Why machine-sounding prose costs lawyers credibility, and how to remove it 

“We have already drafted a first version of the contract. Could you please quickly check it?” Every lawyer knows what “quickly check” can mean. The client has produced something that looks complete enough to require only a few corrections. Ten minutes into the review, the lawyer realizes that the document needs to be rebuilt from scratch. 

The drafting is fluent. The formatting is immaculate. But the "Buyer" is also referred to as the "Acquirer" and the "Purchaser." The liability clause seeks to “strike an appropriate balance” without specifying where that "balance" lies. The contract, governed by German law, contains concepts imported from US litigation. The document has the surface of finished legal work and the substance of an unconstructed conversation between laypeople. The lawyer has little doubt that the client used a generic generative AI tool. For the law firm, the situation has one consolation: the “quick check” can translate into a substantial amount of billable work. 

The reverse situation is less advantageous: A client receives work from their lawyer and assumes at first glance that it is AI output subject to only superficial review. In reality, the strategic advice may reflect years of experience, the reasoning may be sound and the authorities properly researched, but the sentences still sound like every generic AI answer the client has read that week. 

Artificial wrapping can depreciate genuine expertise. 

"Der Ton macht die Musik", as Germans say 

Large language models do not all write identically, but many produce a similar register. Their output tends to be fluent, emphatic, highly structured and "confidently vague". We call this register “AI accent”. 

The accent reflects how the models are built and trained. They learn patterns from enormous collections of existing text, much of it drawn from websites, marketing materials, essays and professional publications. During subsequent tuning, answers that appear polished, helpful and complete are rewarded. As a result, models tend to overproduce the linguistic features associated with those qualities, like an overachieving school pupil who has absorbed every technique for “good essay writing” and then been pumped full of steroids. 

The result is prose that often tastes like artificial banana flavor: immediately recognizable, oddly uniform, clearly detached from the "real thing", and therefore slightly nauseating. 

The phenomenon can be measured. In a 2025 study published in Science Advances (1), Dmitry Kobak, Rita González-Márquez, Emőke-Ágnes Horvát and Jan Lause examined 15.1 million biomedical abstracts published between 2010 and 2024. Their method first showed what happens to language when the world itself changes: During the Covid-19 pandemic, terms such as pandemic, respiratory and remdesivir appeared far more often than their previous trajectories would have predicted. This is to be expected: medical researchers were writing about a new disease, its symptoms and its treatments. In 2020, for example, pandemic appeared in approximately 2 per cent of abstracts, around 21 times its expected frequency. 

But after the release of ChatGPT, the researchers found another abrupt shift, this time predominantly in generic stylistic vocabulary for which no "world change" (... except for the surge in GenAI use) offered an obvious explanation. In 2024, delves appeared 28 times more often than expected, underscores 13.8 times more often and showcasing 10.7 times more often. The scale of the shift was also larger: the researchers identified 190 excess words at the height of the Covid-related change in 2021 (almost entirely content words), compared with 454 in 2024 (overwhelmingly verbs and adjectives associated with LLM writing). On that basis, the authors estimated that at least 13.5 per cent of the abstracts published in 2024 had been processed using an LLM. 

The authors were careful about the limits of their method: Their analysis can show that writing practices changed across millions of biomedical abstracts. It cannot, however, determine how any particular abstract was produced. An author may use "delves" without having used AI, while an AI-assisted article may contain none of the words identified by the study. The method therefore provides evidence of LLM-assisted writing across the collection, but not a reliable test of authorship for an individual text. 

This article therefore does not give lawyers a forensic tool for unmasking AI-assisted drafts from opposing counsel. By making the accent easier to recognize, however, it allows them to choose how they dress their own expertise. 

The study also showed that individual words have a short shelf life as markers. Once delve became notorious, users began banning it and models used it less often. As in fashion, lawyers will have to keep updating their wardrobe. 

What gives legal writing the AI accent? 

Words that steal the reader’s time 

As shown in the above-cited study, the most obvious markers are the words appearing disproportionately often in AI-assisted text. Anyone who regularly reads AI-assisted prose will have their own additions: spearhead, robust, seamless, pivotal, comprehensive, foster, landscape. The author's personal blacklist includes quietly, genuinely and honestly. They are rarely wrong in a strict grammatical sense, but they occupy space without adding useful information: A “robust compliance framework” might consist of approval rights, reporting duties, an audit process... or a single miserable internal policy no one follows. The adjective allows the writer to avoid specifying which of these are actually in place. 

