Playbooks
What legal and compliance teams actually worry about with AI-drafted contracts
The single biggest worry isn't that AI will write a bad clause — it's that AI will confidently write a fabricated one (a citation, a precedent, a term) that looks completely legitimate and slips past review, because the lawyer or paralegal who signed off on it is personally and professionally on the hook when it's discovered, not the AI vendor. That risk is no longer theoretical: a public tracker counted roughly 1,490 court decisions worldwide, over 1,000 in the US, where a party relied on AI-hallucinated material and a court had to respond, and sanctions have climbed from four-figure fines to $30,000 against two attorneys in a single 2026 federal appeals case.
Last updated Aug 21 · 20 min read
The case that started it: Mata v. Avianca
Mata v. Avianca, Inc. is the case every legal-AI governance conversation eventually circles back to, and it's worth understanding in specific detail rather than as a vague cautionary tale. Plaintiff's attorneys used ChatGPT to draft a legal motion. The tool generated entirely fabricated case citations and quotations — cases that did not exist, presented with the confident formatting of real legal authority. When opposing counsel and the court challenged the citations, the attorneys asked the AI tool to verify them, and it falsely reassured them the cases were real and could be found in standard legal databases (Wikipedia).
Judge P. Kevin Castel dismissed the underlying case in June 2023 and, separately, found the attorneys had acted in "subjective bad faith" under Federal Rule of Civil Procedure 11, imposing a $5,000 sanction. The court's written decision specifically identified "gibberish" in the AI-generated legal analysis submitted as part of the record — not just wrong citations, but analysis that didn't hold together on its own terms (Wikipedia).
Two things about this case matter more than the headline. First, the sanction fell on the human attorneys, not on OpenAI or any AI vendor — establishing early and clearly that lawyers are accountable for AI-generated misinformation they submit, regardless of how the tool failed them. Second, Mata v. Avianca is now frequently cited in later cases involving AI-fabricated legal content, meaning it functions as a shaping precedent for how courts approach responsible AI use in legal practice generally — not just a one-off embarrassment.
In direct response to this pattern, the American Bar Association issued its first formal ethics opinion addressing generative AI use in legal practice in July 2024. Commentary on that opinion states plainly that lawyers "cannot reasonably rely on the accuracy, completeness, or validity of content generated by [generative AI] tools" without independent verification (Wikipedia). That's the standard every contract-drafting use of AI in a law firm or legal department is now implicitly measured against: verify, don't trust.
What "AI governance" means for a legal team in practice
The 2025 State of the Legal Industry report frames this directly: given the volume of documented hallucination-driven filing errors, formal AI governance is now treated as being as critical to a firm's risk profile as the AI tool's raw capability (LawNext). That's a meaningful reframing — it means the question a legal team should be asking isn't just "is this AI tool good," but "do we have a process that catches it when the tool is wrong."
In practice, that process tends to include a few concrete elements that recur across the guidance reviewed for this article: a documented policy on which tasks AI may assist with (and which it may not), a mandatory independent-verification step before any AI-drafted language is filed or executed, a record of who reviewed and approved AI-assisted output, and — increasingly — a formal incident-tracking mechanism for when AI-generated errors do slip through, so the firm can show a court or regulator it takes the problem seriously rather than treating each incident as a surprise.
The scale of what's being governed against is no longer small. A public database maintained by legal researcher Damien Charlotin tracks roughly 1,490 court decisions worldwide, more than 1,000 of them in the United States as of May 2026, where a party relied on AI-hallucinated material and a court responded to it (TechNewsWorld). Penalties have escalated accordingly: from four-figure fines toward $15,000-per-attorney sanctions in a federal court of appeals, and the first bar suspensions tied specifically to AI filings. In March 2026, the Sixth Circuit Court of Appeals levied $30,000 in total sanctions — $15,000 each — against two attorneys for briefs riddled with fabricated citations (TechNewsWorld). Crucially, when courts sanction lawyers for AI hallucinations, they hold counsel responsible regardless of which department selected the tool or how sophisticated the vendor's own claims were — governance failures don't get outsourced to IT or procurement any more than the underlying legal error does.
