A polished paragraph or a strangely uniform essay can make you wonder who wrote it, but no single clue can settle the question. If you are learning how to tell if something is written by AI, start with care, not suspicion. A student, teacher, editor, or reader may have good reason to check a piece. Still, style alone cannot establish authorship.
A false positive is a detector result that labels human-written text as AI-generated. The strongest assessment combines close reading, fact checks, comparison with prior samples, and careful use of tools. The goal is to find a reason for a closer review, not to make a fast accusation. A few checks can help you separate a real concern from an unusual choice of words.
Key Takeaways
- A writing style can raise a question, but it cannot prove who wrote a piece.
- Check claims and citations against real evidence.
- Compare the work with prior samples written under similar conditions.
- Treat a detector score as one clue, not a verdict.
- Ask for context before drawing a conclusion.
Treat every clue as a reason to check more carefully, not as proof of who wrote the piece.
How to Tell If Something Is Written by AI From the Text
Repeated stock phrases, unusually even sentences, and formatting that does not fit the situation are useful clues to inspect, not a verdict. Look across the full text, since one odd sentence may reflect editing or personal habit. Patterns matter more than any single word or mark.
A recognizable phrase or punctuation choice alone cannot establish who wrote a piece. People use familiar language, too, and writers often edit their work to sound polished. Text generators such as ChatGPT, Claude, and Gemini can produce different styles, so no one pattern works as a dead giveaway.
Notice Repeated Language and Stock Phrases
Buzzwords such as “delve,” “leverage,” and “unlock” may stand out when they appear again and again. A generic opener like “In today’s fast-paced digital landscape” can also sound detached from the topic. These phrases may point to generated text, but people use them as well.
Notice whether a phrase adds meaning or simply fills space. A writer may repeat a favored term, especially in a short piece. Repetition across several paragraphs deserves a closer read, while a single familiar phrase tells you little.
Em dashes can make a sentence feel polished or dramatic. Some writers use them often, and others avoid them. Frequent em dashes may catch your eye, but punctuation is a weak clue without support from the rest of the writing.
Look at Sentence Rhythm and Formatting
AI writing can have an overly even sentence length, with each line carrying a similar beat. The wording may sound polished yet generic, as if it could fit many topics. That same smooth rhythm can also come from careful human editing.
Formatting should suit the task. Bullet points can feel out of place in a personal reflection or a short letter, though they make sense in a guide. Ask whether the layout helps the reader or seems added by habit.
- Repeated stock phrases
- Uniform sentence length
- Excessive em dashes
- Unexpected bullet points
A style clue can guide your next check, but it cannot name the writer.
Compare the suspected pattern with the purpose and format of the piece. Check whether it appears throughout the piece or only in one sentence.
What Deeper Patterns Can Reveal Generic AI Writing?
A passage can be grammatically smooth and still feel empty because it never explains what the writer thinks or why the details matter. A clear voice gives the reader a reason to care about the point. Fluency is not the same as thought.
Look for meaning and context, not just polished language. A writer’s style may sound formal or plain, yet still show a choice, a motive, or a useful example. A survey of 44 studies found tendencies in AI and human writing, but it also noted limits in the research, including its focus on English and GPT models.
Look for a Missing Point of View
A passage may make sense at first glance but reveal no identifiable stance. It may state that a topic matters without explaining why the writer chose it. Look for a human voice that shows what the writer values or wants the reader to understand.
Lived detail can make a point specific. “The bus was late again, so I missed the first ten minutes of class” tells you more than “transportation can be challenging.” The first line names an event and its effect. It does not prove a person wrote it, but it gives the argument a clear setting.
Check Whether the Logic Has Real Context
Vague words such as “challenges” or “distractions” can gesture at a problem without explaining it. Ask which challenge, who faces it, and what happens next. A broad statement needs details before it can support a useful conclusion.
A metaphor may sound vivid but leave the point unclear. “The idea was a lantern in a storm” paints a picture, yet it does not explain what the idea changed. A strong comparison helps the reader understand the claim, not just picture it.
Check whether each reason supports the next one. If a paragraph jumps from a problem to a sweeping answer, it may lack the link that makes its logic convincing. The gap is worth reviewing, whoever wrote the passage.
Notice Repetition and Forced Triplets
Formulaic transitions and repeated sentence structures can make writing feel built from a template. The rule of three means grouping ideas in sets of three. A three-part list or polished transition alone is not evidence of AI authorship.
Human writers also repeat patterns, especially in speeches, essays, and formal reports. What matters is whether the repeated form helps the argument or simply makes the prose sound complete. A list of three can be useful when each item adds a distinct reason.
A scholarly survey reviews patterns across research, including claims about formal tone and repeated wording. It also describes gaps in that work, so broad tendencies should not be treated as a test for one passage. Ask what a sentence contributes to the argument after you remove its polished phrasing.
