Can AI detectors really tell if something was written by AI? Sometimes they can identify patterns or signals associated with particular AI systems, but a detector score is an inference—not proof of who wrote a passage, how it was produced, or why. Post-hoc classifiers and watermarks look for different evidence, cover different outputs, and can both fail. For a consequential decision, treat a result as a reason to review the work and its provenance, not as a verdict.
What does an AI text detector actually detect?
Most tools fall into two broad categories. A post-hoc classifier examines text after it has been written and estimates whether it resembles examples associated with machine-generated or human-written text. A watermark detector instead searches for a signal that a participating generation system was designed to insert while producing text.
As an Amazon Associate I earn from qualifying purchases.
Neither method identifies the person at the keyboard. And neither resolves the full range of real writing practices: a person may use AI to generate a draft, edit a passage themselves, ask AI to revise their own writing, or combine human and machine-written sections. Those are different forms of AI use, even if a detector reduces them to a label or score.
Post-hoc classification
A classifier makes a statistical judgment from features of the text and its relationship to the examples or signals it was tested on. It may be able to assess output from more than one system, but performance can change when the model, language, genre, subject, length, or editing differs from the evaluation data. Its score is not a direct record of how the passage was created.
Watermark detection
A watermark is a signal embedded by a generation system that participates in a particular watermarking scheme. Detecting it can support a claim that text is consistent with that scheme. It does not establish who prompted the system, whether the user submitted the output unchanged, or what role AI played in the finished work. It also cannot reliably cover output from systems that do not apply that watermark.
How do classifiers and watermarks differ?
| Question | Post-hoc classifier | Watermark detector |
|---|---|---|
| What is examined? | Patterns in text after it exists. | A signal associated with a particular watermarking scheme. |
| Which systems can it cover? | Potentially multiple generators, depending on the tool’s evaluation and how closely the input matches tested material. | Outputs from systems that insert a detectable signal; it does not automatically cover unwatermarked systems. |
| What can a positive result support? | That the text resembles material the classifier associates with AI generation, subject to its error rate and scope. | That the text contains a signal consistent with the watermarking scheme. |
| What can it establish about the writer? | It cannot identify the writer or prove how much assistance they used. | It cannot identify who operated the system or establish the writer’s intent. |
| What can make results uncertain? | Unmatched models, languages, genres, short passages, editing, or other differences from tested material. | Unsupported systems, short or altered text, translation, or changes that weaken the signal. |
What does the evidence show about the detection arms race?
NIST’s text-to-text pilot found uneven results
NIST’s 2024 GenAI pilot evaluated curated groups of articles and human- and machine-generated summaries. Its report, cited in 2025, used measures including AUC and Brier scores. NIST found that generated summaries increasingly resembled human writing, while discrimination remained reasonably effective in the test. Results varied substantially by system: some generators deceived most tested discriminators, while some discriminators detected content from almost all tested generators. Discriminator performance also improved across rounds.
That is evidence of movement on both sides, not a universal leaderboard. NIST’s summary does not give one headline accuracy figure that applies across detectors and real-world writing. Its curated benchmark cannot by itself predict how a tool will perform on a particular student essay, newsroom submission, translated passage, or workplace document.
Free tools Windows power users keep installed
One-click scans. No signup required.
Watermarks have a narrower but different kind of evidence
Google DeepMind’s peer-reviewed SynthID-Text work reported feedback across approximately 20 million Gemini chatbot interactions in 2024. That is a deployment-scale evaluation, not proof that every output was independently authenticated. The paper describes watermarking as complementary to post-hoc methods: it depends on systems applying the watermark and cannot reliably cover decentralized open models. It also identifies ways a watermark may be weakened, including stealing, spoofing, scrubbing, and paraphrasing.
Google’s developer documentation, updated April 9, 2025, describes SynthID Text detection as probabilistic, with three possible states: watermarked, not watermarked, and uncertain. Thresholds can be adjusted to target false-positive and false-negative rates. Google notes that factual responses offer less room to alter generation without reducing accuracy, and that thorough rewriting or translation can lower detector confidence.
Watermark robustness findings depend on the method and test. A separate ICLR 2024 study by Kirchenbauer and coauthors found that some watermarks remained detectable after human or machine paraphrasing, in part because paraphrases sometimes retained fragments of the original. In its particular setup, the authors reported detection after strong human paraphrasing at an average of 800 tokens and a 1e-5 false-positive rate. That result should not be generalized to all watermark schemes or detectors.
OpenAI announced text watermarking in October 2026
On October 5, 2026, OpenAI announced that API customers globally could opt into text watermarking on selected models. It said eligible ChatGPT and Codex text output would receive an invisible watermark in the EU over the following weeks, and that detector access would initially be available to approved researchers and expert organizations. OpenAI describes textGrain as an invisible statistical signal in word choices and cautions that strong performance in ideal evaluations does not guarantee reliability in everyday use.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsOpenAI reported its own evaluations at a target false-positive rate of 1%: watermark detection was about 80% for 200-token psychology passages and about 95% for 400-token psychology passages. It reported substantially lower performance on mathematics text. These are company evaluation results for the stated material and conditions, not independent benchmarks for all writing.