How much are law firms actually using AI for contracts specifically
Here's a detail that surprises people outside the industry: contract drafting is not actually the leading AI use case in law firms right now, despite being the one most associated with "AI replacing lawyers" in popular framing. Among lawyers who use AI at all, contract drafting/templating accounts for only about 13% of use — well behind legal research (40%), drafting communications (25%), and summarizing legal narratives (23%) (LawNext). Due diligence sits even lower, at 8%.
That ordering tells you something important about risk tolerance: firms are more comfortable pointing AI at tasks with a built-in, relatively fast human check (summarizing something you're about to read anyway, drafting a routine email) than at tasks where a subtle, confidently-worded error can become a binding legal obligation if it slips through. Contract drafting sits in an uncomfortable middle zone — high-stakes enough that firms are cautious, but valuable enough (repetitive templating work, first-draft generation) that adoption is still growing rather than staying flat.
The trust gap: adoption is rising faster than confidence
This is the single most striking data point in the research for this topic. 63% of mid-sized law firms have formally adopted generative AI — yet 81% of firm leaders simultaneously express internal concerns about the technology's reliability and risk (LawNext). Adoption and distrust are rising together, not trading off against each other, which is exactly the dynamic that makes formal governance non-optional rather than a nice-to-have: firms are deploying a tool a large majority of their own leadership doesn't fully trust.
Microsoft Copilot is cited as the most commonly adopted generative AI tool across mid-sized firms in the 2025 survey — a detail worth noting because Copilot is a general-purpose productivity tool retrofitted into legal workflows, not a purpose-built legal-AI product with domain-specific hallucination safeguards baked in (LawNext).
And the industry documented 487 instances of AI errors or hallucinations in US court documents during 2025 alone, with licensed attorneys — not just self-represented litigants unfamiliar with legal procedure — responsible for 37.8% of the problematic filings (LawNext). That last figure matters: this isn't purely a pro se litigant problem that trained lawyers are immune to. More than a third of the 2025 incidents came from people who had passed the bar.
Despite all of this, optimism about AI's business impact remains high: 94% of firm leaders in the cited survey predict AI will boost revenue and improve client service, even though structural changes like billing-model transformation haven't yet materialized in practice (LawNext). The picture that emerges is a profession betting heavily on AI's upside while simultaneously documenting, in real time, a rapidly growing catalogue of its failure modes.
Practical examples
Real example — Mata v. Avianca's fabricated citations. The clearest documented instance of the exact failure mode legal teams worry about: AI-generated case citations that looked completely legitimate, cited with confident formatting, that did not correspond to any real case — and the AI doubled down when challenged, falsely claiming the cases were real (Wikipedia).
Real example — the Sixth Circuit's $30,000 sanction. A 2026 federal appeals case where two attorneys were sanctioned $15,000 each for briefs containing fabricated citations — demonstrating that this risk hasn't been solved by three years of industry awareness following Mata v. Avianca; it has, if anything, escalated in penalty severity (TechNewsWorld).
Illustrative scenario — a contract clause, not a citation. None of the documented public cases reviewed for this article involve a fabricated contract clause specifically (most documented hallucination cases involve litigation filings and case citations rather than transactional contract drafting), but the underlying mechanism transfers directly: an AI tool asked to draft an indemnification or liability-limitation clause can generate language that sounds standard and enforceable but subtly misstates the governing law, cites a superseded statute, or invents a defined term inconsistent with the rest of the agreement — and because contract review often happens under deadline pressure, that kind of confidently-wrong output is exactly the failure mode compliance teams should assume is possible even though a headline court case hasn't yet made it famous. Evidence not sufficiently verified: a specific, publicly documented instance of an AI-hallucinated contract clause (as opposed to litigation citations) causing a dispute — this scenario is presented as a reasoned extrapolation from the documented citation-hallucination pattern, not as a reported case.
