Opus 5.5's 2,548 AI Writing Tics: How to Recognize AI-Generated Text

Open an email, an article, or a homework assignment, and you sense something is off. The grammar is fine, the sentence structure is tidy, but you suspect it was written by AI. The hard part is explaining why.

The marketing agency Graphite conducted a quantitative study to try to answer that question. They compared articles generated by ten AI models against human-written articles word by word, cataloging the expressions in Claude Opus 5.5 — Anthropic's latest model — that appear far more frequently than in human writing.

What the study found

Opus 5.5 contains 2,548 writing tics, only 4% fewer than the previous generation, Opus 5, which had 2,666. A "tic" has a precise definition in this study: any word, phrase, or sentence pattern that appears at least twice as often in AI text as in human text.

The research team paired human and AI articles rigorously. On the human side were 10,000 website articles published before November 30, 2022 (the day ChatGPT launched), ensuring no chatbot ghostwriting. On the AI side, each human article was summarized, and only the summary was given to each model, which was asked to write a new article of the same length, avoiding direct copying of the original wording.

Three metrics did improve

The first is the em dash. Opus 5 used 2.92 em dashes per thousand words; Opus 5.5 dropped to 0.015, a 99% reduction — roughly one dash every 67,000 words.

The second is "affected prose," meaning the use of metaphor and embellishment in place of direct statement. On a scale of 0 to 100, human text scored 6.65, Opus 5 scored 16.75, and Opus 5.5 dropped to 10.57 — still 1.6 times the human rate.

The third is the overall distance between the model's vocabulary and human vocabulary, which fell 19% compared to Opus 5. This aligns with Anthropic's promise in its release announcement to use "fewer terms and idiosyncratic expressions."

In one sentence: Opus 5.5 writes more like a human than Opus 5, but it eliminated very few tics — mostly it replaced old ones with new ones.

Writing tics that expose AI text

Opus 5.5's tics fall into four categories: announcing that something matters, defining things through contrast, praising with vague adjectives, and stitching paragraphs together with transition phrases.

Announcing importance

This is the strongest tic. "This matters" appears 116 times more often than in human text, and "why _ matters" appears 92 times more often. The key difference: human writing shows the consequence of a fact, while AI stops to tell you the fact is relevant. For example, instead of "the battery lasts two days, which matters," say "the battery lasts two days, so you don't need to pack a charger for a weekend trip."

Ready-made contrast constructions

After the structure "It's not just X, it's Y" became widely recognized, Opus 5.5 abandoned it. In its place is a variation: "is more than a _, it's a _," which appears 98 times more often than in human text. What these constructions share is setting up a weak point in order to knock it down and appear profound, while actually conveying no new information.

Empty praise and transition phrases

The third category is adjectives that approve rather than describe: "dependable" appears 23 times more often than in human text, and "steady" 11 times. These words can be inserted into any review of any product.

The fourth category is bridges between paragraphs: "looking ahead" appears 40 times more often, "adds another layer" 27 times, and "what comes next" 24 times. Human authors usually move directly from one idea to the next without first announcing that a transition is about to happen.

Every model has its own accent

In the original study, Graphite counted 12,877 distinct tics across nine models, 65% of which appeared in only a single model family (Claude only, GPT only, or Gemini only). Learning one model's tic list won't help you recognize another model.

The differences between models are very specific. GPT-6 Astra tends toward hedges, using "may provide" 18 times more often than Opus 5.5. Opus 5.5 tends toward exaggeration, using "the most popular" 45 times more often than Astra. One is like a cautious lawyer, the other like an enthusiastic salesperson.

The two are also moving in opposite directions. Claude's vocabulary distance from human text fell from 0.064 in Opus 5 to 0.052 in Opus 5.5, while OpenAI's rose from 0.101 in GPT-5.6 Sol to 0.109 in GPT-6 Astra.

What's called a "model accent" refers to systematic bias in word choice and sentence structure within the same model family, similar to the stylistic fingerprint of different authors. Graphite quantifies this with "overall distance between model and human vocabulary": comparing the word-frequency distribution of model-generated text against that of human text as vectors, where a smaller distance means closer to human usage. This metric is relative; it only reflects distributional differences and does not indicate text quality. GPT-6 Astra's widening distance suggests it deviates further from human word habits in the new version, while the shrinking distance of the Claude series suggests its expressions hew closer to human corpora. A caveat: a small distance may also come from "imitating common expressions" and does not equal greater insight.

Why AI writes this way

Language models choose the most probable next word at each step. Expressions that appear frequently in the text they've read and that satisfied evaluators during pre-release assessments become habits. Models have billions of parameters, testing is limited, and labs have less control over tics than you might think.

