A new study conducted by Anidjar & Levine documents an adoption curve that has outpaced oversight. Law firms and legal departments are integrating AI across research and drafting at scale, while courts, bar associations, and state legislatures race to define disclosure, supervision, and malpractice boundaries. The central question is no longer whether AI belongs in the courtroom ecosystem. It is how to codify safe use when error rates can alter case outcomes and public trust.
Adoption versus oversight by the numbers
The study’s synthesis shows a widening gap.
- Firm adoption: About 70 percent of firms report using AI for research, review, or drafting.
- Workflow impact: 64 percent of legal professionals believe AI will significantly affect their work within five years.
- Near‑term reliance: Document review AI usage is tracking from 74 percent in 2024 to 77 percent in 2025.
- Regulatory coverage: Only a handful of states have enacted or proposed explicit AI‑in‑law guardrails to date.
The implication is clear. Practice is moving faster than policy. That mismatch elevates liability exposure and raises the stakes on firm‑level governance.
Where risk emerges in practice
The study delineates the specific failure modes.
- Fabricated citations: Hallucinated authorities can slip into drafts when retrieval is weak or prompts are poorly constrained.
- Misapplied precedent: Summaries without precise jurisdictional anchoring lead to incorrect reliance.
- Confidentiality leaks: Improper use of public models risks data exposure when client material is fed into systems without enterprise controls.
- Over‑automation: Delegating judgment tasks to AI increases the chance of missing nuance in facts or equitable considerations.
Each failure mode maps to a corrective control. The highest risk is not AI itself but the absence of verification and supervision protocols.
Judicial and bar responses are accelerating
The study records measurable changes in courtroom and licensing expectations.
- Disclosure orders: More than 40 federal judges now require attorneys to disclose AI use in filings.
- Supervision mandates: Bar guidance in California, New York, and Florida requires lawyer oversight of any AI‑generated work.
- Emerging statutes: At least eight state efforts focus on malpractice liability, consumer protection, and transparency in legal services that deploy AI.
- Court operations: Some courts are using AI in administrative functions such as evidence summaries, but adoption remains limited by cultural and trust constraints.
These guardrails signal an important norm. AI can assist, but lawyers are accountable for the work product and must understand and manage the tools they use.
Quantifying accuracy risk
The study’s aggregation of model performance sets baselines.
- General models: 58 to 82 percent error rates without domain adaptation.
- Specialized systems: Westlaw AI hallucinated in 34 percent of tests; Lexis+ AI posted error rates above 17 percent even with retrieval augmentation.
- Professional sentiment: 74.7 percent of legal professionals identify accuracy as the primary concern; 56.3 percent cite reliability; 47.2 percent flag privacy.
These figures justify disclosure and supervision requirements. They also justify firm‑level controls that treat AI outputs as drafts requiring validation, not finished work.
Governance that matches the risk profile
Based on the study’s risk taxonomy, the governance toolkit is concrete.
- Model selection: Prefer systems with retrieval‑anchored outputs and citation verification built in.
- Prompt design: Use templates that force jurisdiction, date ranges, and source constraints.
- Verification: Run automated cite‑checkers and parallel human review before filing.
- Access control: Ensure client data remains in compliant environments with audit trails.
- Training: Require periodic attorney training on AI failure modes and supervision duties.
- Disclosure: Follow local rules on AI use and document internal supervision in the matter record.
This framework turns abstract risk into operational steps, aligning firm practice with judicial expectations.
Public trust remains the gating variable
The study highlights a durable disconnect.
- Client demand: 68 percent of clients under 45 expect their lawyers to use AI.
- Lawyer skepticism: Only 39 percent say AI improves client outcomes.
- Market signaling: 42 percent of clients would consider hiring a firm that advertises AI assistance.
Trust increases when firms show how supervision works in practice. The message that wins is not that AI is used. It is that AI outputs are validated by attorneys and disclosed where required.
Bottom line
The study conducted by Anidjar & Levine concludes that regulation and risk in courtroom AI are converging toward a simple principle. Adoption is acceptable when accuracy is controlled and accountability is clear. Disclosure rules, bar guidance, and state statutes are not obstacles to innovation. They are the conditions for durable adoption.