Data and evidence
– ~1,490 court decisions worldwide, 1,000+ in the US (as of May 2026) involved a party relying on AI-hallucinated material, per the Damien Charlotin tracker (TechNewsWorld).
– $30,000 in total sanctions ($15,000 each against two attorneys) levied by the Sixth Circuit Court of Appeals in March 2026 for fabricated citations (TechNewsWorld).
– 487 documented instances of AI errors/hallucinations in US court documents during 2025, with licensed attorneys responsible for 37.8% of them (LawNext).
– 63% of mid-sized law firms have formally adopted generative AI; 81% of firm leaders still express reliability/risk concerns (LawNext).
– Among lawyers using AI, legal research (40%), drafting communications (25%), and summarizing legal narratives (23%) far outpace contract drafting/templating (13%) and due diligence (8%) (LawNext).
– 94% of firm leaders predict AI will boost revenue and client service, despite the reliability concerns above (LawNext).
– Microsoft Copilot is the most commonly adopted generative AI tool across mid-sized firms surveyed (LawNext).
– The ABA issued its first formal ethics opinion on generative AI in legal practice in July 2024 (Wikipedia).
Comparisons
AI contract drafting vs. traditional contract review. Traditional review is slower and depends entirely on the reviewing attorney's own diligence and knowledge, but it doesn't carry the specific hallucination risk documented above — a human reviewer might miss a real error, but won't confidently invent a citation to a case that doesn't exist. AI drafting is faster at producing first-draft language and can surface relevant boilerplate quickly, but every output requires the same independent verification standard the ABA's 2024 opinion describes — meaning the time saved on drafting has to be partially reinvested in verification, not simply banked as pure efficiency gain.
General-purpose AI (ChatGPT, Copilot) vs. purpose-built legal AI for contracts. General-purpose tools like Microsoft Copilot — the most commonly adopted tool in the 2025 survey — weren't built with legal-domain hallucination safeguards or citation-verification features specific to case law and statutory references. Purpose-built legal AI products (not named with the same level of adoption data in the sources reviewed here) are more likely to integrate directly with verified legal databases, reducing but not eliminating the fabrication risk. The tradeoff is cost and workflow integration versus domain-specific safety features — and the 2025 survey data suggests firms are currently prioritizing the former, given Copilot's dominant adoption.
Real-world use cases
– First-draft generation for routine, low-stakes clauses. Legal teams increasingly let AI produce a first pass at standard boilerplate (NDAs, routine vendor terms) precisely because these are the lowest-stakes, most template-like documents — closer to the "drafting communications" use case (25% adoption) than to novel, high-value negotiated terms.
– Summarizing long agreements before human review. Falls under the "summarizing legal narratives" bucket (23% adoption) — using AI to produce a digest of a long contract so a human reviewer knows where to focus attention, rather than replacing the reviewer's read entirely.
– Legal research to support contract negotiation. The single largest AI use case in law (40%) — using AI to quickly surface relevant precedent or statutory context that informs how a clause should be negotiated, with the same independent-verification obligation applying to anything cited.
– Post-incident governance tightening. Firms that experience or narrowly avoid a hallucination incident are reported to respond by formalizing a verification checkpoint and incident-tracking process — the reactive version of the proactive governance approach recommended by the 2025 industry report.
Common mistakes
– Treating AI-generated citations as pre-verified — the single mistake at the center of Mata v. Avianca and a recurring theme across the ~1,490 tracked court decisions involving AI hallucinations.
– Assuming a general-purpose productivity tool (like Copilot) has legal-domain safeguards it wasn't built with — general AI tools weren't designed with the specific citation-verification rigor legal work demands.
– Believing sanctions are a "pro se litigant problem" — licensed attorneys accounted for 37.8% of documented 2025 hallucination-driven filing errors, not just self-represented parties unfamiliar with procedure.
