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 →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 cannot reliably prove from text alone that a passage was written by AI. AI detectors estimate whether writing resembles patterns in their training examples; they can flag human writing and miss AI-generated text. Treat a detector result as a reason to review the work, not as proof or the sole basis for a consequential decision.
What can—and can’t—tell you whether text was written by AI?
There are two broad kinds of signals: text classifiers that estimate authorship from writing patterns, and provenance systems that look for a marker embedded in eligible model output. Neither, by itself, identifies a person or establishes the full process behind a finished passage.
- A classifier score is an estimate. It does not establish who wrote a passage, and a human-written text can trigger a result.
- A negative result is not proof of human authorship. Editing, paraphrasing, short length, and other constraints can make AI-generated text harder to detect.
- Process evidence adds context. Drafts, notes, sources, version history, and an author’s explanation can help show how a piece was produced, but none automatically proves who contributed what.
For a serious decision, combine relevant context with a fair conversation and the applicable school, workplace, or institutional rules. Do not treat a detector score as an accusation.
Can I ask ChatGPT if it wrote something?
You can ask, but its answer is not reliable evidence. OpenAI says ChatGPT does not know what text it generated and may make up an answer about authorship; it describes such responses as “random and have no basis in fact.” OpenAI’s guidance on asking ChatGPT about authorship explains why a yes or no answer should not be trusted.
How reliable are AI-writing detectors?
Reliability depends on the specific tool, text, language, and conditions. There is no universal accuracy figure that applies to all detectors and documents.
OpenAI’s discontinued classifier
OpenAI discontinued its experimental AI text classifier on July 20, 2023, citing low accuracy. On its English challenge set, it identified 26% of AI-written text as likely AI-written and incorrectly labeled human-written text as AI-written 9% of the time. Those figures describe that historical classifier and test set—not current tools generally. OpenAI also warned that the classifier was unreliable on short inputs, performed worse outside English and on code, was poorly calibrated outside its training data, and could be evaded with edits. It said the tool should not be used as a primary decision-making tool. OpenAI’s classifier announcement documents the evaluation and limitations.
False positives and false negatives
OpenAI reported that its former classifier flagged human writing, including Shakespeare and the Declaration of Independence. It also noted indications of disproportionate effects on students who had learned or were learning English as a second language, and on particularly formulaic or concise writing. That is a warning about that classifier’s limitations, not a claim that every detector behaves identically. OpenAI’s educator guidance recommends using process evidence and conversation rather than relying on a detector alone.
Free tools Windows power users keep installed
One-click scans. No signup required.
False negatives matter too: OpenAI said small edits could help evade its former classifier. A low or zero score therefore does not clear authorship. Turnitin likewise advises educators to treat an AI Writing report as one data point alongside their own judgment, other information, and institutional policy. Turnitin’s report-review guidance describes how to interpret its report.
How should I review an AI Writing report?
Turnitin describes its AI Writing report as a percentage of qualifying submission text it judges likely to have originated with a large language model, with passages highlighted. The percentage is a product-specific estimate, not a finding that a named person used AI or a measure of how much assistance they received. Follow your institution’s policies and use the report alongside other evidence.
- Read the highlighted passages in context. A report points to text the system classified; it does not establish authorship or intent.
- Compare with relevant prior work and assignment context. Consider whether there are meaningful differences, while avoiding assumptions based on writing style alone.
- Ask about the drafting process fairly. Invite the student to explain decisions, drafts, notes, and sources. OpenAI suggests asking students to share specific ChatGPT conversations and to log and cite sources used with AI.
- Apply the institution’s rules. Do not make a consequential decision from the report alone. Turnitin says no tool can replace educator judgment combined with other data points in deciding whether a conversation is needed.
For more detail on what the system reports, see Turnitin’s AI writing detection model information.
How are AI text classifiers different from watermarks?
A classifier infers whether text resembles patterns associated with AI writing. A watermark detector looks for a signal deliberately embedded in eligible model output. These are different technologies: a classifier can produce a score without any embedded marker, while a watermark check depends on the signal being present and detectable.
| Evidence type | What it checks | What a result can establish | Main limits |
|---|---|---|---|
| Text classifier | Patterns in the text | An estimate that text resembles the detector’s AI examples | Can produce false positives and false negatives; results depend on text, language, model, and tool. |
| Watermark detector | A signal embedded in eligible generated text | Evidence that a supported model likely generated or processed some content | Coverage is limited; edits, short text, and constrained text can weaken or defeat detection. It does not identify a person or establish the extent of model contribution. |
OpenAI says its text watermarking approach, called textGrain, adjusts token choices to create a pattern that a detector can test for. The signal is not universal. OpenAI’s current help page describes ChatGPT text watermarking as EU-only, while API customers globally can opt in for text outputs in eligible settings; coverage varies by product, model, region, export route, and creation date. OpenAI’s provenance-signals guidance describes current scope and limitations.
In an announcement dated October 5, 2026, OpenAI said API customers globally could opt in for select models and that invisible watermarking would be added to eligible ChatGPT and Codex outputs in the EU over the following weeks. At the time of that announcement, access to its text detector was being opened to approved researchers and expert organizations. These are rollout details, not evidence that every ChatGPT or API output carries a detectable watermark. OpenAI’s October 5, 2026 announcement gives the stated rollout details.
Rank #4
What OpenAI’s watermark evaluation found
OpenAI reported that, at a 1% target false-positive rate, its watermark detector identified about 80% of 200-token passages and about 95% of 400-token passages in content such as psychology; detection was substantially lower for mathematics. In 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%, while replacing 25% reduced it to 17%. These are results from OpenAI’s stated evaluation conditions, not general accuracy estimates for all detectors or text.
A detected watermark is evidence that a supported OpenAI model likely generated or processed some content. It does not identify the person who used the model, determine how much the model contributed, or establish accuracy, ownership, or responsibility. No detected signal does not prove that a person wrote the text.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesWhat evidence should you use instead of a score alone?
Look for evidence that helps explain the writing process, and weigh it in context rather than treating any single item as automatic proof.
- Drafts and version history: These can show how the text changed over time, but a final draft or revision history may not capture every tool or contributor.
- Notes and sources: Ask how the author found and used their sources. In educational settings, OpenAI recommends documenting sources used with AI.
- An explanation of choices: Ask the author to explain specific claims, structure, or revisions. Keep the discussion focused on the work and applicable policy, not on a detector score as a verdict.
- Relevant prior work: Comparisons can provide context, but a difference in style is not proof of AI use.
These kinds of evidence can help establish process, but authorship can be mixed: someone may use AI at one stage and make substantial original contributions elsewhere. A polished passage alone cannot reveal that history.
