Phantom Clauses and Missed Liabilities: What Transactional Lawyers Get Wrong About AI in Drafting and Due Diligence
10/30/20265 min read
Phantom Clauses and Missed Liabilities: What Transactional Lawyers Get Wrong About AI in Drafting and Due Diligence
Almost every conversation about AI and legal ethics is about litigation — fabricated case citations, sanctions, court filings. That focus has left a blind spot the size of the transactional bar. Deal lawyers have adopted AI for drafting and due diligence just as fast as litigators, but the failure modes are quieter, harder to catch, and in some ways more dangerous: no judge is checking the data-room summary, and a hallucinated clause doesn't announce itself the way a nonexistent case does.
Here's what actually goes wrong when AI touches a deal, why it implicates the same professional duties litigators worry about, and how to use these tools without signing off on an error you never saw.
The duties don't change because you're not in court
Competence (Rule 1.1), diligence (Rule 1.3), confidentiality (Rule 1.6), and communication (Rule 1.4) apply to transactional practice exactly as they do to litigation. ABA Formal Opinion 512 and California's 2026 Practical Guidance both speak to AI use across the practice of law, not just to court-facing work. Legal scholarship has made the point bluntly: overreliance on AI to draft and negotiate contracts can breach an attorney's duties of competence and diligence in much the same way that filing fabricated citations does. The absence of a courtroom doesn't lower the standard of care — it just removes the backstop that catches the mistake.
The transactional failure modes no one warns you about
These are the errors specific to deal work, and they're subtle by design — AI output in contracts reads polished and confident even when it's wrong.
Phantom clauses. In due diligence, AI can "find" clauses that aren't there — misquoting contractual language or conflating terms from entirely different agreements in the data room. Picture a tool that flags 47 change-of-control provisions across a diligence set, three of which don't exist; the model manufactured them from neighboring contracts it treated as relevant context. In drafting, the mirror problem is the "phantom" market term: a clause that looks standard and defensible but has no basis in actual deal practice. You can't negotiate accurately against a market that the AI invented.
Cross-reference blindness. AI tends to read provisions in isolation, and contracts don't work that way. An AI summary might report that a license agreement contains a broad IP assignment — without catching that the assignment is conditioned on a defined term whose definition, three cross-references deep, limits it to improvements made in the first contract year. A human reviewer who understands the deal timeline catches that. The model, reading the clause as a standalone sentence, doesn't.
Missed liabilities. The flip side of hallucination is omission. AI tools can overlook important clauses and hidden liabilities entirely, and the polished summary masks the gap — then the memo gets forwarded up the chain, and the omission travels with it. The danger isn't that the output looks sloppy; it's that it looks complete.
Clause-drafting failures. Research evaluating AI clause generation finds that fluency and correctness are different things. A generated privacy clause may fail to distinguish controller and processor obligations; a limitation-of-liability provision may be subtly unenforceable; jurisdiction-specific or compliance requirements may be missed; and highly bespoke deal terms often fall outside the model's training distribution, which is exactly where it's most likely to improvise. Operative contract language isn't an answer to a question — it's enforceable text, and errors in it affect enforceability, compliance, and leverage.
The confidentiality and privilege layer — sharper in deals
Transactional work runs on exactly the material you most need to protect: draft agreements, NDAs, data-room contents, term sheets, client financials. Uploading any of that to a third-party AI platform implicates the duty of confidentiality, and — as courts have begun to hold in the litigation context — disclosing confidential or privileged material to a tool whose terms permit data use can risk destroying confidentiality or waiving privilege. The fix is the same as in any vendor relationship: appropriate legal and technical data boundaries, enterprise tools contractually barred from training on or disclosing your data, and the same vendor due diligence you'd apply to any third party receiving client information.
Communication and supervision still apply
Two quieter duties round it out. Communication (Rule 1.4) means considering whether to tell the client that AI is being used in their matter, and honoring any instruction that limits it — particularly where AI use bears on confidentiality or fees. Supervision means the deal partner remains responsible when an associate or paralegal runs the AI review; "the tool flagged it" and "the junior ran it" are not answers. The licensed lawyer owns the output.
A transactional AI protocol
Build these into how your group actually works:
Verify every AI-flagged clause against the source document before relying on it — open the operative language, don't trust the summary. Chase defined terms and cross-references by hand; that's precisely where AI reads provisions in isolation and misses the limiting language. Treat AI as a first-pass accelerator, never the reviewer of record; a qualified lawyer performs the substantive review. Vet any "market standard" the AI asserts against real precedent before you negotiate against it. Keep client deal documents off consumer-tier tools; use secure platforms with no-training, no-disclosure terms. Decide and document your client-disclosure position on AI use in the engagement. Supervise staff AI use the way you supervise any delegated substantive work.
The competence point, restated for deal lawyers
You cannot catch a phantom clause, a buried cross-reference, or a missing liability in output you've accepted on faith. Understanding where these tools fail — and reading the operative language yourself — is now part of transactional competence, not an optional sophistication. The lawyers who get burned won't be the ones who used AI; they'll be the ones who trusted a confident summary and never opened the contract.
FAQ
Do AI legal-ethics duties apply to transactional work? Yes. Competence, diligence, confidentiality, and communication apply across the practice of law. ABA Opinion 512 and California's 2026 guidance aren't limited to litigation.
What's the most common AI error in due diligence? Two: hallucinated or "phantom" provisions the AI reports but that don't exist as described, and omissions — important clauses or hidden liabilities the tool misses, hidden behind a polished summary.
Why does AI struggle with contracts specifically? It tends to read provisions in isolation, missing defined terms and cross-references that change their meaning, and it improvises on bespoke terms outside its training. Operative language demands precision that fluency doesn't guarantee.
Can I put a draft agreement or data-room documents into an AI tool? Only with appropriate data boundaries. Using a third-party platform whose terms permit data use implicates confidentiality and can risk privilege. Use secure, enterprise tools and vet the vendor.
Does using AI lower my standard of care? No. If anything, the duty to verify is heightened, because transactional errors lack the courtroom backstop that surfaces litigation mistakes.
AI genuinely accelerates deal work — but only for lawyers who know where it breaks and build verification into the workflow. If your transactional group wants a practical AI protocol that captures the speed without inheriting the phantom-clause risk, that's exactly the kind of readiness we help attorneys put in place.