Does AI Use Contractions? What This AI Writing Pattern Really Tells You

AI uses contractions on any task that invites them, and skips them entirely on formal ones. Ask ChatGPT for a chatty email and the contractions show up. Ask the same model for an academic passage and you won't see a single one.

So the habit points to the task you set. We gave three chatbots the same three writing jobs and counted every contraction, and you'll find the numbers further down. Before those, here's what the 2026 research measured and the seven patterns worth reading together.

🤖 Does AI Use Contractions? Yes, but Not Consistently

Yes, AI-generated text uses contractions. Two 2026 studies measured the same feature and reached opposite conclusions, one by Utsav Paneru and one by Karolina Rudnicka and Thomas Juzek, because each tested different models on different kinds of writing.

What counts as a contraction?

A contraction is a shortened form of two words, joined by an apostrophe. Negation contractions make up one group, where do not becomes don't. Others shorten a verb, so it is or it has becomes it's, we are becomes we're, and I have becomes I've.

An apostrophe on its own doesn't make a contraction. In "the writer's laptop," it shows possession instead. Counts have to separate the two, or the totals come out wrong.

Why the old "AI avoids contractions" rule is incomplete

Paneru's preprint shows where the rule came from. He built a parallel corpus of 25,140 paired chunks of AI and human text. In its 1,390-item test subset, Paneru's dataset recorded 0.00 contractions per AI chunk against 0.17 for the human references.

Rudnicka and Juzek read the same feature across model families and found the opposite. In their 2026 model group, rates ran from over 1,200 to more than 30,000 per million words, with Claude Haiku 4.5 recording 30,611.9. GPT-3.5, tested in the older 2024 group, recorded 120.2.

⚖️ What Contractions Can and Cannot Tell You About AI Writing

Contraction habits tell you what kind of writing you are looking at. A high count suggests a conversational brief and a count of zero suggests a formal one, with the author out of reach either way.

Formal human writing drops contractions on purpose, and journals, legal memos and graded essays all reward the expanded forms. An editor can also add or strip them in one pass, so the count may belong to the editor rather than the writer.

A shift partway through a document is worth a closer read.

👇 Try the same passage in the checker before you read on.

🔀 Why Contraction Use Changes Across AI Models

Four things move the rate: the model, the prompt, the genre, and any editing afterward.

Model and version

A Carnegie Mellon study from April 2025 identified which of five large language models wrote a passage with 97% accuracy, using word choice alone. That work compared the five models against each other and never tested AI against human writing. It does explain why lists of common ChatGPT words tend to fail on Claude or Gemini.

Prompt and requested tone

"Write a formal academic explanation" and "explain this to a friend" set two different registers, so they should never return the same apostrophe count.

Does a formal prompt remove contractions?

Every time, in our test. A formal instruction pushes a model toward the expanded forms, like "they are" and "would not".

Genre and audience

Papers and legal memos avoided contractions long before any AI was involved, while emails and blog posts have always used them freely.

Edits and collaboration

A person may rewrite AI output, and AI may edit human writing, and the finished page looks the same either way.

What if a person edited the AI draft?

Then the visible style belongs partly to that editor. Word choice changes first, so the deeper grammar survives the edit.

The Phrasly cross-model contraction test

We gave the same three tasks to three chatbots and counted the contractions in every answer.

Every formal prompt returned zero contractions, whatever the model, and the neutral prompt returned almost none. The three models only separated once the prompt asked for a conversational tone, which puts the task ahead of the model.

🔍 7 AI Writing Patterns to Check Alongside Contractions

Anyone working out how to spot AI writing meets the same problem. Each of these seven patterns appears in human writing too, so read every one as a question, and you recognize AI writing by how many turn up together in the same passage.

1. Repeated stock phrases. Words like delve, tapestry, and underscore show up often in model output. Phrasly's report on common AI phrases lists the most overused ones.

2. Formulaic contrast and parallelism. Mirrored sentence frames and repeated groups of three give prose a template feel. Watch for the same shape reused across several paragraphs.

3. Uniform rhythm. The sentences run to a similar length, or the paragraphs are all cut to the same size, where human drafts usually run uneven. One tidy paragraph on its own still means nothing.

4. Noun-heavy, dense prose. Verbs get converted into nouns and clauses stack into long noun phrases, which is the shape the PNAS 2025 study measured in GPT-4o. The result reads flat on the page.

5. Generic detail. The writing offers smooth generalities where the task called for specific observation. Models can produce specific detail on request, and they can invent it. The test is whether the detail checks out.

6. Over-organized transitions and endings. The same connectors come back on repeat, the sections run to an even size, and the conclusion only restates the article. Common AI words tend to surface in those joins first.

7. Sudden changes in voice. The register jumps from casual to formal and contraction use often jumps with it, though co-authoring, quoted material and ordinary revision all produce the same effect. Mark where the voice changes and read both sides.

That list covers the common AI text patterns worth checking in a short passage. To identify AI writing across a longer document, read our guide on how to tell if text was written by AI.

🧭 How to Check a Passage Without Relying on One "Tell"

Five steps worked in order are how to spot AI-generated text without leaning on a single habit. Any one of them can mislead you alone.

1. Use enough text. Read a complete passage, because one sentence rarely holds enough evidence to judge. Phrasly's detector accepts a 30-word minimum, and longer coherent samples give any tool more to work with.

2. Look for a cluster. Ask whether the same patterns repeat: formulaic phrasing, flat rhythm, generic claims, odd citations, voice shifts. Stylometry, the measurement of writing style, works on groups of features like these rather than on any single one.

3. Run a detector second. Paste the full passage into the Phrasly AI Detector, read the overall estimate, then open the sentence-level highlights. Our guide to how AI detectors work explains what the score measures, and the walkthrough on how to scan text for AI covers the steps.

4. Check provenance. Look at sources, quotations, earlier drafts, version history, and the writer's previous work.

5. Keep the response proportionate. A score starts a review or a conversation. Turnitin's own documentation says an AI writing score "should not be used as the sole basis for adverse actions against a student." Turnitin also withholds scores in the 1% to 19% range, because false positives happen.

✍️ If Your Own Writing Looks "AI-Like," Do This

Revise the writing itself and leave the score alone. Random contractions, slang, typos or invented anecdotes will not settle the question of who wrote it.

• ✅ Revise for clarity and specifics. Name the source, the number, the case you actually saw.

• ✅ Verify every quotation before the draft goes out.

• ✅ Keep your drafts and version history. They answer the question faster than any argument.

• ✅ Follow the AI-use policy your school, employer, publisher, or client has published.

• ✅ Ask for human review if you're flagged, and walk through how you wrote it.

Non-native English writers and heavily edited drafts draw more false flags. Our guide on AI detector false positives covers what to do when one lands on honest work.

❓ Frequently Asked Questions