Evidence · strength 2 of 5 · searched to 2 Nov 2026
Does AI help with legal work?
AI for legal analysis and drafting may help, but the evidence is not yet strong enough for a firmer verdict. 3 of 4 studies found a clear gain in time per task.
What randomised trials found when law students and lawyers did legal analysis and drafting with AI: speed, quality and invented citations.
VerdictMay help
No good evidence. Studies point to a gain, but no factor named in advance explains who gains.
ImpactGain
3 of 4 studies found a clear gain in time per task.
Evidence strength2 of 5
We have low confidence in this estimate.
Cost per user a year?
Legal AI tools are mostly priced by quote; general assistants cost less. Too varied to band.
Pilot rating under our v0.3 rules. The rules are published on our method page but not yet independently reviewed, so ratings may change as the method is checked and new studies arrive. Our method.
What the studies found
Time (headline)
Gain- Choi et al. (2024): −11.8% Time taken, client memo (smallest of four task savings; complaint −24.1%, contract −32.1%, handbook −21.1%). Self-reported.
- Choi et al. (2024): −32.1% Time taken, contract drafting. Self-reported.
- Schwarcz et al. (2026): −12% Completion time on tasks with a significant speed gain: −12% to −28% with o1-preview, −14% to −37% with Vincent AI (smallest shown).
On the legal tasks where speed improvements were identified, they ranged from 14-37% for Vincent and 12-28% for o1-preview.
- Autor et al. (2026): −10 minutes Time on the 10-day drafting task (control mean 110 minutes) (p = 0.05). Not statistically clear.
Treated lawyers spent about 10 fewer minutes on the 10-day drafting task than controls (= 0.05), whose mean was 110 minutes
- Worledge et al. (2026): −25.9% Review time per requirement (p = 2.5e-10).
users were 25.9% faster in reviewing the average requirement than unaided reviewers (p < 2.5e-10)
Output
Large gain- Schwarcz et al. (2026): +50% Productivity (quality score points per minute), significant in five of six tasks.
these AI tools significantly boost productivity in five out of six tested legal tasks, with statistically significant gains of anywhere from 50% to 130%.
Quality
Gain- Choi et al. (2024): +0.24 points on a 4.0 grade scale Blind-graded quality (4.0 grade scale), contract drafting task, GPT-4 vs no AI (95% bootstrap CI 0.07 to 0.41; p = 0.0060 (mean 3.00 to 3.24)).
only contract drafting showed a statistically significant increase in performance at the ninety-five percent level.
- Choi et al. (2024): +0.17 points on a 4.0 grade scale Blind-graded quality, complaint drafting. Not statistically clear.
- Choi et al. (2024): +0.07 points on a 4.0 grade scale Blind-graded quality, employee handbook. Not statistically clear.
- Choi et al. (2024): −0.07 points on a 4.0 grade scale Blind-graded quality, client memo. Not statistically clear.
- Schwarcz et al. (2026): +10.2 Blind-graded overall quality, o1-preview vs no AI (four of six tasks significant) (significant at 1% (three tasks) and 5% (one task)).
o1-preview's statistically significant improvements in scores ranged from approximately 10% to 28% (10.2%, 20.9%, 23.6%, and 28.1%)
- Schwarcz et al. (2026): +8 Blind-graded overall quality, Vincent AI vs no AI (four of six tasks significant) (significant at 10% level only). Not statistically clear.
those for Vincent AI ranged from 8% to 15% (8%, 12%, 13.5%, and 15.1%)
- Autor et al. (2026): +0.34 SD Blinded expert score on benchmark patent drafting task at 10 days (with AI) (p = 0.03).
AI access raised the quality of work delivered on benchmark patent drafting tasks at 10 days (0.34 SD, p = 0.03) and 90 days (0.38 SD, p = 0.01)
- Autor et al. (2026): +0.38 SD Blinded expert score on benchmark patent drafting task at 90 days (with AI) (p = 0.01).
AI access raised the quality of work delivered on benchmark patent drafting tasks at 10 days (0.34 SD, p = 0.03) and 90 days (0.38 SD, p = 0.01)
- Worledge et al. (2026): +6% Accuracy per requirement when reviewing default judgment requests (p = 1.0e-04 (Bonferroni)).
users aided by the Default Assistant were 6.0% more accurate on the average requirement than unaided reviewers (p < 1.0e-4)
Skill
Moderate gain- Autor et al. (2026): +0.32 SD Unassisted redlining of an existing patent application after three months (no AI) (p = 0.04).
Treated lawyers outperformed controls by 0.32 SD (p = 0.04), but this advantage was concentrated entirely among senior lawyers (0.45 SD, p = 0.02).
A minus sign on time means the task took less time. Bands are provisional; see how impact is banded. Effects are not pooled yet: each band is set from one result per study, and where studies use different units we show only the direction.
How sure are we?
- Starting level3 of 5, from 3 grade A or B studies.
- Down one: old models0 of 4 studies show they used a model released since May 2025 (the 18-month window); the rest used older models or do not say which.
- Publication biasnot testable with fewer than 10 studies.
- Result2 of 5. We have low confidence in this estimate.
Behind the average
- Lowest-skilled participants (Choi et al. (2024)): On complaint and contract drafting the with-AI grade line is flatter against baseline grade; slope differences 0.50 (95% CI 0.20 to 0.84) and 0.37 (0.22 to 0.74), significant at 95%. No levelling on the handbook or memo.
- Fast vs slow participants (Choi et al. (2024)): Time slopes roughly parallel; only contract drafting showed a slope difference (0.34, 95% CI -0.08 to 0.68), not significant.
