I pasted a human-written text into an AI detector. It came back labeled AI. The awkward part was that the detector belonged to my own website.
So I tried Pangram, GPTZero, ZeroGPT, QuillBot, Detecting AI, and Sapling with German and English factual writing. I also fed the same full texts I used with Pangram, GPTZero, ZeroGPT, QuillBot, and Detecting AI to my now-retired German Gradually detector.
- Pangram gave the expected class for all six complete texts, but misplaced the beginning of the AI passage in one English mixed document.
- GPTZero and QuillBot recognized the four pure texts. They judged the mixed texts mostly or entirely human.
- ZeroGPT, Detecting AI, and Sapling produced false alarms on human writing. Sapling also cut off longer inputs after 2,000 characters.
If I had to check a text today, I would start with Pangram. GPTZero would be my second choice for a purely human or AI-written text. I would be more cautious with Sapling, especially when it gives human writing a high AI score.
1. Pangram
I entered six factual texts into Pangram. One human, one AI-generated, and one mixed document in each language. The human sources covered photosynthesis in German and the water cycle in English.
All six reports gave the right result. I found the mixed documents more revealing. Pangram noticed that the writing came from both sources, but the English highlight was not precise. The first 19-word AI sentence still appeared in the human section.
My six Pangram checks from September 19, 2026. The source of each text was known before submission.
The interface worked well for me. Reports appeared in the history, and I could reopen the last one. I also imported and analyzed a DOCX file and a text-based PDF. I would still read any highlighted passage myself before drawing a conclusion.
Pangram is where I would start for a second opinion. Six texts do not establish an overall accuracy rate. My separate Pangram review has more on using the tool.
The pricing page lists 2,000 free words per day. Individual costs $20 per month for 300,000 words. I used a license I already had.
2. GPTZero
I pasted the same six full factual texts into GPTZero that I used with Pangram. It classified both human texts as human and both purely AI-written texts as AI. Its AI score for the German AI text was 66%, though, so even that result was less clear than the English one.
The mixed texts told a different story. GPTZero showed 51% human and 49% mixed for the German text. For the English one, it showed 93% human and just 6% mixed. Both documents contained human and AI-written passages.
The same six factual texts at GPTZero on September 24 and Pangram on September 19, 2026. GPTZero's percentages estimate a class for the whole document, not the share of AI-written words.
My free account had 10,000 monthly credits. Each scan used one credit per word. Free sentence highlighting ran out after the first scan, but whole-document classification kept working. You can check the current plans on GPTZero's pricing page.
GPTZero was useful for the purely human and AI-written texts in this small set. If you are checking a revised text or one with only some AI-written passages, I would not rely on its document label alone.
3. ZeroGPT
ZeroGPT let me check all six full texts without signing in. The first result made me stop. It labeled the human-written German photosynthesis text “AI/GPT Generated” and showed 69.8% AI/GPT.
The two purely AI-written texts scored 75.1% and 99.2%. The mixed texts scored 90.2% and 80.9% AI/GPT, even though each contained a human-written passage. I did not get a separate mixed category. The human English text was stranger still. ZeroGPT called it likely human, then showed 62% AI/GPT beside that verdict.
The same six factual texts at ZeroGPT on September 24 and Pangram on September 19, 2026.
gradually.aiThat German false alarm is enough to make me cautious. I would also look closely at the conflicting label and score for the English human text before sharing that result with anyone.
4. QuillBot
QuillBot called both human texts 100% human. It gave the purely AI-written texts AI scores of 52% and 76%. The German AI text was only just over halfway to AI.
The mixed texts told a different story. Each contained a human passage and an AI-written passage. QuillBot still showed 0% AI and 100% human for both. Its separate score for human-written and paraphrased text was also 0%. I would not rely on that reading for a document written partly with AI.
The same six factual texts at QuillBot on September 24 and 25 and Pangram on September 19, 2026.
gradually.aiAfter five scans in one day, QuillBot told me no scans were left. I checked the remaining text the next day. Its product page says the free plan allows up to six scans per day and 1,200 words per scan.
