Stop Making Your Legal Documents Look AI-Generated
This post uses hypothetical scenarios for illustrative purposes only. It does not describe any actual client, transaction, or representation, and is not legal advice.
A pattern has started showing up in transaction documents over the past eighteen months, and it’s visible within the first two pages. The agreement arrives as a Word document with default blue hyperlinks in the table of contents. The definitions section runs three pages long, defining terms that appear exactly once in the operative provisions. There’s a mutual indemnity section even though only one party has any meaningful indemnification obligation. The notice provisions include a paragraph about notices sent by “nationally recognized overnight courier,” copied verbatim from a form that was last updated when FedEx was the default delivery mechanism for signature pages. And scattered throughout the document are phrases like “it is important to note that” and “for the avoidance of doubt,” deployed so frequently that they’ve become structural filler rather than clarifying language.
None of these are wrong in the way a misspelled party name is wrong. They’re wrong in the way a suit that doesn’t fit is wrong: they signal that nobody with judgment reviewed the output. And the people reading your documents, investors’ counsel, diligence teams, opposing transaction lawyers, transfer agents, can see it immediately.
The Formatting Tells Are Specific and Fixable
The first and most obvious tell is hyperlink formatting. Microsoft Word’s default style for hyperlinks is blue, underlined text. When a document is drafted natively in Word by someone who works in legal documents regularly, the hyperlinks in the table of contents and cross-references are formatted to match the surrounding text: black, same font, no underline. When a document is generated by an AI tool and pasted into Word, the default blue hyperlink style persists. This is a five-second fix (select all, change the hyperlink style), but leaving it in place tells the reader that the document went from generator to send without a human formatting pass.
The second tell is excessive bullet-point formatting. Legal documents have used numbered paragraphs, lettered sub-clauses, and Roman numeral sub-sub-clauses for decades, and for good reason: enumerated provisions create unambiguous cross-reference targets. “Section 3.2(a)(iii)” points to exactly one place. A bullet point, by contrast, has no inherent reference identifier. AI drafting tools default to bullet points because they’re optimized for readability in a chat interface, not for legal operability. When those bullets survive into a transaction document, they create provisions that cannot be precisely referenced in a bring-down certificate, a closing condition, or a side letter.
The third tell is the presence of a defined-terms section that defines terms used only once. Defined terms exist for a reason: they allow a complex concept to be stated precisely once and referenced concisely thereafter. If “Material Adverse Effect” appears fourteen times in an agreement, a defined-terms section earns its space. But when an AI-generated draft includes a definitions section that defines “Business Day” (used once, in the notice provision), “Governmental Authority” (used once, in a representation), and “Knowledge of the Company” (used once, in a single disclosure qualifier), the definitions section is performing no useful function. It’s adding two pages of text that a reviewer must cross-reference against the operative provisions, finding each time that the defined term could have been stated in plain English at the point of use without any loss of precision.
The Substantive Tells Are More Damaging
Formatting issues can be fixed in an afternoon. Substantive tells take longer to identify and create real problems when they survive into executed documents.
The most common substantive tell is the mutual indemnity in a transaction where indemnification should run in only one direction. In a typical stock purchase agreement, the seller indemnifies the buyer against breaches of the seller’s representations. The buyer might indemnify the seller against breaches of the buyer’s representations, but the buyer’s representations are usually limited to authority, organization, and investment intent, making the buyer’s indemnification obligation functionally narrow. An AI-generated draft will often produce a fully symmetric indemnification section, with identical baskets, caps, and survival periods for both parties, because the training data includes forms where symmetry was appropriate. In a seed-stage acquisition where the founders are selling and the acquirer is buying, a symmetric indemnity is structurally wrong, and its presence tells the acquirer’s counsel that nobody analyzed the indemnification economics for this specific deal.