Similarly, when technologies quietly transform industries, or risks quietly emerge, the adverb does little except spare the author from explaining when the change occurred, who caused it or how its effects can be observed. 

A Concerto in AI Minor 

Perhaps “multimodal” models have taken the term too literally? They seem to love rhythm, at the expense of substance. Every technique taught by rhetoric professors to hold an audience’s attention has been ingested and is fired into the prose like a row of staples from a mechanical tacker. Lists come in threes. Ordinary statements are recast as: “It is not about X. It is about Y.” The model asks a question and immediately answers it. Short fragments create drama. 

Human legal analysis is the opposite of music to the ear. The law is boring but precise, and a list contains exactly as many points as the analysis requires. 

If your writing suggests that you favour form over substance, the reader may subconsciously hear the voice of a shallow sophist where they should recognize an experienced advocatus. 

Structure without substance: a house of cards 

Again, reproducing every lesson from Essay Writing 101, AI adds headings, signposts, summaries and subtitles. These may be useful in a long submission, but they look like parody in a one-page memorandum whose conclusion repeats the first paragraph.  

Just like you, your clients and opposing counsel will raise an eyebrow at the 50 “em dashes”, the pale-blue quote boxes, the bullets introduced by quirky bold titles and the indispensable "Why this matters" subheading. 

What legal writings require, however, is consistency in defined terms. Any lover of elegant prose may wince when reading that "Seller undertakes to inform Purchaser of any delay by Seller or Seller's Subsidiaries in the production of Purchaser Product", but letting AI smooth away the repetitions by introducing synonyms would be a beginner’s mistake. 

“Better to be roughly right than precisely wrong” pushed too far 

Models are trained to avoid overclaiming and therefore compensate by "hedging". They explain that an issue “may potentially” create risk, that the outcome “will ultimately depend on the circumstances” and that “a careful and comprehensive review is recommended”. They present both sides and avoid deciding between them. 

This substantive cowardice is difficult to reconcile with what law-firm clients are paying for: judgment under uncertainty. So deleting the typical “this does not constitute legal advice” disclaimer displayed at the end of an AI draft is not enough: a legal memorandum written in permanent caveat mode may also diminish the client’s enthusiasm for paying the bill. 

Toxic residues 

Saving the best for last, nothing beats the Schadenfreude of finding a so-called "chat residue" in an opposing party's draft: “Certainly! Here is the revised clause”, followed perhaps by the AI's kind offer to prepare a shorter version. More subtle, placeholders (“[Insert jurisdiction]”) can survive several rounds of review because the surrounding text looks finished. Beware when reviewing your own team’s drafts: Fluency encourages the eye to move on. 

Removing the accent 

In our experience, a good style prompt removes around the major part of the recurring patterns at source.  

Since the model already considers its default output natural, telling it to “write normally” achieves little. Style instructions therefore work better when they identify an observable pattern and prescribe a correction. Some useful examples: 

Include two or three samples of your own writing whose style you still like. They give the AI evidence of your sentence length, register and vocabulary. Choose them carefully. AI has a sense of humour of its own and is good at spotting, and reproducing, your quirks and bad habits. 

In conclusion, navigating the rapidly evolving landscape of AI-assisted legal writing requires a robust and holistic approach. By leveraging innovative tools, fostering authentic communication and embracing meaningful human oversight, legal professionals can unlock the full potential of AI while preserving the uniquely human qualities that lie at the heart of the profession. 

Did that paragraph make you doubt the rest of the article? 

Prof. Marcel Salathé has proposed a useful name for criticizing AI slop in prose that itself appears to have been left largely unedited: slopocrisy.(2) One paragraph of unedited AI English, and suddenly the author seems less competent.  

As creators of a powerful Legal AI that helps legal professionals work more efficiently every day, and which we use extensively ourselves, you can probably guess that this article is not meant to send you back to pen and paper. Nor is it meant to help you conceal your use of AI. By training our eyes to recognize the patterns and understanding the impression an unwanted accent creates, we can ensure that the tool’s contribution remains subordinate to our own analysis and that we remain the real maîtres of the result, in both senses of the French word. 

_____

1 Dmitry Kobak, Rita González-Márquez, Emőke-Ágnes Horvát and Jan Lause, “Delving into LLM-assisted writing in biomedical publications through excess vocabulary”, Science Advances, vol. 11, no. 27, 2025, eadt3813. 

2 Marcel Salathé, “I’d like to propose the term #slopocrisy for the growing genre of posts that deride AI while clearly being written by AI”, LinkedIn post, July 2026.