Last updated: October 10, 2026 · By Ned Walsch

AI Image Detector: Is This Image AI?

An AI image detector can’t give you a trustworthy “% AI” score, because no classifier works reliably across every generator: NIST found that detectors “tend to exhibit substantial error rates” on images from generators they were not trained on, especially after compression or resizing NIST AI 100-4, Reducing Risks Posed by Synthetic Content, Nov 2024, p. 28. What can be checked is the evidence inside the file. Drop an image below and this free tool reads it locally, with no upload: signed C2PA Content Credentials, the IPTC digital source type trainedAlgorithmicMedia (“Created using Generative AI”) IPTC NewsCodes: Digital Source Type (read 10 Oct 2026), Stable Diffusion and ComfyUI generation parameters, generator names in EXIF and comments, and missing camera data. Each finding is marked strong, moderate, weak or info and explained. If the file carries nothing, which is common because platforms strip metadata, use the links for a second opinion: Google opened its SynthID Detector to everyone on 7 October 2026 for content made with AI from Google or partners including OpenAI, NVIDIA and Kakao Google blog, Pushmeet Kohli, 7 Oct 2026, and classifier services give their own scores.

Nothing is uploaded. The image is read inside this browser tab and every check runs here. No request is made with the file or its contents. Links to other detectors only open those sites; you would upload there yourself, under their terms.

Drop an image here, paste it, or JPEG · PNG · WebP · GIF · HEIC/AVIF · TIFF — read locally, never uploaded

Evidence, not a verdict: no percentage is shown because no reliable one exists.

What does each check mean, at a glance?

Each row is a separate piece of evidence. Only the first can be cryptographically verified; everything else is metadata that anyone can edit or delete.

Check	What it reads	What a hit means	Limits
C2PA Content Credentials	Signed manifest: `digitalSourceType` on actions and ingredients, claim generator, software agent, signature, content hash	Strong if it declares `trainedAlgorithmicMedia` and the signature and hash verify	Often stripped on upload. The signer is shown but not checked against the C2PA trust list; use Content Credentials Verify for that.
IPTC digital source type (XMP)	`Iptc4xmpExt:DigitalSourceType` and the IPTC 2025.1 AI fields	Strong, explicit declaration of AI creation or AI editing	Unsigned: can be added, changed or removed with free tools.
Generation parameters	PNG text chunks (`parameters`, `prompt`, `workflow`), EXIF UserComment	Strong: prompt, sampler, seed and model settings from Stable Diffusion tools	Only present on files saved straight from those tools; easy to strip.
Generator names	EXIF Software, Make, Model, ImageDescription, XP tags, JPEG and GIF comments, IPTC captions	Moderate: a program or a person named an AI tool	Captions can mention AI without the image being generated.
Camera metadata	Make, model, lens, exposure, capture date	Weak either way	Platforms remove it from real photos; it can be copied into fakes.
Error level analysis (optional)	Pixel recompression differences	Possible local editing or pasting	A manipulation check, not an AI test: a fully generated image has no pasted-in region to find, so a clean map says nothing about AI.

Can you tell if an image is AI-generated?

Sometimes, with certainty; often, not at all. It depends on what survived between the generator and your screen.

When the file still carries its records, you can. A C2PA manifest or an IPTC label written by the generator states how the image was made. Meta labels images on Facebook, Instagram and Threads when it can “detect industry standard indicators that they are AI-generated”, using “the ‘AI generated’ information in the C2PA and IPTC technical standards” Meta, Nick Clegg, 6 Feb 2024.
Most shared copies have lost those records. NIST: “Many internet platforms strip at least some metadata from uploaded files to protect privacy” NIST AI 100-4, Reducing Risks Posed by Synthetic Content, Nov 2024, p. 20. OpenAI lists the same reason when its own tool finds nothing: metadata “was stripped during upload, download, editing, conversion, or sharing” OpenAI Help: Provenance signals in OpenAI-generated content (read 10 Oct 2026). An older IPTC test of 15 social media sites found that “ten sites removed at least some metadata” and only Behance kept and displayed it IPTC, 19 Jan 2016. A screenshot is a new file and carries none of the original’s metadata.
Unsigned metadata proves little either way. NIST: “anyone copying or editing the file can then manipulate or falsify the metadata”, and “metadata on its own is neither tamper-evident nor reliably attributable” NIST AI 100-4, Reducing Risks Posed by Synthetic Content, Nov 2024, pp. 17–18. A C2PA signature fixes that for the claims it covers, but NIST also notes that “anyone can sign a certificate themselves” NIST AI 100-4, Reducing Risks Posed by Synthetic Content, Nov 2024, p. 21, which is why the signer matters.
Classifiers guess from pixels, with real error rates. NIST: “detectors are often tied to and may only perform well on specific generators”, and “as soon as a new detection method is created, generation models improve, and adversaries learn new ways to avoid detection” NIST AI 100-4, Reducing Risks Posed by Synthetic Content, Nov 2024, p. 23.
Your eyes are not a reliable test either. NIST cites a 2024 study of audio, visual and audiovisual content that “found performance near chance” for people NIST AI 100-4, Reducing Risks Posed by Synthetic Content, Nov 2024, p. 23.

That is why this page shows evidence with its strength instead of a probability. A checker that prints “87% AI” for every file is hiding which of these situations you are in.