– Deploying AI for contract drafting with no documented verification step — exactly the gap formal AI governance is meant to close, and exactly what the ABA's 2024 opinion says lawyers cannot skip.
– Assuming adoption implies confidence — 63% adoption alongside 81% leadership concern about reliability shows firms are deploying AI faster than they're resolving their own doubts about it, which is a governance gap, not a solved problem.
– Re-asking the AI tool to "verify" its own fabricated output — this is literally what happened in Mata v. Avianca, where the tool falsely reassured the attorneys its fabricated cases were real.
Best practices
– Require independent, human verification of every citation, quotation, and factual claim in AI-assisted drafting before it's filed or executed — never ask the AI tool itself to self-verify.
– Document a clear, written policy on which tasks AI may assist with and which require full traditional drafting, rather than leaving it to individual attorney discretion.
– Track AI-assisted drafting with a record of who reviewed and approved the output, so the firm can demonstrate a functioning process if a problem does surface.
– Treat contract drafting as a higher-risk AI use case than research or summarization, and apply a correspondingly stricter review standard, consistent with its comparatively low current adoption rate (13%) relative to lower-stakes use cases.
– Build an internal incident-tracking mechanism specifically for AI-generated errors, so patterns (a particular tool, a particular clause type) can be identified and addressed rather than treated as isolated surprises.
– Stay current on your jurisdiction's specific ethics guidance — the ABA's 2024 opinion is a floor, not a ceiling, and several state bars have issued their own supplementary guidance.
– Don't conflate a general-purpose tool's productivity gains with legal-domain reliability — evaluate contract-drafting AI specifically on its citation and factual accuracy, not just its writing fluency.
Key takeaways
– Mata v. Avianca established the governing standard: lawyers are personally accountable for AI-generated fabrications they submit, and the AI reassuring you it's correct doesn't count as verification.
– Sanctions have escalated sharply — from $5,000 in 2023 to $30,000 against two attorneys in a 2026 federal appeals case — while a public tracker counts roughly 1,490 related court decisions worldwide.
– Contract drafting is currently a minority AI use case in law (13%) compared to research (40%) and communications (25%), reflecting appropriate caution around its higher stakes.
– Adoption (63% of mid-sized firms) is outpacing confidence (only 19% of leaders lack reliability concerns) — a gap that formal AI governance is specifically meant to close.
– Licensed attorneys, not just pro se litigants, are responsible for over a third of documented 2025 AI-hallucination filing errors — this is a trained-professional problem, not a layperson problem.
Relevant tools.scult.in resources
Teams building out internal AI-use policies, review checklists, or contract-drafting guardrails may find useful structured starting points in the legal & compliance prompts collection — built to support a documented, human-verified workflow rather than unsupervised AI drafting.
For a related, free starting point, try the AI Visibility Checker.
Frequently asked questions
What is Mata v. Avianca?
A 2023 case where plaintiff's attorneys used ChatGPT to draft a legal motion that included entirely fabricated case citations; the court sanctioned the attorneys $5,000 and dismissed the underlying case (Wikipedia).
Can a lawyer be sanctioned for AI-fabricated citations?
Yes — Judge P. Kevin Castel's Mata v. Avianca ruling and the Sixth Circuit's 2026 $30,000 sanction against two attorneys both confirm this is an active, escalating risk, not a hypothetical one.
What does "AI hallucination" mean in a legal context?
A confidently generated but factually false output — a case citation, quotation, or fact that does not exist or is inaccurate, presented with no indication of uncertainty.
Does the American Bar Association have guidance on generative AI?
Yes, the ABA issued its first formal ethics opinion on generative AI use in legal practice in July 2024, stating lawyers cannot reasonably rely on GAI-generated content's accuracy without independent verification (Wikipedia).
How common are AI hallucinations in court filings now?
A public tracker counted roughly 1,490 court decisions worldwide (1,000+ in the US) as of May 2026 involving reliance on AI-hallucinated material (TechNewsWorld).
