Spotting AI-isms.

What this guide covers

An AI-ism is a writing habit that shows up again and again in generated text. A quality problem is a real loss of usefulness, accuracy, logic or clarity. The two often travel together, but a habit alone does not make a passage bad. Neither of them tells you who wrote the text.

Use this guide to find a specific pattern, explain what it does to the reader and suggest a small edit. It is an optional reference. It cannot detect authorship, and it gives no universal score.

I made up all the examples below for teaching. Their numbers and product descriptions say nothing about real products.

A five-step review

• Read the task. Note the audience, the purpose, the requested format and the source material. A slogan can suit an advert and still fail a technical explanation.
• Read the whole passage. Understand what the writer meant before you mark single words or symbols.
• Mark the exact span. Name the pattern you see, using the list in this guide.
• Check the effect. Add a quality label only when the wording causes a problem you can name, and explain that problem in context.
• Verify and report. Check facts against the supplied evidence when the task asks for it. Suggest the smallest edit that keeps the meaning, any needed uncertainty and the attribution.

Never turn a count of patterns into an "AI probability" or a quality score. Several labels can describe one span, and they still describe one problem.

Five quality labels

These labels describe the effect of the wording. Keep them separate from the pattern that caused it.

• Clunky. The wording sounds wrong for the audience, the genre or the meaning. Ask whether the intended speaker would really say it this way.
• Ornate. Decoration or length gets in the reader's way. Ask whether a simpler version would keep the useful meaning and the voice.
• Hollow. The claim stays unclear, or general phrases stand where specific information belongs. Ask what exactly the reader learns.
• Filler. The material adds nothing the task needs. Ask whether the reader would lose anything relevant if it disappeared.
• Broken logic. The parts don't fit together, or the conclusion doesn't follow. Ask whether each step leads to the next.

Telling them apart

• Clunky. “Please execute the clicking of the download control.” The action is clear, but nobody talks like that. “Click Download” works better. Reading aloud helps here, though academic and technical prose does not need to sound like speech.
• Ornate. “The elegant button ushers you into a seamless universe of effortless file access.” The decoration buries a basic instruction.
• Hollow. “The platform offers meaningful improvements.” Which ones? If the source supports it, name the feature, for example “The platform now exports PDF files.”
• Filler. A paragraph about the history of downloads before the answer to “Where is the download button?” Remove it unless the task asks for background.
• Broken logic. “The answer is longer, therefore it is more accurate.” Length says nothing about accuracy, and fluent grammar does not repair the reasoning.

Choose the label or labels that say the most. A sentence that merely feels artificial does not need all five.

Seven patterns to check

Each pattern below is something to inspect. Finding one does not prove a quality problem.

Separator overload

Several separators such as dashes, semicolons, dots, bars or arrows compete inside ordinary prose. Ask whether the punctuation organises the information or chops it into slogan-like fragments.

For example, “The workflow — faster; clearer | ready to go → now.” A better version names the actual change, such as “The workflow now includes a preview step”, and only if that step really exists. Never invent a benefit while you edit.

Fine to keep, a semicolon that joins two related clauses, a meaningful dash, a decimal point or arrows inside a process diagram.

Reframe flip

The text rejects one description to announce another, as in “not X, it's Y” or “not just X, but Y”.

For example, “It is not a document. It is a revolution in understanding.” Ask whether the second half clarifies anything or only turns up the volume, and whether rejecting the first half makes sense. A plainer version says what the document does, for example “The document explains tokenisation with worked examples.”

Fine to keep, a contrast that corrects a real misunderstanding, such as “The model predicts tokens, which are often parts of words.”

Missing actor

The sentence hides who did something, as in “The file was deleted.”

For example, “The incorrect settings were selected, and the files were deleted.” Ask whether the reader needs to know who acted, and whether the wording hides responsibility or makes an instruction harder to follow. If the source names the actors and the order, a clearer version says “The script deleted the files after the user selected the incorrect settings.”

Fine to keep, procedure descriptions where the actor matters less, such as “The samples were stored at −20 °C.” This construction is no grammatical error and no reliable AI marker. One study of GPT-4o even found fewer agentless passives than in human writing.

Command heading

A heading gives an order instead of naming the content, such as “Unlock Your Potential”.

For example, “Embrace the Future of Understanding”. Ask whether the heading tells the reader what the section contains or what useful action to take. A descriptive heading such as “How tokenisation works” usually serves better.

Fine to keep, real instructions such as “Download the sample file” or “Check the source”.

Tagline talk

Catchy formulas, buzzwords or balanced fragments carry little concrete meaning.

For example, “Smarter learning. Stronger connections. Limitless possibilities.” Ask which feature, action or supported result the line communicates. A useful version says something the reader can do, for example “Rotate the model and select a block to read its explanation.”

Fine to keep, a short campaign line when the brief asks for one. Even then, audience and purpose decide.

Cushioned figures

The text calls a number, chart or estimate approximate, illustrative or indicative. Ask whether evidence supports the qualifier, whether the text needs it and whether the reader can tell a measured number from an estimate, a hypothetical value or an invented one.

For example, “The system improves accuracy by approximately 47%.” Nothing shows where the figure comes from. A fair note says “The draft gives no source for the improvement claim”, and the edit removes the percentage until evidence exists. The word “approximately” does not rescue an invented result.

Fine to keep, honest disclosures such as “This chart uses made-up values to show how softmax changes a distribution.” Keep real uncertainty, measurement limits and simulation labels.