OpenAI also tested synonym replacement on 400-token passages. In its 2026 evaluation, detection fell from about 92% to 66% after 10% of words were replaced with synonyms, and to 17% after 25% were replaced. The figures show how editing affected that evaluation; they do not predict the exact result for every passage or watermark detector.
OpenAI’s Help Center, accessed in 2026, says text watermarking is being implemented for ChatGPT users in the EU in response to the EU AI Act and the company’s commitments under the EU Code of Practice on Transparency of AI-Generated Content. It describes detector access for qualifying organizations on a case-by-case basis and says the approach may evolve. In the company’s described 24-language evaluation, detection at a 1% false-positive rate was 69.0% for Spanish and 42.2% for Romanian. Those are test results for that evaluation, not general performance estimates for Spanish or Romanian writing.
Can a detector prove I used AI?
No. A detector can return evidence consistent with AI generation, but that is not the same as proof of authorship or use. OpenAI explicitly says, “The absence of a detected watermark does not prove human authorship.” The company lists short, edited, or translated text, unsupported models, text created before watermarking, and output from other providers among reasons its watermark might not be detected.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →A detected watermark is also limited evidence: it points to a signal associated with a watermarking scheme, not to the identity of the person who used the system or the purpose for which they used it. A post-hoc classifier’s result has a different limitation: resemblance to patterns in its evaluation does not prove that a person used AI. In both cases, the result must be interpreted within the tool’s tested scope.
It also matters what “used AI” means in the situation being judged. Generating a complete passage, using a tool to brainstorm, asking for copy edits, and translating or polishing a human draft are not equivalent acts. A binary detector generally cannot establish those distinctions from a final text alone.
Why can false positives matter even when a tool reports a low rate?
A false-positive rate is not the same as the probability that a flagged passage was written by a human. The share of flagged passages that are false positives depends on how common AI-generated text is in the population being screened, as well as on the detector’s sensitivity and error rate.
For illustration only, imagine screening 1,000 passages where 10% are AI-generated, using a hypothetical detector that catches 80% of AI passages and has a 5% false-positive rate on human passages. It would flag 80 AI passages and 45 human passages. Of its 125 flags, 45—more than one in three—would be false positives. Those assumed rates are not a claim about any real detector; the example shows why a tool’s operating threshold and the screening population matter.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Rank #4
The harms are not hypothetical simply because a detector reports a number. A 2026 in-the-wild audit by Dutta and coauthors, published in the Proceedings of the International AAAI Conference on Web and Social Media, reported considerable false positives across the systems it examined. In the audited data, liberal political discourse was flagged as AI-generated more often than conservative political discourse. This is a finding from that audit, not evidence that every detector has the same bias or error pattern.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How should you evaluate a detector claim?
Before using a score to make a decision, find out what the number actually measures. A vendor’s best result on a selected test set is not independent validation, and an overall accuracy figure can hide errors that matter for the particular language, genre, or population being screened.
- Identify the evidence type: Is the tool classifying text after the fact, checking for a specific watermark, or combining methods?
- Check the tested material: Look for the languages, genres, subjects, text lengths, models, and human-versus-AI comparison used in evaluation.
- Ask about editing: Determine whether paraphrasing, translation, synonym replacement, or ordinary human revision was tested.
- Read both error rates: Ask for false-positive and false-negative rates at the threshold the tool uses, not just a detection rate or accuracy headline.
- Check for uncertainty: Find out whether the system can abstain or return an “uncertain” result rather than forcing a binary label.
- Establish access and permitted use: Confirm who can use the tool, where it is available, and whether the provider permits the intended purpose.
- Preserve context: Keep drafts, revision history, source notes, and relevant provenance records where appropriate.
What is a fair process for schools, employers, and publishers?
For a consequential decision, a detector output should initiate a review, not end one. Ask what the score can and cannot establish, and give the person whose work is being assessed a fair chance to explain their process. Consider the relevant policy and the work itself alongside the output.
- Define the question first. Decide whether the concern is undisclosed AI generation, a specific form of assistance, or a breach of a stated rule. A detector cannot supply that policy distinction.
- Review evidence beyond the score. Where relevant and lawfully available, consider drafts, revision history, notes, citations, and the person’s account of how the work was produced.
- Check whether the tool fits the case. Compare the passage’s language, length, genre, editing history, and likely generation system with the detector’s evaluation conditions.
- Document uncertainty and allow response. Record the tool and threshold used, explain the limits of the result, and allow the affected person to respond before taking action.
- Use proportionate consequences. Do not treat an uncorroborated score as conclusive proof in discipline, hiring, or publication decisions.
Why transparency needs more than one detector
NIST’s AI 100-4, published in 2024 and updated April 8, 2026, surveys content authentication and provenance, synthetic-content labeling such as watermarking, detection, testing, and auditing. That broader framework matters because no single detector can establish every fact readers may want to know. Transparency can combine signals and records with clear policies, evaluation, and review procedures—each addressing a different part of the problem.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
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.
