You cannot reliably prove from text alone that a passage was—or was not—written with AI. The fairest assessment looks at the writing process, relevant policy, available provenance signals, and the writer’s explanation. A detector result can prompt questions, but it is not an authorship verdict.
Start by defining what you need to establish
“Did AI contribute?” is not the same question as “Was AI use disclosed as required?” or “Is this claim or source inaccurate?” Identify the applicable rule and the specific concern before reviewing the work. A detector score cannot decide policy, responsibility, or accuracy. Turnitin cautions that its AI-writing model can misidentify human, AI-generated, and AI-paraphrased text, and should not be the sole basis for adverse action against a student (Turnitin AI Writing Report guide).
As an Amazon Associate I earn from qualifying purchases.
Review how the writing developed
When appropriate and permitted by the relevant rules, look at evidence of the work in progress: drafts, version history, outlines, notes, source lists, and intermediate revisions. These materials may show how a piece developed, but no single item necessarily proves who wrote it.
Ask the writer to walk through their research, explain key choices, and describe what changed during revision. In an educational setting, OpenAI suggests that students document and cite sources used with AI; relevant interactions may also give educators a chance to observe critical thinking and problem-solving (OpenAI education resources). Treat this as contextual evidence, not a universal requirement. Follow institutional rules and privacy expectations, and apply the same approach consistently.
#1 Best Overall
Compare writing only with a fair baseline
If you have earlier work to compare, choose examples that are genuinely comparable: similar genre and assignment, language, time constraints, and level of editing support. A change in style can justify a neutral question; it does not establish AI authorship.
Polished, predictable, concise, or formulaic prose is not proof. OpenAI has noted that its former classifier wrongly flagged some human writing, including Shakespeare and the Declaration of Independence, and warned of possible disproportionate effects on English learners and people whose writing is especially formulaic or concise (OpenAI’s announcement about its retired AI classifier). Style impressions are especially weak when separated from context.
Understand what a detector result actually means
A detector estimates whether text resembles patterns its model associates with AI. Its output is a product-specific signal, not a probability that a particular person cheated or a forensic identification of an author. Turnitin, for example, says its report percentage concerns qualifying prose sentences that its model estimates could be AI-generated, including text that may have been modified by an AI paraphraser or bypasser. That percentage is separate from the similarity score (Turnitin’s detection capabilities and 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 →Interpret any report in light of whether the text fits the tool’s supported conditions. Turnitin’s current documentation specifies a minimum of 300 words of prose, a maximum of 30,000 words and 100 MB, supported languages of English, Spanish, Japanese, and Arabic, and accepted file types DOCX, PDF, TXT, and RTF. It says detection is not reliable for non-prose such as poetry, scripts, or code, or for short-form and unconventional writing such as bullet points, tables, and annotated bibliographies. Requirements and model behavior can change, so check the product’s current documentation before interpreting a report.
Scores can also be suppressed or unavailable. Turnitin says false positives were more frequent in its 0–19% range; it now suppresses scores and highlights above zero and below 20%, showing an asterisk instead. Reports created before July 8, 2024 may show older numerical results below 20%. This display policy is intended to reduce misinterpretation; it does not establish that every higher score is correct (Turnitin’s detection capabilities and limitations).
Performance varies across tools and changes as generators and detectors evolve. OpenAI retired its own classifier on July 20, 2023, citing low accuracy. In an English challenge set, that classifier identified 26% of AI-written text as “likely AI-written” and incorrectly labeled human-written text as AI-written 9% of the time. OpenAI said it was unreliable on short text, weaker in languages other than English, vulnerable to edits, and potentially overconfident on inputs unlike its training data. Those figures describe that retired classifier’s evaluation—not today’s detectors generally (OpenAI’s announcement about its retired AI classifier). NIST’s 2025 evaluation likewise found substantial variation: some tested generators could deceive most discriminators, while some discriminators detected content from almost all tested generators (NIST, 2025 evaluation).
Rank #4
Treat provenance watermarks as narrow signals
A verified watermark can indicate that a participating system generated or processed some text under conditions its watermark supports. It does not identify the person who used the system, quantify human contribution, settle ownership or responsibility, or verify factual accuracy. An absent watermark does not prove that a person wrote the text.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteOpenAI’s October 2026 announcement describes textGrain, an invisible statistical signal in word choice. At announcement, API customers globally could opt in for select models while the feature remained off by default; OpenAI said eligible ChatGPT and Codex output in the EU would receive watermarks, with detector access initially limited to approved researchers and expert organizations. Those rollout and access conditions are time-sensitive; consult the announcement for current details (OpenAI’s textGrain announcement).
OpenAI reports that, at a target 1% false-positive rate, its detector identified watermarks in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology. Detection was substantially lower for mathematics, where word choice is more constrained. In the company’s 400-token tests, replacing 10% of words with synonyms reduced detection from about 92% to 66%; replacing 25% reduced it to 17%. These are OpenAI evaluation results under the described conditions, not universal authorship-test rates. Short, edited, translated, unsupported, pre-watermark, or other-provider text may have no detectable watermark.
Use a detector only as corroboration
If you use a detector, first check its supported language, length, format, and text type. Keep the report and its context, including which version of the text was submitted. Do not read a percentage as a probability of misconduct, and do not use a score by itself to make a consequential decision.
There is no universal test that identifies the author of arbitrary text, and the evidence here does not establish a current independent, apples-to-apples ranking of consumer detectors. Compare a tool’s fit to the specific text, how it handles false positives and uncertain results, what systems and edits it covers, and what its output actually represents. A score, highlighted passage, or watermark each answers a different and limited question.
Give the writer a fair chance to respond
For a high-consequence decision, consider the process evidence consistently, apply the governing policy, and let the writer explain. Separate what you can establish about the work’s process from what the rules require and from what conclusion the evidence can support. If the evidence is inconclusive, describe it that way rather than making a definitive authorship claim.
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.
