No AI detector can tell you for certain whether a person or a model wrote a text. The tools below, GPTZero, Originality.ai, Winston AI, Pangram, Sapling, ZeroGPT, Copyleaks and Turnitin for text, and Reality Defender and Sightengine for media, give you a probability score. Use that score as one signal and never as proof, especially when a person’s grades, job or reputation depends on it.
How AI detectors work and why they get it wrong
Text detectors look at patterns in writing: word choice, sentence structure and how predictable each next word is. A model trained on examples of human and machine text then produces a score. Image and video detectors look for traces left by generators, such as odd textures, lighting or face edges.
That approach has built-in limits, and the vendors say so themselves. GPTZero states that no AI detector is fully accurate and that results should not be used to punish anyone. Originality.ai says its score should be one signal and not a final decision. Sapling says no detector, its own included, should be the only basis for deciding whether text is AI-generated.
These are the common failure points.
- False positives. A detector can flag a human’s writing as AI. Plain, formal or heavily edited text is at higher risk. Writing by people who learned English as a second language is often cited as a concern, and several vendors say they work to reduce it.
- False negatives. Paraphrased or lightly edited AI text can pass as human.
- Short text. A sentence or a short paragraph gives the tool too little to judge. Sapling says results get more reliable after about 50 words, and Originality.ai sets a minimum length.
- Edited text. Grammar and rewriting tools can change a human draft enough to trigger a flag. Originality.ai names some grammar tools as a cause of false positives.
- New models. Detectors have to keep up with each new generation of AI models, and there is always a lag.
- Language. Most text detectors work best in English. Sapling lists the languages it supports and says others are less reliable.
One telling example: OpenAI released its own AI Text Classifier in early 2023 and took it down a few months later, citing a low rate of accuracy. If the company behind the most popular chatbot could not make a dependable detector, treat every score with care.
Quick comparison
| Tool | Detects | Free option | Best for |
|---|---|---|---|
| GPTZero | Text | Yes, with a free account | Teachers and students |
| Originality.ai | Text, plus plagiarism checks | Yes, limited free scans | Publishers and SEO teams |
| Winston AI | Text and images | Free trial | Educators and publishers |
| Pangram | Text and images | Yes, with a free plan | Schools and editors |
| Sapling | Text | Yes, no sign-up for the free check | Quick checks and developers |
| ZeroGPT | Text, images and video | Yes, free detector | Casual, quick checks |
| Copyleaks | Text, plus plagiarism checks | See the vendor site | Schools and businesses |
| Turnitin | Text, inside its academic platform | Sold to institutions | Universities and schools |
| Reality Defender | Deepfake audio, video and images | Free API key offered | Security and media teams |
| Sightengine | AI images, video and deepfakes | See the vendor site | Moderation on apps and platforms |
1. GPTZero
GPTZero is one of the best-known text detectors and was among the first to reach a wide audience. Its site says it detects text from models such as ChatGPT, Claude and Gemini, and it highlights the sentences it flags. It also has a plagiarism checker, writing feedback and a feature called Writing Replay that shows how a document was written. A Chrome extension and Google Docs support are available, and it connects to Canvas and Google Classroom.
Who it fits: teachers who want to check a document and see which parts triggered the score, and students who want to check their own work before they submit it. Longer inputs give stronger results, and the vendor says accuracy is best on English prose.
2. Originality.ai
Originality.ai scores text as likely AI, likely original or mixed. It adds a plagiarism checker, readability and grammar checks, a fact-checking feature and a deep scan that suggests edits. You can use it through a web app, a Chrome extension, a Moodle plugin and an API.
Who it fits: publishers, agencies and content teams that want to check writer submissions at scale, with team roles. The site offers a limited number of free scans without an account, so you can test it on your own samples first.
3. Winston AI
Winston AI scores text on a scale from human to AI and shows a sentence-by-sentence colour map. It also has an image detector for pictures from tools such as Midjourney, a plagiarism checker and downloadable reports. Integrations include a Chrome extension, WordPress, Zapier and an API.
Who it fits: educators who need a report to attach to a case file, and publishers who want one tool for both text and images. The site offers a free trial with no credit card.
4. Pangram
Pangram detects text from models such as ChatGPT, Claude and Gemini and returns a prediction with a score. It also estimates how much of a document is AI-generated or AI-assisted and flags AI-edited segments. It has a plagiarism checker and an image detector. The free plan lets you run checks after signing up.
Who it fits: universities, publishers, recruiters and developers who use its API. Pangram publishes its own false-positive figures, but those are the vendor’s numbers. Test it on writing you know is human before you trust any figure.
5. Sapling
Sapling offers an AI detector that estimates the chance a text came from a model such as ChatGPT or Gemini. It gives an overall score and a per-sentence score, and it checks pasted text, PDF and Word files, web pages and selected text through a browser extension. The free version works without signing up, and there is a free API key for developers.
Who it fits: anyone who wants a fast check with no account, and developers who want to try detection in their own tools. The free check has a limit on length, so longer documents need a paid plan.
6. ZeroGPT
ZeroGPT is a free detector that highlights sentences it believes an AI wrote and shows an overall percentage. It also offers PDF reports, batch file uploads, a plagiarism checker, AI image and video detectors and an API. Some of its extra tools, such as paraphrasing and humanizing, sit on the other side of the problem, so judge the detector on its own.
