You are reading a brief and something seems off. The cases are real. The citations check out. There is no reference to a decision from a court that does not exist. But you know.
They used AI.
Or do you?
Much of the attention surrounding lawyers’ use of generative artificial intelligence has focused on its most serious failures. We previously wrote about lawyers who found themselves before courts after submitting briefs containing fabricated cases, inaccurate quotations, and other AI-generated material that had not been properly verified.
But hallucinated cases are not the only reason AI-assisted legal writing may stand out. As lawyers become more familiar with these tools—and, importantly, better at checking their work—the signs may be getting more subtle.
Sometimes, the first clue is as small as a quotation mark.
Start With the Punctuation
Consider a brief drafted in Microsoft Word that uses curly quotation marks throughout. Halfway through, a paragraph suddenly contains “straight quotation marks.” The apostrophes change too. Then an em dash appears—followed shortly by another. And another.
None of these features proves that the paragraph was generated by AI. Straight quotation marks can result from copying text from any number of sources. Lawyers were also using em dashes long before ChatGPT came along.
But the change may be worth noticing.
Text generated by AI platforms and pasted into a document can introduce typographical characteristics that differ from the surrounding text. In addition to straight quotation marks and apostrophes, look for changes in the use of hyphens, en dashes, and em dashes. There may also be unusual spaces or other characters that look ordinary on the page but behave differently when edited in Word.
Formatting can provide other clues. A paragraph may suddenly have slightly different spacing, indentation, font, or hyperlinks. And sometimes the evidence is considerably less subtle: stray asterisks, backticks, or other remnants of markdown formatting may survive the trip from an AI platform into a filing.
Individually, these things mean very little. Together—particularly when they suddenly appear in one section of an otherwise consistent document—they may give you a reason to keep reading carefully.
First, Second, and Third
The writing itself may offer additional clues.
First, generative AI likes organization. Second, it likes transitions. Third, it really likes lists of three.
Arguments become “unsupported, unexplained, and ultimately unpersuasive.” Evidence is “clear, consistent, and compelling.” Courts are asked to consider “text, structure, and purpose.”
Certain transitions may also begin to appear with surprising frequency. “Moreover.” “Furthermore.” “Importantly.” “Notably.” And, of course, “In sum.”
Then there is the construction in which something is not merely one thing; it is something more. An error is not merely procedural; it is fundamental. An omission is not simply relevant—it is dispositive. And an argument is not merely incorrect; it misunderstands the governing law entirely.
AI-generated text can also have a certain symmetry. Paragraphs may begin with a clean topic sentence, proceed through two or three supporting points, and end by neatly restating the conclusion. The next paragraph then does much the same thing. Sentences may repeatedly appear in similar lengths and structures.
Sometimes the repetition is substantive too. A proposition is stated. The next sentence restates it in slightly different language. A third sentence explains why the proposition, now stated twice, matters.
None of this is necessarily bad writing. In fact, many of these devices work precisely because they are effective writing devices. Lawyers use transitions. Lawyers use lists of three. Lawyers use em dashes. And, notably, lawyers use the word “notably.”
The more useful clue may be that the writing has changed.
Does It Still Sound Like the Same Brief?
A brief may refer to “Plaintiff” for twenty pages before suddenly referring to “the Plaintiff.” A company identified by a defined term may become “the Company.” A party name may be abbreviated differently. A section of an otherwise direct advocacy piece may suddenly spend several paragraphs carefully presenting both sides of an issue before reaching a conclusion.
Vocabulary can shift too.
Suddenly, facts “underscore” conclusions. Issues are “multifaceted.” Considerations are “crucial.” An approach is “robust.” A point that might once have been described as important is now, inevitably, “notable.”
Again, there are many explanations for these differences. Legal briefs often have multiple authors. Lawyers incorporate language from research memoranda, prior filings, emails, client materials, and drafts prepared by colleagues. A different writing style does not mean that a chatbot has entered an appearance as co-counsel.
But a change in voice may tell you where to look.
And that is where the more consequential signs begin.
Check the Cases
The most publicized examples of AI misuse in legal filings have involved nonexistent authorities. But a citation does not need to be fabricated to create a problem.
The case may exist but not support the proposition for which it is cited. The quotation may sound exactly like something a court would write but not actually appear in the decision. The proposition may appear somewhere in the case, but the pincite may be wrong.
There are subtler problems too. A district court decision may be described as establishing a rule more broadly than the opinion supports. An unpublished decision may be presented without appropriate context. Conflicting authority may disappear as “courts consistently hold.” A qualified proposition may become categorical.
Sometimes an authority may seem almost too good. The facts are remarkably similar. The court addresses precisely the issue presented. And the holding fits the argument perfectly.
Perhaps it does. But that is a good reason to open the case.
Check the Record Too
The same problem can arise with facts.
An exhibit may exist but not support the sentence citing it. Testimony may be summarized a little more broadly than the witness actually gave it. A date may change. A job title may shift. A dollar amount may be slightly different the second time it appears.
More concerning, a fact that logically should be true can find its way into a factual narrative even though nothing in the record establishes it.
These errors can be particularly difficult to spot because they are plausible. And generative AI is very good at plausible.
That matters in litigation because the facts that do not fit neatly together are often the important ones. The contract may not contain the provision one would expect. The witness may not have given the obvious answer. The chronology may have a gap. The case that initially appears directly on point may turn on a factual or procedural distinction buried later in the opinion.
A tool designed to generate coherent text may smooth over precisely the inconsistency that matters.
So, Did They Use AI?
Maybe.
But the better question may be whether it matters.
If the quotation marks change, the em dashes multiply, and “notably” makes its fourth appearance in two pages, the most useful response is probably not to prove that opposing counsel used ChatGPT.
Check the work.
Pull the case. Confirm the pincite. Search the quotation. Open the exhibit. Compare the factual assertion against the cited testimony.
There may be nothing improper about using generative AI to assist in preparing a filing, subject to applicable court rules and professional obligations. AI-generated text that has been carefully reviewed may be entirely accurate. Conversely, a brief drafted entirely by humans can contain bad citations, inaccurate quotations, and factual mistakes.
The source of the language does not determine its reliability. The review does.
Did You Spot Them?
Now, as you read this article, you may have noticed something: it sounds like AI drafted it. That’s because it did. We asked AI to draft this article using the very tells discussed above, including straight quotation marks, em dashes, lists of three, familiar transitions, and other common AI writing patterns. We wanted to show how those clues can come together on the page, rather than just tell. And if they were enough to raise your suspicions here, they should do the same when reviewing a legal brief. The key is that you don’t stop at “we know you used AI.” As we have written before, lawyers remain responsible for every citation, quotation, and legal argument they submit to the court. So, check the work and confirm whether the attorney did theirs.
Recent Comments