One em dash, a polished sentence, or "delve" on its own is weak evidence. The useful question is whether several AI writing patterns cluster together, depart from the writer's usual voice, and point to something you can check.
What are the signs of AI writing?
The signs range from easy-to-spot word choices to repeated sentence shapes and generic structure. Treat these AI writing tells as review prompts with different strengths, not as an authorship test. Wikipedia's WikiProject AI Cleanup page lists candidate patterns and points out that similar elements appear in ordinary prose.
| Tell | Level | How much weight to give it | What to check |
|---|---|---|---|
| A dated cluster of words such as delve, pivotal, tapestry, or underscore | Word | Mild by itself; more useful when several appear in a short passage | Is the word doing precise work here, or decorating a general claim? Wikipedia notes these favorites change over time. |
| Repeated openers, transitions, or sentence lengths | Sentence | A useful reason to compare with the writer's other work | Do several sentences start the same way or land with the same rhythm? |
| A tidy three-part list, "not X but Y" turns, or an ending that restates the opening | Structure | Stronger as a cluster across the whole piece, still not proof | Does the structure fit the actual argument, or could the same frame fit any topic? |
| Vague expertise, unsupported numbers, or citations that do not resolve | Evidence | The most actionable clue: verify the claim or source directly | Does the source exist, and does it support the exact sentence? |
| An em dash, perfect grammar, or a list | Surface | Weak alone; these are ordinary writing choices | Look for repeated, unusually uniform use alongside other clues. |
Vocabulary studies support the narrow claim that certain words rose in scientific writing after chatbots became common. One analysis of more than 15 million biomedical abstracts estimated that at least 13.5% of 2024 abstracts had been processed with an LLM; that is a field-specific estimate, not a way to label one paragraph. Kobak et al.'s study looks at biomedical abstracts, while Juzek and Ward's study identifies 21 focal words in scientific abstracts. Neither makes a single word a fingerprint.
How do you spot AI writing without overreading one word?
Use a four-pass review. Keep a note of what you can point to instead of guessing at the writer's process.
- Compare with a baseline. Read another recent piece by the same writer, if you have one. A sudden change in vocabulary, rhythm, or level of detail is more useful than a general "AI feel."
- Mark a cluster. Circle repeated transitions, abstract praise, identical sentence openings, and conclusions that summarize what was just said. Several different kinds of pattern matter more than one repeated word.
- Check the claims. Open the links. Verify names, dates, figures, and quotations. A fabricated or mismatched citation is a concrete problem whether or not AI was involved.
- Ask what the structure is doing. Underline each paragraph's job in five words. If every section follows the same "problem, three points, summary" frame, ask whether the subject actually needs that frame.
Wikipedia's list of AI writing signs changes as model usage changes. It says "delve" dropped off sharply in 2025, and distinguishes newer patterns from earlier ones. Check its dated word groups before treating any internet list of "AI buzzwords" as current.
The page's caveat also warns against treating these signs as the problem to fix: scrubbing them can just make detection harder while the deeper concerns stay.
How reliable are AI writing detectors?
A detector score is a lead to review, not proof. Its result depends on the tool, text length, language, model generation, and whether the text has been edited.
OpenAI retired its own classifier in July 2023. In its English challenge set, it correctly marked 26% of AI-written text and falsely marked 9% of human-written text; OpenAI also warned that short text was especially unreliable. Those numbers describe that classifier's test, not every current product. OpenAI's notice says it should not be used as the primary decision tool.
False positives can matter to real writers. A 2023 Patterns study ran seven detectors over two essay sets: 91 TOEFL essays collected from a Chinese forum, and 88 essays by US eighth-graders. On the TOEFL samples, the average false-positive rate was 61.3%; the result belongs to that study's specific tools and sample. Liang et al. also found that simple prompting could make generated writing harder for those detectors to catch.
A person's fluency or first language is not evidence of AI use.
If a detector flags a draft, check provenance if available, verify the factual claims and citations, and ask the writer about the work. Do not turn a score into an accusation. For your own draft, use the result to find a sentence worth editing, not to chase a "human" score.
Which signs of AI writing survive editing?
A word or punctuation tell can be changed in one edit; checking a repeated structure takes comparison across drafts. To see how output shape can vary, compare repeated runs of the same example prompt — not as a test of who wrote them, but as a small illustration.
In our showcase, the plain Claude run for an example candle post starts: "Poured 40 candles yesterday and 12 of them came out with cracked tops." A protocol run for the same example input starts: "12 of yesterday's 40 candles came out with cracked tops." The subject and facts stay put; the opening changes. These are example inputs written for the showcase, not customer posts. The showcase page has more side-by-side outputs.
I measured the six example inputs in threads_posts and x_posts, separately for Claude and Codex. Each input has three plain runs and three outputs for three protocol plays. I counted a repeated opening only when all three outputs for one input shared the same first three words. "Same ending shape" means the three last sentences still matched after lowercasing, preserving punctuation and a fixed set of function words, and replacing other words with X.
| Model | Same first three words in all plain runs | Same first three words in all protocol plays | Identical ending shape in either set |
|---|---|---|---|
| Claude | 4 of 6 input groups | 0 of 6 | 0 of 6 plain; 0 of 6 plays |
| Codex | 3 of 6 input groups | 0 of 6 | 0 of 6 plain; 0 of 6 plays |
No output in this sample used an em dash or the fixed Wikipedia-derived buzzword list. That does not show those tells were edited away: each arm shared rules that already banned em dashes and several stock phrases. The sample is only six example inputs per model; it does not establish a general rate or prove that any text was AI-written. It does show why a word checklist can miss a structural question: do the same task and facts keep getting arranged the same way?
We make Unsame. Its launch post explains the playbook protocol behind the example; you can review the samples without using the product. If you want to make AI posts sound less like AI, start with the claim, detail, and point of view you actually want to keep.
How should you review your own AI-assisted draft?
Keep a short, evidence-based edit list: one repeated pattern, one claim to verify, and one specific detail that belongs to you. Then read the revised text aloud and decide whether it still says what you mean.
- Replace vague praise with a fact, example, or delete it.
- Change a repeated opener only if it sounds wrong in your voice.
- Check every quote, number, and source link.
- Keep punctuation you normally use. An em dash is punctuation, not proof.
- Do not add fake typos or awkward phrasing to satisfy a detector.
A good review leaves a clearer draft and a small list of checked claims. It does not leave a verdict about authorship.
— Don