Who it fits: quick, casual checks. Treat the percentage as a rough guide, and compare the result against a second tool before you act on it.
7. Copyleaks
Copyleaks began as a plagiarism checker and now sells AI detection alongside it. It supports many languages, works with learning management systems and offers browser and Google Docs integrations. Schools can adjust the sensitivity level, which matters because stricter settings flag more text.
Who it fits: schools, universities and businesses that want plagiarism and AI checks in one workflow. Review the sensitivity setting with your team so everyone reads a result the same way.
8. Turnitin
Turnitin is the plagiarism platform that many schools and universities already use, and it includes AI writing detection in its reports. You cannot buy it as a single user in the way you can with a web tool, because it is sold to institutions.
Who it fits: schools and universities that already have it. Its own representatives have said the detector can have a higher incidence of false positives when only a small share of a document is flagged, and it marks those cases with an asterisk. Read the guidance from your institution and talk to the student before drawing any conclusion.
9. Reality Defender
Reality Defender focuses on deepfakes rather than text. Its site lists detection for audio, video and image, plus detection of synthetic voices on calls. It is aimed at organisations such as banks, telecom companies, media groups and identity verification providers, and it offers a free API key to start.
Who it fits: security and fraud teams, newsrooms and platforms that need to check media, not essays. Most individuals will not need it.
10. Sightengine
Sightengine offers AI image detection, AI video detection and deepfake detection through an API. Its site names content moderation, online marketplaces and profile photo checks as use cases.
Who it fits: developers who run an app, a marketplace or a community and want to flag AI-generated images that users upload.
Other routes worth knowing
Some AI makers add invisible watermarks to what their tools produce. Google announced a SynthID Detector portal that checks whether content carries its SynthID watermark. It can only find content made with Google’s own AI tools, and it cannot tell you anything about text or images from other systems. Access was limited when it launched, so check Google’s page for the current status.
Developers can also find open detection models on the Hugging Face model hub. Those need technical skills to run, and the quality varies from one model to another, so test them against samples you know before you build on them.
How to use a detector responsibly
- Treat the score as a prompt for a conversation. If a writer or student gets flagged, ask them about the draft, the sources and how they wrote it.
- Use more than one tool. If two tools disagree, that tells you the evidence is weak.
- Look at version history. Google Docs and Word keep revision history. A document that grew in steps over days is strong evidence of human work.
- Set a clear policy first. Say in advance what counts as acceptable AI use, so the rule is about behaviour and not about a score.
- Test on your own samples. Run a few pieces you know are human and a few you know are AI through the tool, and see how it behaves on your kind of writing.
- Protect people from a bad flag. Never make a decision that harms someone on the strength of a detector result alone.
How to check content without a detector
You can often learn more by reading. AI text tends to be smooth and generic, repeats the same sentence shapes and avoids specifics such as names, dates, places and first-hand detail. Ask the author for sources and check that they exist. Ask for a draft or an outline. A person who wrote a piece can usually explain how and why they wrote it.
For images, look for the source. A real photo usually has an original uploader, an earlier appearance online or metadata. A reverse image search can show whether a picture has appeared before. For video and audio, look for the original publisher and check whether other credible outlets show the same footage.
If you run a site that accepts submissions
Blogs, communities and marketplaces face this problem at a different scale. You may receive guest posts, reviews or user profiles that a model wrote. A detector helps you triage, but a clear submission policy and a human review step do more. Ask contributors to disclose AI use, require a real author profile, and review anything that gets flagged before you reject it. For related reading, see our guides to the best AI models, the best AI bots and the best AI companion apps.
Frequently asked questions
What is the best AI detection tool?
There is no single best tool, because each one makes mistakes. GPTZero, Originality.ai, Winston AI, Pangram and Sapling are common choices for text, and Reality Defender and Sightengine cover deepfakes and images. Test two or three on writing you know is human before you rely on any of them.
Are AI detectors accurate?
Not reliably. They return a probability, and they can flag human writing as AI or miss edited AI text. Even the vendors warn against using a score as the final verdict.
Can an AI detector flag my own writing by mistake?
Yes. Plain, formal or heavily edited writing is more likely to be flagged, and so is writing by people who use English as a second language. Keep drafts and version history so you can show how your work developed.
Is there a free AI detector?
Yes. GPTZero, Sapling, ZeroGPT, Originality.ai and Pangram all have a free way to try them, with limits on length or the number of checks. Check each site for the current limits.
Can I use an AI detector to prove a student cheated?
No. A score is not proof. Use it to start a conversation, look at version history and the student’s other work, and follow your institution’s policy.
How do I detect deepfakes?
Use a deepfake detection service such as Reality Defender or Sightengine, and also check the source of the video or audio. Detectors help, but tracing where the file first appeared is often more reliable.
Does Google penalise AI-written content?
Google has said that it focuses on whether content is helpful and original, not on how it was made. Publishing low-value pages at scale is the risk, whoever or whatever writes them.
What happened to OpenAI’s AI text classifier?
OpenAI removed it in 2023 and cited its low rate of accuracy. It has not released a replacement as of the sources we checked.
Where to go next
Pick one text detector and one backup, run a few samples you know are human and a few you know are AI, and write down how each tool behaves. Then set a plain policy on AI use and a review step for flagged work. If you also run a site where members publish their own content, our guide to building an online community covers moderation and posting rules.