What percentage of lawyers use AI for contract drafting specifically?
About 13% of lawyers who use AI at all use it for contract drafting/templating — well behind legal research (40%) (LawNext).
Do law firm leaders actually trust the AI tools they've adopted?
No — 63% of mid-sized firms have adopted generative AI, but 81% of firm leaders still express reliability and risk concerns (LawNext).
Which AI tool is most used at law firms?
Microsoft Copilot is cited as the most commonly adopted generative AI tool across mid-sized firms in the 2025 survey (LawNext).
What's the difference between AI governance and AI capability in a legal context?
Capability is how good the tool is at the task; governance is the process (verification, documentation, accountability) that catches the tool when it's wrong — the 2025 industry report treats the latter as now equally critical to a firm's risk profile.
Is it only pro se litigants who submit AI-hallucinated content?
No — licensed attorneys accounted for 37.8% of the 487 documented AI-error filings in US courts during 2025 (LawNext).
Why did the judge in Mata v. Avianca call the AI-generated content "gibberish"?
Because the AI-generated legal analysis submitted didn't hold together substantively, beyond just citing nonexistent cases — the court identified analytical incoherence as part of its sanctions finding.
Why does Mata v. Avianca keep coming up in newer cases?
It's frequently cited as precedent in subsequent cases involving AI-generated false citations, making it a foundational reference point for how courts assess AI misuse.
What is Federal Rule of Civil Procedure 11 and why does it matter here?
It's the rule under which Judge Castel found the Mata v. Avianca attorneys acted in subjective bad faith; it generally requires attorneys to certify that filings are well-grounded in fact and law, and it's the mechanism used to sanction AI-hallucination-driven filings.
How much have AI hallucination sanctions grown since Mata v. Avianca?
From a $5,000 sanction in 2023 to a $30,000 combined sanction against two attorneys in a single 2026 federal appeals case — a significant escalation in penalty severity (TechNewsWorld).
Who tracks AI hallucination cases in court filings?
Legal researcher Damien Charlotin maintains a public database tracking AI-hallucination-related court decisions worldwide (TechNewsWorld).
What's the practical difference between using AI for legal research versus contract drafting?
Research findings are typically checked against the underlying source before being relied upon in an argument; contract drafting output can become a binding document if not caught before execution — the stakes and verification urgency differ even though both need independent checking.
Are firm leaders optimistic or pessimistic about AI overall?
Both simultaneously — 94% predict AI will boost revenue and client service, while 81% express reliability/risk concerns about the same technology (LawNext).
Has AI adoption changed law firm billing models yet?
Not according to the 2025 report — despite high optimism about AI's business impact, structural changes like billing-model transformation haven't yet materialized in practice (LawNext).
What's "shadow AI" in a law firm context?
Unofficial, unsanctioned use of AI tools by staff without firm governance or oversight — a risk explicitly named in 2026 industry coverage of law firms grappling with hallucinated legal logic (TechNewsWorld).
Does AI hallucination risk apply only to litigation, or to contracts too?
The most extensively documented public cases involve litigation filings and case citations; the same confident-fabrication mechanism plausibly extends to contract drafting, though a headline contract-specific case wasn't identified in the sources reviewed for this article.
How do I review AI-drafted contract language safely?
Independently verify every citation, defined term, and factual claim against a primary source before relying on it — never treat AI output as pre-verified, and never ask the AI itself to confirm its own accuracy.
How do I keep client data confidential when using AI contract tools?
Confirm the specific tool's data-handling and retention policy before inputting client information, and prefer tools with contractual confidentiality commitments and enterprise-grade data controls over consumer-facing general chatbots.
How do I build AI governance for a legal team from scratch?
Start with a written policy on approved use cases, a mandatory human-verification checkpoint, a review-and-approval record, and an incident-tracking process for when errors do surface — the elements the 2025 industry report frames as now essential to a firm's risk profile.
