AI detector free: what text and image checkers catch and how far to trust them

Plans, prices and features checked on the official pages on October 17, 2026.

Yes, you can check text or images for AI for free: GPTZero, Grammarly and ZeroGPT scan pasted text, and ImageDecoder and Google’s SynthID Detector examine images. None of them can prove anything. Every maker says so on its own page, and a score is best treated as a reason to look closer.

This guide covers what each tool checks, the free limits stated on the official pages, and what the schools’ detectors say about their own reliability.

How AI detectors work and why they can be wrong

A text detector estimates how likely a passage is to follow the patterns of machine-written language. Grammarly describes its tool as splitting the text into sections and checking each one for language patterns linked to AI writing. It then reports the share of the text that looks machine-generated.

That design explains the errors. Grammarly’s page says no AI detector is 100% accurate, that the tool cannot conclusively determine whether AI was used, that human-written text can be flagged, and that lightly edited AI text may go undetected. GPTZero’s page adds that longer inputs give stronger results and that a result should not be the final verdict.

Image detectors work differently. Some look for visual artifacts, but others read provenance data: metadata, Content Credentials (C2PA) and invisible watermarks. A missing label proves nothing, because most photos never carried one.

Our pick for checking images: ImageDecoder

Our pick for checking images is ImageDecoder. According to its own page, it is a web tool that weighs several kinds of evidence together: AI-related signals in the picture, embedded metadata, C2PA provenance records and traces of editing.

The sample report on the page shows an AI signal level, a metadata status, a content credentials status (for example “Not found”) and an editing clues field (for example “Exported image”). The page says it reports signals and likelihoods rather than a certainty, and that missing metadata is never held against an image.

A free preview is available with no account. Accepted files are JPG, PNG and WebP up to 15 MB, or a direct image URL (not a webpage URL). The page states that images are kept private and deleted once analysis is done, and that results are signals, not proof. It lists no prices or usage limits beyond file size and format. ImageDecoder is made by the team behind Aiholics.

Two other routes exist for images:

• Google SynthID Detector. Google’s page says the verification portal, launched in 2025, is being opened to everyone. You upload an image, video or audio file to see whether Google AI or partners such as OpenAI, NVIDIA and Kakao made it, with Apple listed as coming soon. It only recognizes content carrying those watermarks.
• Content Credentials. C2PA describes Content Credentials as a framework that lets people trace where content came from and how it has changed, much like a nutrition label. They are present only if the creating tool added them.

Prices checked on October 17, 2026.

Free text detectors compared: GPTZero, Grammarly and ZeroGPT

GPTZero. The scan box on its homepage shows a limit of 10,000 characters, and its FAQ says longer scans need a free account. GPTZero claims 99% accuracy based on its own testing and benchmarks. On the RAID benchmark it reports detecting 95.7% of AI texts with 1% of human texts wrongly flagged. These are the company’s figures, not independent audits.

Grammarly. The free detector gives a percentage score. Grammarly says the fuller AI Detector agent, which highlights which parts may read as AI-generated and offers rewrites and citations, requires Grammarly Pro. It claims 99% detection accuracy and a #1 ranking on RAID’s benchmark, and also says the model was trained on texts including AI-generated text created before 2021.

ZeroGPT. Its text box is capped at 15,000 characters. It claims 98.4% accuracy, a false positive rate under 1% and coverage of 35+ LLM families, with no testing method or third-party benchmark given on the page.

Other names people search for, such as Scribbr, QuillBot and Copyleaks, also offer free checkers. Their official pages blocked automated access while this guide was prepared, so check limits and prices on their sites before relying on any figure you see elsewhere.

What Turnitin and school detectors say about their own accuracy

Turnitin’s detector is sold to institutions, so the number that matters is the one your school licenses. Turnitin’s chief product officer wrote in June 2023 that the document-level false positive rate is under 1% for documents with 20% or more AI writing. The sentence-level rate is about 4%, meaning a highlighted sentence has roughly a 4% chance of being human-written.

Turnitin also says errors are more common in mixed documents, especially at the transitions, and that about 54% of wrongly flagged sentences sit next to actual AI writing. It states there is no right or target score, and advises treating highlights as starting points for a conversation, not conclusions. These are 2023 figures from the company, so ask your institution which version it uses.

No free tool replicates Turnitin. Each vendor trains its own model, so scores differ between tools. For more context on classroom use, read how AI cheating is surging at UK universities.

Plans and pricing: what is free and what needs an account

Prices checked on October 17, 2026. Where a cell says no price, the official page did not give one.

How to check a text or an image step by step

1. Pick a tool for the format: a text detector for writing, an image checker for pictures.
2. For text, paste at least a few full paragraphs. Short snippets give weak results according to GPTZero.
3. For images, upload the original file if you can. Re-saved or screenshotted copies lose metadata and credentials.
4. Run a second tool and compare. Disagreement between detectors is common and shows how uncertain the answer is.
5. Look for evidence outside the tool: drafts, version history, the source of the image, who posted it first.
6. Treat the result as a likelihood, never as a verdict about a person.

Why scores mislead: limits worth knowing

• Mixed text is hardest. Turnitin itself reports more false positives in documents blending human and AI writing.
• Edits change the score. Grammarly says lightly edited AI text may go undetected, which is why paraphrasing tools exist.
• Vendor accuracy figures are self-reported. The 95% to 99% numbers above come from the makers, using benchmarks they chose.
• Absence of a signal is not proof. No watermark, no credentials and no metadata only means nothing was found.
• Your data goes to a server. Check whether a tool stores what you upload. ImageDecoder says it deletes images after analysis.

The stakes rise as machine-written text grows; see more articles are written by AI than humans. Even OpenAI debated whether to release its own detector, as covered in OpenAI’s secret ChatGPT detector.

Common questions