5. Detecting AI
I could check all six full texts with Detecting AI without creating an account. It gave the human-written German photosynthesis text a 7% AI score and called it likely human. But it classified the human English text as mixed, with a 42% AI score. It even marked 10 of its 32 sentences likely AI. That is a false alarm I would not ignore.
The two purely AI-written texts scored 65% and 67%, yet both received a mixed overall verdict. The German mixed text surprised me more. It scored 7% and was labeled likely human. Only the English mixed text got the expected mixed label. Of the four texts written entirely by a human or AI, Detecting AI gave just one the right overall class.
The same six factual texts at Detecting AI on September 25 and Pangram on September 19, 2026.
gradually.aiAfter these results, I would not treat its document label or colored sentence marks as proof of who wrote a text. I ran these checks through the public input page.
6. Sapling
When I first pasted a text into Sapling, the score from its sample text stayed on screen. I had to click “Check Again” to run my own check. Then I tried longer inputs and ran into another problem. Sapling evaluated only the first 2,000 characters. In a mixed document with an AI section at the end, that can leave out the part you wanted to check.
I then used six shorter texts, each under 1,600 characters. Sapling processed all six in full. Two came from older human-written factual sources. Sapling still showed 97.7% and 84.8% AI for them.
My six Sapling checks from September 19, 2026.
gradually.aiThe colored sentence highlights look decisive. After those two false alarms, I would not use them to accuse a writer. Even a displayed 100.0% remains a model score, not certainty.
Sapling's pricing page lists 2,000 characters per free check and Pro at $25 per month. I could not establish why the longer Pro allowance did not apply in my account. What mattered in my check was the text Sapling actually evaluated.
7. Gradually
My own detector received the same six full texts as Pangram, GPTZero, QuillBot, and ZeroGPT. It did not do well. It labeled both human factual texts as AI. In the separate short-text check, it also missed a German AI-written text.
There was a more worrying interface problem. When I replaced or cleared the input, the old result remained visible. The history showed disabled restore buttons. It would be easy to connect an old verdict to a newly pasted text.
The German version has since been removed. I would not recommend it after this experience.
8. My verdict
If you want to try a detector now, start with Pangram and a human text you wrote yourself. It handled my six texts well, but the misplaced highlight is a reminder to read the passage behind any label.
GPTZero and QuillBot also found the right class for all four pure texts. GPTZero leaned toward human for both mixed texts; QuillBot called them entirely human. ZeroGPT labeled three of the four pure texts correctly, but displayed a 62% AI/GPT score even for the English text it called likely human. Detecting AI got the overall class wrong for three of the four pure texts. My own Gradually detector got two of the four pure texts right.
gradually.aiSapling flagged both of my human factual texts with high AI scores. With longer inputs, check which part was actually analyzed. Sapling received a different, shorter set of texts, so its numbers are not directly comparable with Pangram and GPTZero.
My advice is simple: begin with writing you know is human because you wrote it. If the detector gets that wrong, you have learned more than ten marketing claims could tell you.
9. How I tested the AI detectors
I wanted to answer three questions. Can a detector recognize a fully AI-written text? Does it leave human writing alone? And does it notice when I add an AI passage to a human text?
I used six factual texts, three in German and three in English. The German texts explain photosynthesis; the English texts cover the water cycle. Each language has one human text, one AI text, and one mixed text.
- Human texts. Both excerpts come from Wikipedia revisions dated 2021. The German text contains 406 words; the English text has 578.
- AI texts. I generated these on September 12, 2026, using the configured
gpt-5.6-lunamodel. The outputs contain 506 German words and 496 English words. - Mixed texts. I combined the beginning of each human text with an AI passage. In German, 111 of 451 words come from AI; in English, 121 of 472. That puts the AI share at roughly a quarter.