How do I train associates to avoid AI hallucination risk?
Make the ABA's 2024 opinion and Mata v. Avianca required reading, and build "verify before you file/send" into onboarding as a non-negotiable habit rather than an optional best practice.
How do I decide which contract tasks are safe to delegate to AI?
Lower-stakes, highly templated language (routine boilerplate) is a safer starting point than novel, heavily negotiated, high-value terms — matching the low current adoption rate (13%) for contract drafting relative to lower-stakes use cases like summarization.
Advanced: does using a purpose-built legal AI tool eliminate hallucination risk?
No — it can reduce it through better integration with verified legal databases, but the ABA's verification standard applies regardless of which tool is used; no tool is described as eliminating the risk entirely in the sources reviewed.
Advanced: is a law firm liable if a client submits an AI-hallucinated contract clause the firm reviewed and approved?
The sanctions precedent in Mata v. Avianca suggests the reviewing attorney bears responsibility for what they submit or approve, regardless of who or what originally drafted it — the accountability follows the human sign-off.
Advanced: how should a firm respond after an AI hallucination incident is discovered internally before filing?
Treat it as a near-miss worth formal review — document what happened, tighten the verification process at the point it failed, and consider whether it reveals a broader gap in how the tool is being used across the firm.
Advanced: does cloud-based AI contract drafting create additional data-security obligations for regulated industries?
Yes in principle — regulated-industry clients (financial services, healthcare) typically carry additional data-handling obligations that should inform which AI tools are approved for use on their matters, though specific regulatory requirements vary by sector and aren't detailed in the sources reviewed here.
Advanced: what's the state of AI-hallucination case law outside the US?
The Charlotin tracker's 1,490 total decisions include activity outside the roughly 1,000 US cases, indicating this is an international, not purely domestic, judicial issue, though country-by-country detail wasn't broken out in the sources reviewed.
Advanced: how are bar associations responding beyond the ABA's 2024 opinion?
Several state bars have reportedly issued supplementary guidance following the ABA's lead, though the specific content of each state's guidance wasn't itemized in the sources reviewed for this article.
Advanced: is there a standard AI-governance framework the legal industry has converged on?
Not a single named standard identified in the sources reviewed — the 2025 industry report describes governance as now critical without pointing to one universally adopted framework, so most firms appear to be building their own policies rather than adopting a shared template.
AI contract drafting vs. traditional contract review — which is actually faster once verification time is included?
Not independently quantified in the sources reviewed — AI drafting is faster at producing a first draft, but the mandatory verification step required by the ABA's guidance offsets some of that time savings, and no source here measures the net difference.
General-purpose AI (Copilot/ChatGPT) vs. purpose-built legal AI — which carries less hallucination risk?
Purpose-built legal AI tools integrating directly with verified legal databases are plausibly lower-risk for citation fabrication specifically, but Microsoft Copilot — a general-purpose tool — remains the most widely adopted option in practice according to the 2025 survey, suggesting cost and workflow fit are currently outweighing that theoretical safety advantage for many firms.
Is AI contract drafting a bigger risk at a law firm or an in-house legal department?
Not directly compared with data in the sources reviewed — the underlying hallucination risk mechanism is the same in either setting, though in-house teams may have fewer formal verification resources than a law firm's associate/partner review chain.
My firm is considering Microsoft Copilot vs. a legal-specific AI tool for contract work — what should we weigh?
Copilot offers broader workflow integration and is the most commonly adopted option industry-wide, but wasn't purpose-built for legal citation verification; a legal-specific tool may offer stronger domain safeguards at a narrower workflow fit and likely higher cost.
Is it worth paying more for a legal AI tool marketed as "hallucination-resistant"?
This is a reasonable question to press vendors on directly — ask for their specific verification methodology and any independent audit of accuracy, rather than accepting a marketing claim of "hallucination-resistant" at face value.
How does AI contract review compare to traditional outside-counsel review on cost?
