The post below rails against the vices of AI-generated prose and helps you avoid them — but first, a short plug for my upcoming talk about how you should use AI:
I’ll show you legal-AI workflows you haven’t seen before — ones requiring no software beyond a subscription to Claude or ChatGPT. Join me for a live Zoom presentation and Q&A.
AI Workflows for Lawyers · November 19 · Noon–1 p.m. Pacific · $149
Court dockets, like social-media sites and online news outlets, are increasingly swamped with AI slop. I can usually detect any long stretch of AI-generated prose instantly, especially if it hasn’t been edited. When my internal AI-detector goes off, I lose interest in the piece and confidence in the author. I’m especially disturbed when AI text shows up in mainstream publications1 since I’ve been going out of my way to prefer those over online sources.2
Our writing styles evolve in response to what we read. The more we read AI-generated prose, the more its tics will enter our own prose, coming to mind (as clichés do) as default choices while we draft. Many word choices and style patterns are so overused by AI models that they are becoming skunked, to use Bryan Garner’s term. Whenever I use the previously unobjectionable formulation not X, but Y, for instance, I’m distracted by the thought that it sounds like AI and tempted to restructure the sentence.
I see three main remedies. First, don’t use AI to brainstorm for or draft briefs, personal correspondence, or anything else nontrivial; reserve it for editing, formatting, error-catching, and similar assistant-like tasks. Second, decrease your consumption of social media and online news while muting or unfollowing anyone you believe is posting slop. Use the time you save to read well-written and well-edited books (especially literary fiction) and periodicals (this also has the advantage of slowing the 24-hour news cycle and reducing its associated anxiety and time-wasting). Third, learn to detect and avoid common tells of AI prose.
This post concerns the third of these. I’ll give you a sense of some of the most common (and obnoxious) syntactic differences between AI-generated text and human prose. Learning to notice and fix such defects will also improve your writing even if you never use AI. Many of the same points are discussed in Elegant Legal Writing, which I wrote before AI was capable of plausible drafting.
The legal-brief quotations below are drawn from the 53 amicus briefs filed so far in In re OpenAI Copyright Litigation; as you’ll see, some of the amici’s attorneys took advantage of OpenAI’s wares when weighing in on the case. My firm, Rushing McCarl LLP, wrote an amicus brief for ten AI startups (including Splice, Boomy, and ProRata) in support of the plaintiffs.
AI tells and cures
Here are some of the longstanding clause and phrase patterns, common in high-quality writing and usually unobjectionable, that AI prose has made less useful by making them ubiquitous, putting them at risk of becoming clichés.
1. Antithesis and repeated sentence structures
Antithesis (not X, but Y) balances contrasting phrases through parallel syntax and rhythm. AI often uses it to end a static sentence built on be: “X is not Y, but Z.”
That is not a defect in copyright. It is how copyright’s system of limited rights produces progress.
The two-sentence contrast makes the second sentence sound like a revelation; in fact, the AI is just splitting one sentence’s worth of content into two by setting up a straw man and then knocking it over.
Variants:
• Yet the copyright question remains whether the new textbook appropriates protected expression, not whether it succeeds in taking market share from incumbents.
• The Supreme Court explained that the same copyrighted work can be used for different purposes and that the analysis turns on the challenged use, not an abstract description of the secondary work.
• These approaches do not assume that a final model can explain its own development history. They require organizations to preserve information during development.
• These instruments do not supply the governing copyright rule. They demonstrate the practical importance of preserving facts that legal institutions need to apply existing rules.
2. Balanced coordinate phrases, especially tricolons
Balanced coordinate phrases (A and B or A, B, and C and their variants) are ubiquitous in human writing, and their items ought to be parallel in rhythm, kind, and importance (ELW §§ 4.14, 4.17). But AI prose uses them reflexively, often adding vague or low-information items to fill the structure rather than shaping the structure to the content as the best writers do. This leads to monotonous and vapid prose whose excessive balance smooths out the variations found in human writing, lulling the reader to sleep.
This is syntax flawed by its flawlessness: its overpolish smooths out the human surface. The same defect occurs in AI-generated series and lists.
Lists are indispensable in legal prose. The most common form is a tricolon, a parallel series of three items — ubiquitous in legal writing long before AI. AI’s reflexive use of this structure reflects its utility, but as with any pattern, its value diminishes with overuse.
Try to detect the overexact rhythm and the inclusion of low-information items to fit a preconceived pattern:
Example 1
That flexibility has repeatedly enabled transformative technologies to emerge and flourish, notwithstanding predictions that they would harm authors, disrupt established markets, or undermine copyright itself.
The balanced pair emerge and flourish leads into a three-part series: harm, disrupt, and undermine. All five parallel verb phrases — emerge, flourish, harm authors, disrupt established markets, and undermine copyright — are vague, and their meanings overlap: it’s precisely by undermining copyright and disrupting established markets that authors are harmed. The sentence thus prioritizes syntactic balance at the expense of substance and winds up reading as filler.