The writing instructions were straightforward. The German prompt asked how photosynthesis works, which substances and forms of energy are involved, and why it matters to ecosystems. The English prompt covered evaporation, condensation, precipitation, infiltration, and water storage, including the roles of solar energy and gravity.
Both prompts requested 450-550 words for interested adults without specialist knowledge, in connected paragraphs without headings or lists. I did not use a humanizer or ask the model to evade detection.
Between September 19 and 25, 2026, I checked the same six complete texts with Pangram, GPTZero, ZeroGPT, QuillBot, Detecting AI, and my own Gradually detector. Sapling received six different, shorter texts because of its character limit. Its results therefore appear separately.
The comparison chart counts only the four purely human or fully AI-generated texts. For the mixed texts, I also look at whether the detector finds the added AI passage. I do not convert the percentage displays into a shared detection rate, because the providers use them to mean different things. Six texts on two topics reveal specific strengths and mistakes, not an overall accuracy rate for every writing style and AI model.
10. How do AI detectors work?
A detector cannot see who wrote a text. It looks for patterns that resemble its examples of human or AI-generated writing. There are several ways to do that.
Text features and probabilities
Simple methods count repeated words and phrases or compare sentence lengths. Readability measures such as the Flesch score describe how easy a text is to read. That tells you something about the writing, but it does not establish its origin.
Other methods use a language model to estimate how predictable the words are. “Perplexity” describes how well a model can predict a sequence. In discussions of detection, “burstiness” often refers to variation in these patterns or in sentence structure. But predictable language is not automatically AI language. A human-written instruction manual can be highly predictable, too.
Trained classifiers
A classifier learns from labeled examples which patterns tend to appear in human and generated texts. It does not have to use the same model that produced the writing. The Pangram 4 model card, for example, describes an architecture with multiple specialized models and categories for human, AI-assisted, and AI-generated writing.
What matters is how closely the training examples match the text being checked. Good results on English school essays do not automatically carry over to German product descriptions. GPTZero also notes limitations involving text types, languages, and short inputs. A polished percentage display does not remove those limits.
11. Can AI text contain a watermark?
Yes. Google uses SynthID for text in Gemini. Anthropic also embeds text watermarks in responses from certain Claude models. The blanket claim that AI writing has no watermarks is no longer accurate.
But providers do not mark every response in the same way. A watermark in an image tells you nothing about whether the accompanying chat text was marked. Even within one provider’s products, the answer can depend on the model and how you access it.
Which providers mark their AI text?
For a copied article or an essay, the important question is whether the mark is embedded in the text itself. This overview separates that from image watermarks, file metadata, and labels in a chat interface.
As of September 30, 2026. Confirmed means documented by the provider for the specified output and access route. Unconfirmed does not mean proven watermark-free.
“Unconfirmed” does not mean “watermark-free.” A missing public explanation is different from a provider’s assurance that its text contains no marking at all.
Google Gemini and SynthID
Google’s SynthID page explicitly identifies the Gemini app and web experience as places where text watermarks are used. This is a deployed method, not just a research proposal.
The public mapping to individual models is less clear. I would not read “Gemini uses SynthID” as a guarantee that every Gemini Pro, Flash, or Flash-Lite version is marked identically through every access route. That page does not provide a complete table covering the Gemini API, Vertex AI, and all model versions.
Generation and verification are separate, too. Google describes public SynthID verification for images, video, and audio on that page. This does not establish a generally available text checker where you can paste any paragraph and receive an official SynthID result.
Google’s open models are another case. The SynthID Text reference implementation includes examples using Gemma 1.0 2B-IT and 7B-IT, as well as GPT-2. That shows these models can be run with the method. It does not mean every locally generated Gemma or GPT-2 passage automatically carries a Google watermark.
Anthropic Claude and the specific models
Anthropic’s model table distinguishes the following Claude versions.
Text watermarks on first-party products/API and AWS, Google Cloud, and Microsoft Foundry. As of September 30, 2026.