Symbol clutter

This narrower label covers only these symbols, : | ; · • — ← ↑ → ↓. Full stops, commas, hyphens, brackets, slashes and emoji fall outside it. Separator overload is the broad family, and symbol clutter is the strict version. Count them once when they describe the same span.

Look for chains of separators, arrows inside prose, decorative bullets in the middle of a sentence or punctuation that hides how ideas relate. For example, “Learn → explore → improve | clarity • confidence — results.” A clearer line would be “Explore the model to see how a token moves through its stages.”

Fine to keep, Text → tokens → embeddings inside a labelled diagram, bullets in a real list or a colon that introduces an explanation. Mark the effect, never every symbol by reflex.

More patterns from research

These patterns extend the list above. The evidence behind them comes from specific models, languages and kinds of text.

• Pet words. Repeated favourites such as “delve”, “intricate” and “underscore” where a simpler precise word fits. Researchers measured this in particular English corpora, so treat it as a tendency and never as a banned-word list.
• Noun pile-up. Actions buried inside nouns, as in “the implementation of the optimisation process”. Researchers measured this preference in particular GPT-4o and Llama 3 variants.
• Inflated meaning. An extra clause claims that an ordinary action proves commitment, importance or progress. Check whether that link really exists.
• Mirror rhythm. Triads, equal-length fragments or identical paragraph shapes that repeat. Useful parallel structure is still fine.
• Throat clearing. A passage announces an explanation, restates it and then summarises it without moving the answer forward. Judge what each part adds to the task.
• One-size voice. The same formal, promotional tone appears in a personal note, a tutorial and a critical review. Genre fit matters.

Read words, structures and punctuation together with the task. Change a recurring technical term only when clarity improves, because forced synonyms can make a technical explanation less precise.

Content comes first

A fluent sentence can still be wrong, and a plain one can still be useless. Before you polish a small stylistic habit, check the problems the assignment cares about.

• Unsupported claims. Numbers, quotes or citations that you cannot trace to the supplied evidence. A working link does not prove that the page supports the claim. Studies of early ChatGPT versions documented invented and incorrect citations, and those older error rates say nothing about current models.
• Faulty reasoning. A conclusion that doesn't follow from the facts, or two parts that contradict each other.
• Irrelevance. The answer discusses the topic but misses the actual question or action.
• Sycophancy. Agreement or praise takes the place of a judgment based on evidence. Friendly wording alone is fine.
• Leftover scaffolding. Placeholders, notes to the writer or lines such as “Here is the revised version” left in a finished text.

Use the project's own content labels when it has them. Factual and instruction-following errors belong in those categories, so keep them out of the style labels.

A worked review

The task asks for an explanation of tokenisation for a beginner, in two direct sentences. The candidate answer reads like this.

• “Unlock a new era of understanding” is a command heading and tagline talk. It explains nothing about the operation, so it is hollow and filler.
• “isn't just splitting text — it's the key to limitless intelligence” is a reframe flip with a dash. The real problem is the unsupported exaggeration, which makes it hollow and ornate. The dash matters less.
• “facilitating the transformation of linguistic inputs into actionable representations” is a noun pile-up that weighs down a beginner explanation, so it is clunky and ornate.
• “a seamless learning journey is enabled” hides the actor, sounds like an advert and has no clear link to the task, so it is hollow and filler.

Report the spans and their effects under the project's rules. Never sum them up as “four AI signals, so a machine wrote it”.

Useful feedback

When the project allows free-text comments, a short structure keeps them checkable.

• Task and audience. What the text had to do and for whom.
• Exact span. The words you mean, quoted.
• Pattern. Which habit from this guide you see.
• Quality label. Only if the wording causes a real problem.
• Effect. What the wording does to the reader or the task.
• Evidence. What you checked and what stays uncertain.
• Edit. A suggested fix, only if the task asks for edits.

A weak comment says “Sounds like AI. Too many dashes.” A useful one says “In ‘it is the key to limitless intelligence’, the contrast ends in an unsupported promotional claim. For a beginner explanation, describe what tokenisation does instead. The dash itself is no error.”

A practice question

I wrote these questions for practice, and none come from a platform assessment. Think through your answer before you open the suggested approach.

Final check

Your browser stores your progress, and nothing reaches this site.

Research behind this guide

The list of patterns, the examples and the review steps are this site's own editorial summary. The labels are working terms and no validated classifier, and judging slop always involves some opinion. Most evidence comes from English and from specific models, genres and datasets. Another language needs its own local examples.

• Kobak and colleagues, 2025, Delving into LLM-assisted writing in biomedical publications through excess vocabulary ↗. Word-use changes across English biomedical abstracts, with no verdict on single authors.
• Reinhart and colleagues, 2025, Do LLMs write like humans? Variation in grammatical and rhetorical styles ↗. Grammatical patterns of specific GPT-4o and Llama 3 variants across genres.
• Shaib and colleagues, 2025, Measuring AI “Slop” in Text ↗. A research framework for usefulness, information quality and style. The five labels in this guide come from FacinateAI Hub and do not copy that paper's categories.
• Sharma and colleagues, 2023, revised 2025, Towards Understanding Sycophancy in Language Models ↗. Assistants that agree with users at the cost of the truth, in the tested settings.
• Walters and Wilder, 2023, Fabrication and errors in the bibliographic citations generated by ChatGPT ↗. Citation errors in tested GPT-3.5 and GPT-4 outputs. The old percentages do not describe current models.
• Liang and colleagues, 2023, GPT detectors are biased against non-native English writers ↗. Bias and weak spots in the detectors tested at that time.