Example 2
Rather than using works for their original expressive purpose, LLM training involves analyzing vast collections of works to learn patterns, relationships, and structures in language, images, and other forms of expression.
This sentence uses two three-item lists whose items conceptually overlap — patterns, relationships, and structures and language, images, and other forms of expression — creating a monotonous stream of vague abstractions.
3. Excessive syntactic emphasis, especially involving colons
Once everyone noticed AI’s overuse of em dashes to create syntactic emphasis, top AI models started overusing colons instead:
These examples demonstrate a recurring truth: technologies once criticized as threats to copyright markets often become powerful engines of public benefit when fair use permits their development.
This sentence opens with a vague generalization — indeed, a meaningless one (what’s a recurring truth?) — and pauses to emphasize it before making any point. Another flaw: phrases written in elevated registers (like recurring truth, a phrase reminiscent of a sermon or political address), already inadvisable in legal writing, sound especially discordant and amateurish when surrounded by AI slop rather than an authentic human voice.
Two more examples:
• Generative AI training fits comfortably within these recognized justifications: it uses copyrighted works to derive information about patterns and relationships across vast bodies of material and, in turn, enables a new technological capability that can furnish valuable information and services to the public.
• The real-world applications discussed above demonstrate the point: LLMs are already helping disaster victims find resources, researchers explore potential cancer treatments, healthcare providers interpret complex information, and underserved communities access essential services.
Each example uses an announcement-then-explanation structure that makes the reader wait to learn the paragraph’s propositional content, then fails to reward that wait.
The sentence structure itself, besides being tedious through overuse, breeds more flaws. In each example, nothing of substance is said before the colon, so that material is just a wind-up or throat-clearing phrase (ELW § 2.1). This opener culminates in the abstract noun phrase placed before the colon, assigning heavy mid-sentence emphasis to such phrases as recurring truth, recognized justifications, and the point and asking the reader to focus on something meaningless. The colon then introduces more abstractions. Some of the resulting phrases are merely verbose and inflated (powerful engines of public benefit); others, like derive information about patterns and relationships, are undecipherable.
Further reading
Ryan McCarl, Can You Fool Me with AI-Generated Prose?, Elegant Legal Writing (Aug. 24, 2026), link.
Ryan McCarl, AI Workflows and Ground Rules for Attorneys, Elegant Legal Writing (Aug. 31, 2026), link.
Sam Kriss, Why Does A.I. Write Like … That?, N.Y. Times Mag. (Dec. 3, 2025), link.
My book Elegant Legal Writing is available on Amazon and Audible. Please help me sustain the book’s momentum by rating it five stars and sharing my posts.
Please share this post with your networks:
Thanks for reading Elegant Legal Writing! Subscribe for free to receive new posts and support my work.
Connect with Ryan McCarl: Order the Elegant Legal Writing book | LinkedIn | Rushing McCarl LLP | Linktree | Second Stage (blog) | Bluesky | Substack | Twitter (X) | Threads
Photo by Aaron Burden on Unsplash.
My most recent suspected encounter with AI-generated text in a reputable publication was quoted passages in Rebecca Solnit’s “(Mis)Adventures in the Manosphere”, The Nation (Oct. 13, 2026). Solnit quotes Reset Tech and Equimundo’s 159-page The Grift Economy twice. I don’t believe Solnit herself used AI, but suspect that she unwittingly quoted it and the Nation’s editors failed to spot it. The quotations’ repeated antitheses, tricolons, and polished declarations are examples of the patterns I discuss:
• “There is a multi-billion-dollar industry built on the loneliness, economic anxiety, and identity of men. It does not call itself an industry. It calls itself a movement, a brotherhood, a path to success. It presents itself as the antidote to a society that has failed men—and in doing so, it sells them the very things that keep them isolated, financially drained, and psychologically dependent on the young men profiting from their struggle…. This is exploitation in plain sight. Likewise, the biggest platforms— YouTube, TikTok, X, Instagram and others—are not just passive participants in the grift economy. They carry out inconsistent and often after the fact accountability, sometimes taking down individual postings but leaving the revenue streams and ecosystem operating as it is.”
• “[The] tools needed to act on this industry are already on the books—consumer-protection law, platform-accountability frameworks, and a track record of enforcement against the same playbook aimed elsewhere. The barrier was never authority; it was framing, and naming this as fraud rather than culture war puts the existing tools within reach.”
My current traditional-media subscriptions, for anyone curious and to promote traditional media, which I feel an ethical obligation to support: The Economist (print), The New Yorker (print), The New York Times (digital), The Washington Post (digital), The New York Review of Books (print), n+1 (print), The Paris Review (print), Los Angeles Review of Books (print), Bookforum (print), Poetry (print), and Asterisk (print) — I don’t claim to read all these, of course. I also follow high-quality Substack blogs about law and AI.