Anthropic’s technical explanation describes a version of SynthID Text. It marks the generated word sequence. The Fable 5.1 documentation also confirms that it adds neither hidden characters nor information about you or your organization.
I would not translate a Claude watermark into “Claude wrote everything from scratch.” Anthropic’s technical explanation explicitly discusses processed content, too. Translation or editing can produce text carrying a Claude signal even when a human supplied the original writing and ideas.
According to the Fable 5.1 and Mythos 5.1 announcement, official text verification is initially available to selected organizations in a private preview. An individual user does not automatically have a freely available Claude text-watermark scanner. The Content Credentials file checker is a different tool.
ChatGPT and OpenAI
For OpenAI, saying “ChatGPT text now carries SynthID” would go beyond the evidence. Its current provenance documentation describes SynthID for image and audio outputs, plus Content Credentials for images. It names text as a goal for further expansion.
That does not confirm statistical watermarking in the text responses of every GPT model in ChatGPT, Codex, or the API. This source does not provide a specific GPT model list for that use. Equally, I would not treat it as a guarantee that every ChatGPT passage is watermark-free.
A screenshot makes this particularly easy to confuse. A watermark in a generated image would be a signal about that image. It would not establish that a separately copied image description or chat response contains the same watermark.
Meta, Grok, DeepSeek, Mistral, and Qwen
With the other providers, it is worth checking exactly what is being marked.
- Meta. The Meta AI image-generation page describes Content Seal as an invisible image watermark. This does not confirm equivalent marking in Meta AI chat text or all Llama models.
- xAI. The Grok FAQ explicitly discusses watermarks in generated images and videos. That is not a statement about copied Grok text responses.
- DeepSeek. Section 5.3 of its Chinese terms of use confirms labels for AI content within the platform. It explains neither a SynthID-like method nor a mapping to specific text models. “Labeled as AI” and “statistically marked in copied text” are not interchangeable.
- Mistral. Section 2.5 of its partner-served deployment terms mentions watermarks for tracing model copies under a particular arrangement granting access to weights. The model itself is the subject here. This does not establish universal text watermarking in Le Chat.
- Qwen. The watermark parameter in the Qwen-Image documentation adds a visible label to the bottom-right corner of an image. It is not an equivalent feature for the word sequence of Qwen chat text.
How is a statistical text watermark created?
A language model assembles its response from tokens. These are text units that can be shorter than a word. For many phrases, several continuations are possible. A text watermark uses that flexibility to leave a detectable pattern across many generation decisions.
You do not have to see an odd word or a particular character sequence. SynthID influences selection during generation. Google’s developer documentation describes configuration using keys and a matching detection method. A plain text editor therefore cannot reveal the pattern like an extra space.
This also explains why removing invisible characters is not a universal solution. That operation deletes characters. A statistical watermark resides in the selection decisions behind the readable text units themselves.
I would keep four things separate.
- Statistical text watermark. A pattern in the generated word sequence, checked by a suitable detection method.
- Invisible Unicode characters. Extra characters in the text. Their presence alone does not prove SynthID or identify a particular AI.
- Content Credentials. Signed provenance information associated with a file. Copying its visible contents does not automatically produce a text watermark.
- Visible AI label. A notice in an interface or on an image. The notice and the actual text contents are different things.
Can locally run models generate watermarks, too?
Yes, if their generation software is configured to do so. Downloadable weights do not automatically make an open model watermark-free. What the software does while generating the response matters, too.
vLLM provides a concrete example. Its developers show how to enable a statistical Gumbel-max watermark while serving Mistral-7B-Instruct-v0.3. That is an operator configuration, not a claim that Mistral supplies this watermark to every user by default.
Likewise, Google’s developer guide explains how to add SynthID explicitly to text generation. For a third-party service running Gemma, Llama, Qwen, or another open model, I would therefore ask about the actual deployment, not just the model name.
How reliable is watermark verification?