How Can You Check the Claims and Citations?
A concrete factual error is more useful to investigate than a general feeling that a paragraph sounds artificial. Check what the writing says against direct evidence, such as a cited report or data set. Evidence can reveal a weak claim without revealing who wrote it.
A wrong fact or strange reference is a reason to investigate, not proof that AI was used. Current-event claims may be stale or wrong. Some AI systems can use search or uploaded documents, so do not assume every system has the same access.
Verify Specific Claims and Dates
Start with names, dates, quotations, and numerical claims. Compare each one with the source that the text names. A date can look plausible and still refer to the wrong event, while a quote may leave out words that change its meaning.
Figures deserve a separate check. Recalculate a percentage from the numbers given, and confirm that the text uses the right units. If a study reports a group of 80 people, for example, a claim about 800 participants changes the meaning.
Check whether current information has changed since the passage was written. A once-correct policy, price, or event detail may no longer be current. That can explain an error without answering how the text was produced.
Follow Each Citation to Its Source
A citation needs two checks: confirm that the source exists, then see whether it supports the claim. A real paper may be cited for a point it never makes. A made-up-looking reference may also result from a typo or poor citation work.
Use this short check for each factual passage:
- Confirm that cited sources exist.
- Compare each source with the claim it should support.
- Verify names and dates against the source.
- Recalculate any figures from the stated numbers.
Do not stop at a search result or a reference list entry. Open the source and locate the relevant text or data. That is where you can see whether the support matches the claim.
A source may discuss a related topic without backing the exact sentence. Follow the citation all the way to the passage or data that supports the specific claim.
How Does Comparing Prior Samples Help?
A sudden shift from a student’s usual voice is worth checking against earlier work, especially when the assignment and conditions are comparable. A prior sample gives you a point of reference, not a perfect record of someone’s ability. Context matters.
Compare at least two earlier samples before treating a change in style as meaningful. A new topic, extra editing, or more time can change a piece. A polished assignment may reflect revision, help from an editor, or a change in circumstances.
Compare Voice, Vocabulary, and Sentence Habits
Read the new piece beside earlier work written for a similar task. Compare tone, sentence construction, word choice, complexity, and recurring errors. Look for a cluster of changes rather than one unfamiliar phrase.
A short response may differ from a research paper because the form calls for different choices. A student may also learn new terms or try a more formal style. The work should be judged with its topic, audience, and instructions in mind.
Consider how much help the writer had. A teacher’s feedback, editing support, or more time can make an assignment clearer. Those changes can explain a shift without proving how every sentence began.
Review Drafts and Version History
Drafts, notes, and saved files can show how a piece developed. Google Docs Version history records document edits over time, so it may show added paragraphs or revisions. That record supports a picture of the process, but it does not prove authorship on its own.
Look for a writing process that fits the assignment. A writer might start with notes, add a claim, then revise after checking a source. A draft can also show edits made in one session, which does not by itself make the work suspicious.
One unusually polished piece can reflect revision, an editor’s help, or a change in the writer’s circumstances. That exception matters when you compare samples. Compare like with like by choosing earlier pieces written for a similar task and under similar conditions.
Can Detection Tools Prove Who Wrote It?
A detector can estimate whether text resembles patterns in its training data, but it cannot verify who wrote a passage. Treat the result as a signal that may guide a closer review. No AI detector can conclusively determine whether AI was used.
A detector score is a tool’s estimate, not a fact about authorship. A false positive labels human writing as AI-generated, while a false negative misses AI-generated text. Grammarly also says a detector cannot conclusively establish AI use and should not be the sole basis for a decision.
What Detector Scores Can and Cannot Show
Different detectors can give different results for the same passage. Their scores depend on the detector, its version, the writing type, and test conditions. A high result does not tell you the writer’s intent or show how the text was made.
A guide to AI content detectors explains how scores estimate text patterns and why human review still matters. A score can help you decide which passage deserves attention. It cannot replace the surrounding evidence.
Compare the September 2026 Results
A September 2026 study tested 1,000 essays with 15 detectors. Each essay went through each detector three times, creating 45,000 data points. The human set included 500 pre-2022 academic essays with verified authorship.