The data supports this. The 11 most famous tics lost 53% of their strength from Opus 4 to Opus 5.5, while the total number of tics fell only 32%, from 3,746 to 2,548. Famous tics fade quickly; the overall set changes slowly.

The fate of the em dash illustrates this process. After the signal was widely mocked online, all three labs adjusted their usage. Opus 5.5 and Gemini 3.1 Pro nearly eliminated it, and Astra's usage is even lower than human. The overcorrection means that today, a long article with no em dashes at all deviates from human patterns.

A language model's training objective is to predict the next word given context, and expressions that appear frequently in the training data get reinforced repeatedly, becoming the model's default "smooth option." At the same time, models undergo reinforcement learning from human feedback (RLHF) before release, and evaluators tend to prefer responses that are polite, clearly structured, and seemingly professional, further amplifying tics like "announcing importance" and "using vague praise." Because models have billions of parameters and training and evaluation cover limited samples, labs struggle to control every specific expression. The result: certain "accents" aren't deliberately planted by designers but are byproducts of the training objective and data distribution. This also explains why new versions often replace tics rather than eliminate them entirely.

Why AI detectors are unreliable

A detector is essentially another AI program that gives a probability, not evidence. In 2023, research led by Debora Weber-Wulff tested 14 detection tools (including Turnitin), and the conclusion was direct: they are neither accurate nor reliable, tending to misclassify AI text as human-written.

The reverse error is worse. Research by Weixin Liang and colleagues showed that detectors flag text written by non-native English speakers as AI-generated. For students, this can lead to accusations of academic misconduct.

Watermarking is the theoretical solution: the model slightly prefers a secret set of words during generation, imperceptible to ordinary readers but detectable by an informed program. But this only works if model companies enable it voluntarily, and there is no mechanism to force anyone intent on deception to use a watermarked model.

AI detectors typically judge whether text is AI-generated based on statistical features or another classification model, such as perplexity and burstiness. But these metrics have no evidentiary weight in legal or academic terms, because human writing is extremely diverse and model outputs are constantly changing. Tests by Weber-Wulff and others showed that detector accuracy in real scenarios is far below marketing claims, and false positives cluster among non-native English speakers, those with formal language styles, or those using translation tools. Watermarking requires the model provider to actively embed a detectable signal during generation; once a user employs a non-watermarked model, rewrites the text, or copies it across models, the watermark fails. Detectors are therefore better used as a supplementary reference, not as the sole basis for determining academic misconduct or cheating.

Recognizing AI text without memorizing word lists

The most reliable approach is to focus on what these tics share: sentences that fit any topic, announcing importance without giving consequences, praising without describing, and lacking details only a firsthand witness would know. These signals depend on the ideas in a sentence rather than specific words, so they work more easily across languages and versions.

The "signs of AI writing" guide maintained by Wikipedia volunteer editors is currently the best public reference. TechCrunch called it the best guide in the field.

Next time you suspect a passage, ask five questions:

1. Could this sentence be used in another article? "An important milestone for the industry" applies to any news story. Human sentences are usually tightly bound to their subject.
2. Is the text saying something matters, or showing the effect? Look for specific consequences. If you can't find any, it's empty emphasis.
3. Who said it? "Experts point out" and "research shows," with no names or links, are the vague attributions the Wikipedia guide lists as signals.
4. Are there names, dates, places, numbers? Text written by someone who did reporting contains verifiable details. Generic text only talks about "researchers" and "companies."
5. Does the rhythm vary? Uniform paragraph lengths and every paragraph ending on a punchy line are marks of formulaic writing.

Reverse signals are worth noting too. Graphite's study found that the most human markers are informality: first person, loose phrasing, and exclamation points. Exclamation points appear more than 100 times more often in human articles than in AI ones.

But note that the Wikipedia guide itself warns these patterns are only possible problem signals, and many people write this way without using AI. A single isolated signal proves nothing; it's when multiple signals appear together that it's worth asking questions.

What it means for people who write with AI

The real risk isn't getting "caught" by a detector; it's delivering generic text and letting the model's quirks replace your own voice.

If you write with AI, when you reread, delete every announcement of importance and every non-descriptive adjective, and add details only you know. If you're a teacher or manager evaluating someone else's text, don't use detectors as evidence — instead have a conversation about the content: the person who wrote it should be able to explain what they wrote.

Switching models won't solve the problem, because every model has its own accent. What to remember: look less at em dashes and more at sentences that promise importance but never deliver.