An embedded pattern is not an indestructible stamp. Google’s explanation of SynthID’s limitations identifies short answers, tightly constrained wording, extensive rewriting, and translation as challenges.
An answer of “Paris” to a question about France’s capital leaves little room for many different word choices. A longer, freely written paragraph offers more opportunities. Still, there is no universal minimum length at which every provider’s watermark is guaranteed to be detected.
Translation is not a simple yes-or-no case, either. It can weaken the signal in the original word sequence. But if a model with active text marking generates the translation, that new output can carry a new watermark. Anthropic explains this distinction in its technical description.
Finally, the checker must match the method. A tool’s ability to identify some AI writing does not mean it has the right key and detection method for a particular watermark. The label “AI detector” is not enough.
What does this mean for Pangram, GPTZero, and your writing?
A general AI score and a watermark check answer different questions. An AI detector evaluates how the text looks to its model. A watermark checker looks for a specific intentionally embedded mark.
A high Pangram or GPTZero score therefore does not establish that the tool found a SynthID or Claude watermark. If a tool claims such a finding, I would expect a report explicitly naming the method and its result.
For an important decision, I would ask these five questions.
- Which provider and specific model generated or processed the text?
- How was the response generated, such as through a chat app, an API, or a self-hosted server?
- Was text marking documented for that route at the time of generation?
- Does the checking tool verify that watermark, or does it only return a general AI score?
- Is this the original text, or has it been shortened, translated, edited, or combined with other passages?
A positive signal may indicate processing by a particular provider. It does not automatically identify the person behind the text, the complete writing process, or the proportion of original ideas. A negative result is not proof of human authorship, either. For an accusation or an assessment, I would still consider drafts, sources, and revision history.
12. How can you recognize AI writing yourself?
Not from a single word. “Delve,” “in today’s digital world,” flawless grammar, and equally sized paragraphs are not enough. They might prompt a closer look, but they are not proof.
When reading a text, I would pay more attention to these questions.
- Lots of words, little added information. Does the text repeat the same point across several paragraphs without explaining it more clearly?
- Claims without usable evidence. Does it mention studies, tests, or personal experiences that you cannot verify?
- Contradictions and incorrect sources. Do the numbers, names, and examples agree? Do the linked sources exist, and do they actually support the claims?
- Unexpected changes in style. Does the vocabulary or level of knowledge shift noticeably midway through? That may suggest multiple editing steps, but those steps need not involve AI.
Keep the alternative explanations in mind. Humans write formulaic sentences and make factual mistakes. AI can produce casual language, unusual coinages, and deliberate typos. Treating polished writing as a sign of AI is an easy way to accuse the wrong person.
When the outcome matters, check the actual claims and sources first. Then ask for drafts, revision history, or an explanation of how the text came together. A detector can provide another clue. It cannot replace that conversation.
Frequently Asked Questions About AI Text Detectors
Yes. Pangram’s pricing page lists up to 2,000 free words per day with no payment method required as of September 2026. That is enough to try a few of your own texts. I also used GPTZero, QuillBot, and the other detectors shown here for free within their respective limits.
Look beyond the number of free scans. Check how much text each scan covers and whether you can see the highlighted passages. Sapling cut my long input off after 2,000 characters.
It does not automatically mean that AI wrote 80% of the words. Tools use percentages for different things. GPTZero’s class probabilities describe its classification of the whole document. Pangram describes its document fractions as the share of characters it assigns to each category.
Those assignments can still be wrong. Read the explanation for the score and inspect the highlighted passages. I would not directly compare percentages from different detectors or average them into a supposedly more accurate result.
No. AI detection estimates whether AI may have been involved in writing a text. Plagiarism checking looks for matches with existing sources. QuillBot explains the same distinction.
Copied human writing can pass an AI check. Conversely, a high AI score does not show that someone copied a particular source. When assessing an article, check its origins, citations, and factual accuracy separately. One reassuring detector result cannot answer all three questions.