Not directly compared with independent data in the sources reviewed — the tradeoff in principle is AI's speed and lower marginal cost per document against outside counsel's judgment and accountability, but a verified cost comparison specific to contract drafting wasn't available in the sources used here.
My associate submitted an AI-drafted brief with a fabricated citation — what do we do now?
Treat it with the same urgency as any other filing error: notify the court promptly if it's already been submitted, correct the record, and review your firm's verification process to identify where the checkpoint failed, following the pattern set by Mata v. Avianca's aftermath.
We just got hit with a Rule 11 motion over AI-generated content — what's our exposure?
Exposure has historically ranged from four-figure sanctions (Mata v. Avianca's $5,000) to $15,000 per attorney in more recent 2026 federal appeals rulings — consult counsel immediately and be prepared to demonstrate what verification process was or wasn't followed.
Our AI tool keeps generating citations we can't verify — is this normal?
This is the exact documented failure mode behind the ~1,490 tracked hallucination cases; treat any citation you can't independently verify as unusable, regardless of how confident or well-formatted the AI's output looks.
How do we know if our current AI governance is actually adequate?
If you can't produce a documented policy, a verification checkpoint, and a review-approval record for AI-assisted work on demand, your governance likely falls short of what the 2025 industry report frames as now standard practice.
Our AI tool "confirmed" a citation was real when we asked it directly — can we trust that?
No — this is precisely what happened in Mata v. Avianca, where the AI falsely reassured the attorneys that its fabricated citations were real; independent verification must happen outside the tool itself.
We're seeing internal "shadow AI" use we didn't approve — how big a risk is this?
Significant — unofficial AI use without governance oversight is explicitly named as a current risk area in 2026 industry coverage, and it means your firm may be exposed to hallucination risk you have no visibility into (TechNewsWorld).
Our leadership is split on whether AI contract drafting is worth the risk — is that unusual?
No — it mirrors the industry-wide pattern where 63% of firms have adopted generative AI while 81% of leaders still express reliability concerns; disagreement at this stage is the norm, not an outlier.
Should our legal team adopt AI for contract drafting given all these risks?
Adoption itself isn't the risky part — 63% of mid-sized firms have already done so — the risk is adopting without the governance (verification, documentation, accountability) that the 2025 industry data suggests most firms are still building out.
What should we ask a legal AI vendor before buying a contract-drafting tool?
Ask specifically how the tool sources and verifies citations, what its documented hallucination rate is (and how that was measured), what data-confidentiality guarantees it offers, and whether it integrates with a human-review workflow rather than encouraging unsupervised use.
Is Microsoft Copilot a good choice specifically for contract drafting?
It's the most widely adopted generative AI tool at mid-sized firms, but that reflects broad productivity adoption more than a specific endorsement for contract-drafting accuracy — evaluate it on its own citation-and-fact accuracy for legal work rather than assuming general popularity implies domain-specific reliability.
Is it worth investing in a dedicated legal-AI governance consultant or tool given the sanctions trend?
Given the documented escalation from $5,000 to $30,000+ sanctions and roughly 1,490 tracked hallucination cases, formal governance investment is increasingly framed by industry reporting as a risk-management necessity rather than a discretionary upgrade.
What's the single highest-leverage governance change a legal team can make right now?
Make independent verification of every AI-generated citation and factual claim a mandatory, documented step before anything is filed or executed — it's the one control that would have prevented the core failure in Mata v. Avianca and in the more recent Sixth Circuit sanctions case alike.
Sources
- https://en.wikipedia.org/wiki/Mata_v._Avianca
- _Inc.
- https://www.lawnext.com/2025/legal-industry-reaches-ai-tipping-point
- https://www.technewsworld.com/story/law-firms-grapple-with-hallucinated-legal-logic-shadow-ai-180338.html
- https://www.nortonrosefulbright.com/en-us/knowledge/publications/792d8bf3/ai-in-litigation-update-on-gen-ai-sanctions-in-2026
