How language models rewrote legal drafting and review

A first-year associate opens a mutual non-disclosure agreement (NDA) at 9 a.m., billing sheet ready. Ten hours later, after cross-checking every defined term, hunting for a missing indemnification clause, and reformatting the survival provisions, the markup is done. The same document, run through an AI-assisted review tool in 2026, produces a comparable first pass in about forty minutes.

That gap is the story of this lesson. But the interesting part is not the speed. It is knowing exactly where the model earns its keep and where it quietly makes things up.

What a language model actually does to a contract

A large language modellarge language modelA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → (LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition →) is a system trained to predict the most likely next chunk of text given everything before it. It has read enormous volumes of documents, including contracts, statutes, and case law. It does not "understand" your deal. It recognizes patterns.

For legal drafting, that pattern-matching is more useful than it sounds. Contracts are highly structured. An indemnification clause looks like an indemnification clause across thousands of agreements. A governing-law provision follows a recognizable shape. The model has seen the template a million times.

So when you paste an NDA and ask "what standard protections are missing here," the model compares your text against its internalized sense of what a complete NDA usually contains.

Where it genuinely shines

Catching omissions. The model is excellent at noticing that a document lacks something common. Missing survival clause (the provision stating which obligations continue after the contract ends)? It flags it. No carve-out for information already public? Flagged. No return-or-destroy provision for confidential materials? Flagged.

Consistency checks. Defined terms are a classic source of drafting error. If a contract defines "Confidential Information" in Section 1 but then references "Proprietary Information" in Section 4, the model catches the mismatch faster than a tired human at hour nine.

First-draft generation. Ask for a mutual NDA governed by New York law with a two-year term, and you get a competent starting draft in seconds. It will not be final, but it eliminates the blank page.

Plain-language summaries. Handing a client a three-sentence summary of what a clause means, in readable English, is now trivial. This alone changes how associates communicate with non-lawyer stakeholders.

The dangerous part: silent invention

Here is the failure mode that keeps general counsel awake.

Because the model predicts *plausible* text, it can insert protections that sound right but were never negotiated. Ask it to "clean up" a one-sided NDA and it may quietly add a mutual indemnification obligation, because most NDAs it has seen are mutual. Your client never agreed to indemnify anyone.

This is often called hallucinationhallucinationA hallucination is when an AI model generates output that is fluent and confident but factually wrong, fabricated, or unsupported by its source data.View full definition →: the model generating confident, fluent output that is factually wrong or unsupported. In legal work, hallucinationhallucinationA hallucination is when an AI model generates output that is fluent and confident but factually wrong, fabricated, or unsupported by its source data.View full definition → is not a typo. It is a liability.

Concrete examples of silent invention:

• Adding a limitation of liability cap that was never part of the term sheet.
• Inserting a governing-law clause for a jurisdiction nobody selected.
• "Correcting" a deliberately asymmetric clause into a symmetric one, erasing a negotiated advantage.
• Citing a statute or case that does not exist. This has already produced real sanctions. In the widely reported *Mata v. Avianca* matter, lawyers submitted a brief with AI-fabricated case citations and were sanctioned by the court. You can read the actual court opinion imposing sanctions to see how seriously judges treat this.

The lesson: the model is a strong assistant for finding what is *missing* and a dangerous author for deciding what *should be there*.

A practical workflow that respects both

The professionals getting real value are not asking the model to "do the contract." They are decomposing the work.

Step 1: Separate detection from generation

Use the model to *identify* issues, not to *rewrite* silently. A detection prompt keeps you in control:

Review the attached NDA. Do NOT rewrite it.
Produce a table with three columns:
1. Clause or issue
2. What is present vs. what is missing
3. Why it matters (one sentence)

Do not add or suggest new obligations unless flagged
as "MISSING - optional."

This forces the model into a checklist role. You still decide what goes in.

Step 2: Verify every citation and every added clause

Treat any statute, case, or regulation the model mentions as unverified until you check it in a real source (your firm's research database, the primary text). If the model proposes new language, ask: "Did the client agree to this?" If you cannot answer yes from the term sheet, cut it.

Step 3: Keep a human on defined terms and cross-references

Models are good at flagging these but can also introduce new inconsistencies when editing. A final human pass on definitions is non-negotiable.

🎬 [VIDEO: "How AI is Changing Legal Work" — youtube.com/results?search_query=AI+legal+contract+review — an accessible overview of AI-assisted contract review workflows in law firms]

Confidentiality: the constraint people forget

Before pasting a client contract into any tool, ask one question: where does this text go?

Consumer chatbots may retain inputs for training. That can breach client confidentiality and, depending on jurisdiction, professional conduct rules on protecting client information. Most bar associations have issued guidance; the American Bar Association's Formal Opinion 512 addresses generative AI and lawyers' duties directly, including competence, confidentiality, and billing.

Practical rules many firms now enforce:

• Use enterprise tools with a no-training data agreement, meaning the vendor contractually agrees not to use your inputs to train its models.
• Never paste privileged material into a free public tool.
• Redact client-identifying details when a secure environment is not available.

Reading the economics honestly

Does forty minutes replace ten hours across the board? No. Be precise about where the savings land.

The first-pass review of a standard, moderate-complexity document compresses dramatically. That is real. But the tasks that carry the most risk (deciding whether an unusual indemnity is acceptable, negotiating a carve-out, judging whether a client should accept an asymmetric term) still require experienced judgment.

So the work does not vanish. It shifts. Junior time spent on mechanical review shrinks. Senior time spent on judgment stays or grows in relative importance. This has implications for the billable-hour model, since clients increasingly resist paying associate rates for work a tool does in minutes. Many firms are moving toward fixed fees for commoditized document review as a result. Treat any specific efficiency figure you hear as an estimate; results vary hugely by document type and tool.

A quick self-test for any AI output

Before you rely on an AI-assisted markup, run this three-question check:

1. Did it invent anything? Compare every added clause against the deal terms.
2. Did it verify anything external? Every citation must be confirmed independently.
3. Did it change the balance of the deal? Look specifically for clauses that became "fairer" without instruction.

If you cannot answer all three confidently, the review is not finished.

Where this is heading in 2026

The frontier is not bigger models. It is retrieval-grounded systems: tools that pull from a verified, firm-approved clause library and cite the source of every suggestion, rather than generating from memory. This reduces hallucinationhallucinationA hallucination is when an AI model generates output that is fluent and confident but factually wrong, fabricated, or unsupported by its source data.View full definition → because the model is quoting a known document instead of inventing one.

The mature stance is neither hype nor refusal. It is supervised delegation: let the model do the tireless detection work, and keep the irreducibly human work (judgment, negotiation, accountability to the client) firmly with the lawyer.

Key Takeaways

• AI excels at detection, not decision. It reliably flags missing clauses, inconsistent defined terms, and formatting errors. Let it find problems; you decide the fix.
• Silent invention is the core risk. Models add plausible clauses and fake citations that were never negotiated or never existed. Verify every addition against the actual deal terms and every citation against a primary source.
• Confidentiality comes first. Use enterprise tools with no-training-data terms, and never paste privileged client material into free public chatbots. Check your bar's guidance.
• The savings are real but uneven. Mechanical first-pass review collapses; high-judgment negotiation does not. Expect the work to shift, not disappear, and pressure the billable-hour model.
• Prefer retrieval-grounded tools that cite an approved clause library over tools that generate from memory alone.