The AI set contained 300 ChatGPT-4o essays, 150 Claude Sonnet 4.6 essays, and 50 Gemini 2.5 Pro essays. Those samples give useful test conditions, but they do not represent every kind of text or every detector version. GPTZero v3.4’s 8.0% false-positive rate applies to this study’s human essays.
| Detector or model | Overall accuracy | False-positive rate | Model-specific detection |
|---|---|---|---|
| GPTZero v3.4 | 91.2% | 8.0% | 90.4% AI; 92.6% ChatGPT-4o; 88.1% Claude Sonnet 4.6; 87.4% Gemini 2.5 Pro |
| Turnitin v3.1 | 90.8% | 9.6% | 91.2% AI; 93.8% ChatGPT-4o; 89.7% Claude Sonnet 4.6 |
| 15-detector average | Not reported in this comparison | 16.8% | Not reported in this comparison |
The average false-positive rate across all 15 detectors was 16.8% on human writing. GPTZero v3.4 had a 14.2% false-positive rate on a sub-batch of 60 essays by non-native English writers. That smaller group shows why averages may hide differences among writers.
Model results also varied. The next table lists reported detection rates for specific AI systems rather than all AI text.
| Detector | AI model output |
|---|---|
| GPTZero v3.4 | ChatGPT-4o: 92.6% |
| GPTZero v3.4 | Claude Sonnet 4.6: 88.1% |
| GPTZero v3.4 | Gemini 2.5 Pro: 87.4% |
| Turnitin v3.1 | ChatGPT-4o: 93.8% |
| Turnitin v3.1 | Claude Sonnet 4.6: 89.7% |
Match Tools to Your Checking Needs
Free checks can offer a preliminary signal, but they cannot confirm authorship. Originality.ai’s free plan offers a one-time 10,000 words, described as 100 credits. Paid plans set higher monthly limits for people who need repeated checks.
| Plan | Word allowance | Monthly billing | Annual billing |
|---|---|---|---|
| Originality.ai Free | One-time 10,000 words, 100 credits | Free | Not offered |
| Originality.ai Pro | 200,000 words per month | $14.95 monthly | $12.95 per month, billed annually at $155.40 for the year |
| Originality.ai Enterprise | 1.5 million words per month | $179 monthly | $136.58 per month, billed annually at $1,638.96 for the year |
GPTZero’s team plan offers 500,000 words per member per month, shared team credits, centralized billing, and seat management. Winston AI’s 14-day trial includes 2,000 credits for AI-content, AI-image, and deepfake detection.
Winston AI plans offer 100,000, 200,000, or 500,000 credits per month. Its 200,000-credit plan allows up to five team members, while the 500,000-credit plan allows unlimited members. These product features do not establish who wrote a passage.
A free result may help you choose what to inspect, but the text and its context still matter. Use a detector result to decide what to inspect next, never as the final authorship decision.
What Should You Do If Your Writing Is Flagged?
If your own work is flagged, ask for the specific passage and the policy used to assess it before responding to the accusation. Keep your reply calm and focused on the facts. A detector score alone is not proof of authorship.
Ask what triggered the concern and what review process applies. You can request the relevant passage, the policy for the assignment, and the steps for review or appeal. Keep the exchange respectful, even when you disagree with the result.
Ask for a Specific Review
Ask which lines raised concern, what rule applies, and who will review the work. A specific request keeps the conversation tied to the text instead of a broad claim about your character. Save messages and responses as you go.
Policies differ by school, class, and workplace. A university may set its own steps for reviewing an academic integrity concern. For example, Missouri State’s academic integrity process describes discussion and appeal steps for its students.
Explain what you did, using clear details rather than guesses about the detector. If you used an allowed tool, describe how and where. If you did not, say so plainly and offer records that help show your process.
Gather Evidence of Your Process
Useful support can include several kinds of evidence:
- Dated drafts that show how the piece changed.
- Research notes and sources that shaped the work.
- A clear explanation of the argument and its key choices.
Keep source trails, outlines, and feedback with the drafts. Be ready to explain a specific claim or word choice in your own terms. You do not need to recite the whole essay from memory to discuss how you wrote it.
Nervousness during a meeting or forgetting an older assignment does not, by itself, prove that you did not write the work. A reviewer should consider the records and the explanation together. Keep copies of your drafts and correspondence together so you can provide a clear record if a review continues.
What Is the 30% Rule of AI?
The 30% rule is a guideline people interpret in different ways, not a legal or enforceable standard. One educational interpretation suggests AI-generated content should make up no more than about 30% of a student’s final work. That figure is not universal.
Follow the policy for your specific assignment or institution instead of treating 30% as a fixed limit. Some teachers may allow certain kinds of AI help, while others may set a stricter rule. The same number does not apply to every class or workplace.
What the 30% Guideline Means for Students
The phrase “30% rule” can refer to different ideas. In one educational suggestion, it means that AI-generated content should make up approximately 30% or less of a student’s final work. It does not create a shared standard for schools.
Check the assignment instructions for what counts as AI assistance. A teacher may treat brainstorming, editing, and generated paragraphs differently. The rule may also require disclosure, citation, or limits on which tools you can use.
Do not assume that a detector score measures the share of a paper written by AI. A score is a tool’s estimate, while a classroom percentage is a policy choice. They answer different questions.