Another substantive tell is the inclusion of notice provisions that don’t match the transaction’s actual communication patterns. A hypothetical early-stage company with three employees doesn’t need a notice provision requiring delivery “to the attention of the General Counsel, with a copy to the Chief Financial Officer, at the following address.” The company doesn’t have a general counsel or a CFO. It has a founder who reads email on a phone. A notice provision should reflect how the parties actually communicate, and an AI-generated provision that includes a multi-officer notice chain signals that the drafter didn’t tailor the document to the actual parties.
Generic hedging language represents another category. Phrases like “it is important to note that,” “for the avoidance of doubt,” and “notwithstanding the foregoing” have legitimate uses, but AI drafting tools deploy them as transitions between paragraphs, the way a student uses “moreover” in an essay. In a well-drafted legal document, “notwithstanding the foregoing” creates a specific exception to a specific preceding provision. When it appears at the start of a paragraph that doesn’t actually conflict with anything above it, a careful reader knows the drafter didn’t trace the logical relationship between provisions.
What to Actually Remove
General advice to “keep it lean” or “tighten up the drafting” doesn’t help because it doesn’t tell you what to cut. Here is what to cut, specifically.
First, remove every defined term that is used fewer than three times in the operative provisions. Replace it with the substantive language at the point of use. If “Business Day” appears once, write “any day other than a Saturday, Sunday, or federal holiday” in the sentence where it matters, and delete the definition.
Second, remove or rewrite every notice provision to reflect the actual officers and communication methods of the actual parties. If the company’s founder is the only person who will receive notices, the notice provision should say so. Email notice should be permitted if that is how the parties will actually communicate, which for early-stage transactions it almost always is.
Third, examine every indemnification, representation, and covenant provision for inappropriate symmetry. If the transaction’s risk profile runs primarily in one direction, the document’s protective provisions should run primarily in that direction. Delete the mirror-image provisions that the AI added for structural completeness.
Fourth, search the document for “it is important to note that,” “for the avoidance of doubt,” “it should be noted that,” and “as discussed above.” Delete each instance. Read the sentence without the phrase. If it still makes sense (it almost always does), leave the phrase out.
Fifth, update all hyperlink formatting to match the document’s body text style. This takes less than a minute in Word’s styles pane and eliminates the single most visible formatting tell.
Why This Matters Beyond Aesthetics
The concern here isn’t that AI-assisted drafting is inherently problematic. Used well, AI tools can accelerate first-draft production significantly, and the guardrails available for AI-assisted contract drafting are improving. The concern is that an unreviewed AI draft communicates something about the drafter’s process, and what it communicates is unflattering.
When an investor’s counsel receives a stock purchase agreement with symmetric indemnities, default-blue hyperlinks, and a three-page definitions section for a five-page agreement, the message received is: this document was generated and sent without substantive review. That assessment, whether or not it’s fair, colors the entire negotiation. Requested changes get scrutinized more carefully. Representations get pushed back on more aggressively. The transaction’s overall friction increases because the counterparty has lost confidence in the document’s quality.
In a diligence context, the effect is even more direct. A potential acquirer’s counsel reviewing a data room full of corporate documents can distinguish between documents drafted by experienced counsel and documents generated by AI tools. As the ABA Model Rules of Professional Conduct make clear, competent representation requires the legal knowledge, skill, thoroughness, and preparation reasonably necessary for the representation. A document that contains structural tells of unreviewed AI output raises questions about whether that standard was met.
The fix isn’t to stop using AI tools. The fix is to review AI output with the same rigor you’d apply to a first draft from a junior associate: check the formatting, verify the substantive provisions against the actual transaction, remove provisions that don’t belong, and make sure the document reads like it was prepared for this deal, by someone who understood it.
Counterparties can tell the difference. So can courts reviewing corporate governance documents. The goal is a document that reflects judgment, not just generation.
If you are reviewing transaction documents and want to ensure they reflect the professionalism your deal requires, feel free to reach out to our firm manager, Magda, at Magda@montague.law, or fill out our contact form. Mention you read this post.
