

If you've tried using ChatGPT for legal research, you already know the problem. The output sounds authoritative. It cites cases confidently. And sometimes — often enough to matter — those citations are fabricated, the rule is from the wrong jurisdiction, or the analysis misses the specific statutory element that changes the answer.

This isn't a user-error problem. It's an architectural one. General-purpose AI tools are trained to produce statistically plausible text, not legally accurate analysis. They're predicting what a legal answer should look like — not reasoning from actual primary law.

The result is a tool that's useful for drafting, summarizing, and brainstorming, but unreliable for the work that actually requires you to be right: research that informs a client opinion, analysis that goes into a filing, or any output you'll put your name on.

For large law firms, the solution has been purpose-built legal AI platforms backed by verified legal databases — Westlaw's CoCounsel being the most prominent. The problem for solo and small firm lawyers: those tools cost $225/user/month. They're built for the economics of a 50-lawyer firm, not a 3-attorney practice.

That gap is what DescrybeLM is targeting.

---

What DescrybeLM Actually Does

DescrybeLM is a legal reasoning and drafting workspace built on a curated corpus of more than 100 million structured primary law records. It's purpose-built to produce what Descrybe calls "verification-friendly" legal outputs — answers structured around rule statements, key facts, and step-by-step analysis that you can check against actual sources.

The bar exam claim: Descrybe ran DescrybeLM through all 200 multistate bar exam questions. It answered them all correctly. In the same test, ChatGPT, Claude, and Gemini each missed double digits.

Benchmark claims from AI companies should always be read carefully — the multistate bar exam tests whether a model can apply legal rules to structured fact patterns, which is a specific and somewhat narrow competency. It doesn't test judgment, client-specific context, or the ability to navigate procedural complexity. But it does test whether the model applies legal rules correctly, which is the foundational requirement for any legal AI tool.

What the benchmark tells you: DescrybeLM was built with legal accuracy as the primary optimization target. General-purpose models were not.

The specific use cases DescrybeLM is designed for:

• Legal research drafts: Run an issue through DescrybeLM to get a structured first analysis. Review, verify the key authorities, and refine. Faster than starting from a blank search.
• Rule checking: Pose a specific legal question in a jurisdiction. DescrybeLM's structured output tells you the rule, the exceptions, and the analysis — in a format you can verify rather than just accept.
• Structured argument drafts: For standard legal arguments in transactional or litigation work, DescrybeLM can produce a first draft that follows sound legal analysis structure (rule → fact application → conclusion) rather than the narrative-summary output of general AI tools.

---

How DescrybeLM Fits the Legal AI Landscape in 2026

The Issue #7 edition of The Crossing Report introduced a framework that's worth repeating here: in 2026, there are two fundamentally different categories of legal AI.

Authoritative legal AI (CoCounsel, backed by Westlaw and Practical Law) produces outputs designed to be cited in filings and defended in court. Its research is sourced from verified primary materials with traceable citation chains. Price: $225/user/month.

Operational legal AI (Anthropic's Claude Cowork, Harvey's document tools, August) automates high-volume, lower-judgment tasks: NDA triage, document flagging, first-draft review. Price: $30–free/user/month.

DescrybeLM sits in a third category that hasn't had a good representative until now: reasoning-structured legal AI for small firms. It's not a citation-verified research database like CoCounsel, and it's not a general-purpose document automation tool like August. It's a purpose-built legal reasoning workspace where the architecture prioritizes accuracy over fluency — and it's free.

For a 1-5 attorney firm that has been relying on general-purpose ChatGPT or Claude for legal research because CoCounsel is out of the budget: DescrybeLM is the first purpose-built legal reasoning AI at zero cost. That's a meaningful change.

---

How to Use DescrybeLM for a Specific Research Task

Here's a practical workflow for the first time you use it:

Step 1: Identify a specific legal question from an active matter. Not a broad topic — a specific issue. "What are the requirements for a valid mechanic's lien in Texas?" or "Does a non-compete clause survive an asset purchase in Illinois?" Specific questions get specific structured answers.

Step 2: Run the question through DescrybeLM and review the output structure. The output should include a rule statement, the applicable statutory or case authority, and a structured analysis. Read the rule statement first — does it match your understanding? If not, that's a flag to investigate, not a reason to dismiss the output.

Step 3: Verify the key authority. DescrybeLM's accuracy benchmark is strong, but no AI tool is infallible. Before anything goes into a filing or a client opinion, verify the primary authority the model cites. This is a 5-minute check — you're not redoing the research, you're confirming the foundation is solid.

Step 4: Use the structured analysis as your first draft. DescrybeLM's analysis format — rule, facts, application, conclusion — maps directly to how legal analysis is structured. Edit for your client's specific facts, add jurisdictional nuance, and you have a first draft that took 20 minutes instead of two hours.

The human review step is not optional. DescrybeLM reduces the research and structuring time. Your professional judgment is still the final authority on anything that leaves your office.

---

Where DescrybeLM Still Needs Lawyer Review

DescrybeLM is not a replacement for authoritative legal research when the stakes are high. Be explicit with yourself about the distinction:

Use DescrybeLM for: Research drafts that a lawyer will verify, structured analysis for standard issue types, argument frameworks for typical transactional or litigation patterns, and research in jurisdictions where you have solid baseline knowledge.

Do not rely on DescrybeLM alone for: Anything going directly into a court filing without independent verification, novel legal questions where the rule is contested, multi-jurisdictional analysis with significant variation, or matters where a citation error would create malpractice exposure.

The professional judgment standard doesn't change because the tool is better. It changes the economics of applying that judgment: DescrybeLM handles more of the first-draft work so your judgment can focus on the verification and application layer.

---

The Access Question Is Settled — Start Testing

The practical reason most small law firms haven't adopted purpose-built legal AI isn't skepticism about the technology. It's the access model: enterprise pricing, sales calls, seat minimums, and implementation requirements that assume the budget and IT resources of a 100-person firm.

DescrybeLM eliminates the access barrier for legal reasoning AI. Free tier. Direct sign-up at descrybe.ai. No sales call.

If you've been waiting for purpose-built legal AI to become accessible for a 1-5 person firm — this is the moment that just happened.

The workflow: sign up, take the most recent legal research task you completed, run the same question through DescrybeLM, and compare the output to what you produced. That comparison is your evaluation. You'll know within one task whether this belongs in your workflow.

---

The Crossing Report covers AI adoption for professional services firm owners every Monday. Subscribe free — the first three insights every week are always free.

---

Related reading:

• The Legal AI Split: Two Categories, Two Use Cases, One Decision Framework
• The AI Billing Conversation: How Professional Services Firms Should Handle It in 2026