How Copyright Questions Differ
Copyright asks a separate question from a classroom guideline: whether work made with generative AI can receive copyright protection. A percentage threshold does not answer that question. The U.S. Copyright Office published Part 2 of its Copyright and Artificial Intelligence report on January 29, 2025.
The report addresses the copyrightability of outputs created using generative AI. It does not set a 30% test for students or establish a classroom rule. Part 2 of the Copyright and Artificial Intelligence report concerns copyright questions, not an allowed share of AI text.
For a class assignment, the instructor’s rule controls the work you submit. A workplace may have its own terms for AI use and disclosure. Check the specific course or workplace policy instead of assuming that one percentage applies everywhere.
How Can You Use AI Without Misrepresenting Your Work?
There is no dependable way to guarantee that AI use will escape detection, and hiding it can conflict with the rules for your work. Whether the help is acceptable depends on the assignment or workplace policy. Be clear about what you used.
Follow the relevant rules for disclosure and citation before using an AI system. ChatGPT and Claude are generative AI systems, but permission to use them depends on the task. Check whether the policy allows brainstorming, editing, drafting, or some other use.
Can You Use AI Without It Being Detected?
No tool or writing method can guarantee that AI use will go undetected. Detector results vary with the text and the detector, so trying to beat a score is not a dependable plan. Focus on whether the use follows the rules.
For example, an assignment might allow brainstorming but require you to write and check the final argument yourself. You could disclose that AI helped generate ideas, then verify each claim and explain the reasoning in your own words. That keeps the help distinct from the work you submit.
Some schools and publishers set their own citation or disclosure rules. Purdue’s guide summarizes ways to cite or disclose AI use under APA, MLA, Chicago, and IEEE styles. Check the format that applies to your class or publication.
Follow the Rules for Your Assignment or Workplace
Read the instructions before opening a tool. An instructor may allow idea generation but ban AI-written answers, while a workplace may permit editing but require disclosure. The same tool can be allowed in one setting and barred in another.
If a policy is unclear, ask for guidance before submitting the work. Keep your final text accurate and make sure you can explain its claims. A tool can help with a task, but it does not take responsibility for the finished work.
A guide to AI humanizer tools and detector limits also stresses that revised text still needs review for meaning and accuracy. Before using an AI system, check the relevant policy and disclose assistance in the form it requires.
Conclusion
The safest assessment comes from checking the writing, its evidence, and its context together. A repeated phrase or a detector score may raise a question, but neither can settle authorship alone. Look for support across the text, facts, and process records before reaching a view.
Choose one questionable passage and test its claim against a source. Then compare it with a relevant prior sample or draft. Start with one passage, verify what it says, and ask for context before drawing a conclusion.
FAQ
What are the common signs of AI writing?
Common signs include generic phrasing, repeated structures, uniform sentence rhythm, and shallow or unsupported claims. Each can appear in human writing, too, especially after heavy editing or when a writer follows a template. Check whether several patterns appear across the piece and whether its claims hold up. No single word, sentence shape, or formatting choice proves authorship.
What is the 30% rule of AI?
The 30% rule is an informal guideline, not a legal standard or a universal school requirement. One interpretation suggests that AI-generated content should make up about 30% or less of a student’s final work. Schools and instructors may define AI help differently, including brainstorming, editing, and drafting. Follow the policy for the specific assignment rather than assuming that this percentage applies.
Can AI texts be detected?
AI detectors can estimate whether text resembles patterns linked to generated writing, but they cannot conclusively establish who wrote it. Human writing can receive a false positive, while AI writing can receive a false negative. A score is one clue, not proof of misconduct. Compare the flagged passage with its evidence, context, and writing process before making a judgment.
How do I prove I didn’t use AI?
No single test can prove who wrote a piece, so build a clear record of your work. Dated drafts, research notes, source records, and outlines can show how your ideas developed. You can also explain the argument, key claims, and word choices in your own terms. Keep those records together so you can share them if a review asks for evidence.
Can you use AI without it being detected?
There is no reliable guarantee that AI use will go undetected. Detector results vary, and trying to hide assistance may break the rules for a class or job. Check whether the assignment or workplace permits the intended use. Follow its disclosure and citation rules, and make sure you can stand behind the final claims and reasoning.
What apps don’t detect AI?
No app can guarantee that a text will not be flagged or prove who wrote it. Results can change with the wording, the detector, and its version. A free check may offer an early signal, but it cannot settle authorship. Use tools to decide what deserves a closer look, then consider the writing and its context.
Can an AI detector flag my own writing?
Yes, a detector can flag human writing as AI-generated, creating a false positive. That may happen even when you wrote the work yourself. Keep drafts and notes that show your process, and ask which passage raised concern if someone reviews the result. A detector score alone is not proof that you used AI.