- Highest baseline skill (scores on tasks done without AI) (Schwarcz et al. (2026)): AI access can lower quality for the strongest participants; o1-preview's quality gains were larger for lower-GPA students.
- Junior lawyers, drafting with AI (Autor et al. (2026)): Larger gains: +0.58 SD at 10 days (p = 0.04) and +0.60 SD at 90 days (p = 0.02).
- Senior lawyers, drafting with AI (Autor et al. (2026)): Smaller gains: +0.22 SD at 10 days (p = 0.24) and +0.30 SD at 90 days (p = 0.08).
- Junior lawyers, unassisted redlining after three months (Autor et al. (2026)): No average gain (-0.03 SD, p = 0.89); scores split into more poor and more good results.
- Requirements needing extensive document search (Worledge et al. (2026)): Largest gains: errors down by up to 62% and time by up to 34% (p < 0.05).
How, where and why it works
Why it works
1 of 4. Studies show speed gains on drafting tasks; why quality changes is less clear.
For whom and where
2 of 4. Results differ by task and by the participant's skill; the strongest participants sometimes did worse with AI.
What it takes
1 of 4. Participants got brief training; tools with legal sources behaved differently from general chatbots.
The full cost
0 of 4. No study counts cost.
Four separate scores from 0 to 4, never added up: how much the studies tell us about why it works, for whom, what it takes and what it really costs (the EMMIE method).
What to do next
Measure your pilot →If you are trying it yourself
- Never file or send AI-drafted legal text without checking every authority.
- Use AI for first drafts and summaries, not final advice.
If you are rolling it out
- Check professional conduct rules in your jurisdiction.
- Pilot with blind quality grading and a citation check.
What it takes
- Use tools that cite real sources, and check every citation.
- Keep a qualified lawyer responsible for the work.
What can go wrong
- Invented case citations: one trial found more assignments with fabricated citations when using a general reasoning model.
- The strongest participants' quality could fall with AI.
The studies
| Study | Grade | Design | People | AI model (released) | Company role |
|---|---|---|---|---|---|
| Lawyering in the Age of Artificial Intelligence Jonathan H. Choi, Amy B. Monahan, Daniel Schwarcz, 2024. Minnesota Law Review 109 (journal) | B | RCT, realistic task | 59 University of Minnesota law students completed (about 60 enrolled; 22 dropouts replaced from a waitlist); two randomised groups of about 30, each using GPT-4 on two of four tasks | GPT-4 (Mar 2023) | none |
| AI-Powered Lawyering: AI Reasoning Models, Retrieval Augmented Generation, and the Future of Legal Practice Daniel Schwarcz, Sam Manning, J.J. Prescott, Patrick Barry, David R. Cleveland, Beverly Rich, 2026. Journal of Law & Empirical Analysis 3 (journal) | B | RCT, realistic task | 127 upper-level law students (Minnesota and Michigan), six legal tasks each | OpenAI o1-preview (Sep 2024); Vincent AI (vLex, retrieval-augmented legal tool) | none |
| Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting David Autor, Tanya Rodchenko, Josh Martin, Zanna Iscenko, Scott Strand, David Pearl, Melissa Ferere, 2026. NBER Working Paper 35720 (working paper) | B | field RCT, realistic task | 133 practising patent lawyers at 11 US IP firms (12 firm cohorts); 105 completed the 10-day task, 98 the 90-day drafting, 91 the redlining task | Unreleased Google Labs AI writing assistant for professional services (underlying model not named in the sections read) | vendor-run: The study was conducted by Google |
| AI Assistance for Human Review of Default Judgments Theodora Worledge, Othman Bensouda Koraichi, Daniel Bernal, Aviv Caspi, Tatsunori Hashimoto, Carlos Guestrin, David Freeman Engstrom, 2026. arXiv:2607.01256 (preprint) | C | RCT, realistic task | 66 law students randomised to two arms; 204 case reviews per arm (12 real debt-collection cases each) | OpenAI gpt-4.1 inside a citation-grounded 'Default Assistant' (Apr 2025) | none |
Grades: A, a well-run randomised trial in real work; B, a strong natural experiment in real work or a well-run trial on realistic tasks; C, a trial with clear limits or a matched comparison; D, before-and-after, surveys or case studies. How studies are graded.
Studies we found but do not count
- Better Bill GPT: Comparing Large Language Models against Legal Invoice Reviewers (Whitehouse et al. (2025), grade D). AI alone against human reviewers, written by the vendor; nobody worked with the AI (section 4, rule 1).
- Better Call GPT, Comparing Large Language Models Against Lawyers (Martin et al. (2024), grade D). AI alone against lawyers on ten contracts, written by the vendor; nobody worked with the AI (section 4, rule 1).
These studies are real and often cited, but they fall outside our rules for this rating. What counts as a study.
Does the rating hold up?
- Run without company-run studies3 studies: strength 2, verdict no good evidence, headline band small gain.
- Run without working papers and preprints2 studies: strength 2, verdict no good evidence, headline band small gain.
How up to date is this evidence?
Method →Evidence searched to 2 Nov 2026. Studies cover models released Mar 2023 to Apr 2025. Rating computed 5 Nov 2026 under Module 15 v0.3 (pilot).
How this page was made: AI research agents searched for studies and read each paper; a second pass checked every number on this page against the paper itself, with the quote and where it appears. The rating is computed by fixed rules, not chosen. No person outside AI Adoption has reviewed it yet. This page is not yet in search engines: health, legal and finance pages wait for a named domain reviewer, and none is appointed yet. Editorial owner: AI Adoption editors. Spot a mistake? Corrections.
Disclosure: AI Adoption uses AI tools, including Claude from Anthropic, in its research and editorial work. No company pays for or has any say in a rating.