Recommended Free Tools
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
You can spot writing that feels formulaic or unlike its supposed author, but no phrase, punctuation habit, or polished structure proves that AI wrote it. Treat stylistic clues as reasons to look for more context—not as evidence for accusing someone. AI detectors also make mistakes, and a chatbot cannot reliably certify whether it wrote a passage.
What can make writing seem AI-generated?
Readers may notice prose that feels generic, repetitive, or unusually unlike the writer’s established voice. Those impressions can be useful prompts to investigate how a passage was produced, but they are not a dependable authorship test. People can write in formulaic ways, and AI-generated text can be revised, translated, or mixed with human writing.
Researchers can measure stylistic patterns. A 2025 study of Japanese-language public comments and text from seven large language models examined phrase patterns, part-of-speech pairs, and function-word use. In an online task involving 403 participants, the study found that people’s ability to distinguish human from AI text was limited in that setting. Its results apply to the study’s language, samples, models, and task—not to every kind of writing. Read the Japanese-language stylometry study.
There is no validated universal checklist of visible “AI tells.” A tidy outline, generic transition, polished tone, or particular word may occur in human writing too. A separate 2025 paper examines whether frequent ChatGPT users can recognize outputs from several models, but its available summary does not establish a reliable accuracy figure or a checklist that generalizes. See the ACL Anthology paper.
Can AI detectors tell who—or what—wrote a passage?
Detectors classify text based on patterns or other signals; a result is not proof of authorship. Their performance can depend on factors such as language, genre, passage length, model, and whether a person edited or combined the text. OpenAI advises educators that, “In short, not in our experience,” when asked whether AI detectors work. It also cautions that its earlier classifier should not be a primary decision-making tool. OpenAI’s 2023 classifier announcement.
What the 2023 OpenAI classifier figures mean
On an English challenge set, OpenAI’s classifier identified 26% of AI-written samples as “likely AI-written” and incorrectly labeled 9% of human-written samples as AI-written. The first figure is a true-positive result for AI-written samples, not an overall accuracy score. OpenAI discontinued that classifier on July 20, 2023, citing its low accuracy. These are historical results for that specific tool and evaluation, not a score for current detectors generally. Read the announcement and limitations.
When detection is especially uncertain
OpenAI’s published guidance identifies short text, constrained writing, editing, and translation as conditions that can make assessment harder. A detector score should therefore be interpreted in light of what the passage is, how long it is, and how it may have been produced. Classifier limitations.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →What does a text watermark show?
A watermark is a provenance signal, not a certificate of authorship. OpenAI says its textGrain approach can indicate that an OpenAI system generated or processed text, but it cannot identify the user, quantify a person’s contribution, establish responsibility, or verify that the passage is true. OpenAI also states: “The absence of a detected watermark does not prove human authorship.” OpenAI’s text provenance announcement, October 5, 2026.
Controlled results are not everyday guarantees
In OpenAI’s controlled evaluations of watermarked English psychology answers, textGrain detected about 80% of 200-token passages and about 95% of 400-token passages at a target 1% false-positive rate. In a separate 400-token evaluation, replacing 10% of words reduced detection from about 92% to 66%; replacing 25% reduced it to 17%. These figures describe those specific evaluations, not a guaranteed result for arbitrary text. Evaluation details and scope.
Can ChatGPT tell you whether it wrote something?
No. OpenAI says ChatGPT may make up an answer about whether it generated a passage, and that answer has no factual basis. Asking the chatbot is not a reliable authorship check. OpenAI’s guidance on whether ChatGPT tells the truth.
Rank #4
How to assess a suspicious passage fairly
If authorship matters, look for information beyond a stylistic impression or a detector score. Keep the consequence of a mistaken judgment in view: a false accusation can harm a human writer, while a missed detection can overlook AI assistance.
- Compare the passage with the writer’s established work, while treating differences as a question rather than proof.
- Ask about the writing process and any relevant drafts, notes, or revision history; do not assume a detector can reconstruct that process.
- If a tool is used, check its stated language, passage-length, and text-type scope, and whether its result is a probabilistic classification or a provenance signal.
- Do not use a detector result as the sole basis for a consequential decision about a person. OpenAI’s guidance for its discontinued classifier explicitly warned against relying on it as the primary decision tool. OpenAI’s guidance.
No independent, current comparison established here supports recommending one commercial detector over another. A tool’s confident label does not remove the need to consider its scope, errors, and the consequences of getting the judgment wrong.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