Is this image AI? How to check it in five steps

Get the best copy. Save the original file rather than a screenshot. Ask the sender for the file they received, or download it from the earliest post you can find.
Read the file with the checker above. A signed trainedAlgorithmicMedia declaration or Stable Diffusion parameters usually settle it. For the full manifest and certificate chain, open the same file in the C2PA Inspector.
Check for watermarks. Metadata is fragile; invisible watermarks are designed to survive edits. Try SynthID Detector for Google and partner models and OpenAI Verify for OpenAI images (both are uploads to those companies).
Find where it came from. A reverse image search across several engines often finds the earliest post, the account that made it, or a caption saying it was generated. Older copies may still carry metadata the shared copy lost.
Then weigh classifiers and visual details. A classifier score is one more opinion. For a full documented case (provenance, metadata, earliest copy, forensics, context and a conclusion), use the Image Verification Workbench; for edits to a real photo, Photo Forensics.

What is SynthID and can I check it?

SynthID is Google DeepMind’s watermarking tool: it “embeds digital watermarks directly into AI-generated images, audio, text or video”, imperceptible to people but detectable by SynthID’s technology, and Google says the image and video watermark is “designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression” Google DeepMind: SynthID (read 10 Oct 2026).

Can you check it as of 10 October 2026? Yes, for some generators.

SynthID Detector: on 7 October 2026 Google wrote, “Now, we’re expanding access to everyone – with the tool available globally in English starting today.” It checks “if an image, video, or audio file was made with AI from Google or our partners, including OpenAI, NVIDIA, Kakao, and soon, Apple.” Google had introduced an early version for media professionals the year before Google blog, Pushmeet Kohli, 7 Oct 2026.
Gemini, Search and Chrome: “Simply upload the image, video or audio clip to your chat, and ask if it’s been created or altered by Google AI”; they “will check for a SynthID watermark, and let you know if it finds one” Google DeepMind: SynthID (read 10 Oct 2026). Google says these built-in checks handle over 1 million requests a day Google blog, Pushmeet Kohli, 7 Oct 2026. Google added the Gemini app check in November 2025 and said images from Nano Banana Pro would also carry C2PA metadata Google blog: AI image verification in the Gemini app, 20 Nov 2025.

What it cannot tell you. A SynthID check answers one question: was this made or edited with a model that applies SynthID? Images from generators that don’t use it will show no watermark, so “no SynthID found” is not “real”. NIST also notes that “most current watermarking schemes have been consistently found vulnerable to removal” NIST AI 100-4, Reducing Risks Posed by Synthetic Content, Nov 2024, p. 13. This page does not check for SynthID; use Google’s tools above for that.

OpenAI runs a similar, narrower check. Its Verify page looks for a SynthID watermark or a trusted OpenAI C2PA manifest, and OpenAI says it is “not designed to detect content generated by other AI services” OpenAI Help: Provenance signals in OpenAI-generated content (read 10 Oct 2026); it also says uploaded files “are not stored unless legally required and are not used to train our models” OpenAI: Verify OpenAI-generated content (read 10 Oct 2026).

What do C2PA and IPTC digital source types say?

Both standards use the same vocabulary. IPTC maintains a controlled list of digital source types, each identified by a URI such as http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia IPTC NewsCodes: Digital Source Type (read 10 Oct 2026):

Code	IPTC label	IPTC definition	How this checker reports it
`trainedAlgorithmicMedia`	Created using Generative AI	“Digital media created algorithmically using an Artificial Intelligence model trained on captured content”	Strong AI declaration
`compositeWithTrainedAlgorithmicMedia`	Edited using Generative AI	“Augmentation, correction or enhancement using a Generative AI model, such as with inpainting or outpainting operations”	AI editing of part of the image
`compositeSynthetic`	Composite including generative AI elements	“Mix or composite of several elements, at least one of which is Generative AI”	AI element in a composite
`algorithmicMedia`	Pure algorithmic media	“Media created purely by an algorithm not based on any sampled training data”	Not generative AI (reported as info)
`algorithmicallyEnhanced`	Algorithmically-altered media	“Modification or correction by algorithm without changing the main content of the media”	Neutral
`digitalCapture`	Digital capture sampled from real life	“The media was captured from a real-life source using a digital camera or digital recording device”	Capture declaration
`computationalCapture`	Multi-frame computational capture sampled from real life	Multiple frames captured from a real-life source and merged, for example HDR	Capture declaration
`composite`	Composite of elements	“Mix or composite of several elements, any of which may or may not be generative AI”	Neutral: says nothing about AI

In XMP (unsigned)

The IPTC Photo Metadata Standard stores the value in Iptc4xmpExt:DigitalSourceType, defined as “the type of the source of this digital image”. Version 2025.1 added four AI fields: Iptc4xmpExt:AISystemUsed (“the AI engine and/or the model name used to generate this image”), AISystemVersionUsed, AIPromptInformation and AIPromptWriterName IPTC Photo Metadata Standard 2025.1, 26 Nov 2025. The checker reads all of them from the raw XMP packet, including extended XMP split across several JPEG segments.

In C2PA Content Credentials (signed)

C2PA added a digitalSourceType field to actions in version 1.2 (October 2022), and since version 2.1 a standard manifest must contain either a c2pa.created or a c2pa.opened action; version 2.4 (April 2026) also allows a digitalSourceType on ingredients that arrive without their own manifest C2PA Technical Specification 2.4, Apr 2026. The Content Authenticity Initiative’s SDK documentation says that to mark generative AI origin you “use the c2pa.created action with digitalSourceType” set to one of the IPTC AI values Content Authenticity Initiative SDK docs: Actions (read 10 Oct 2026). The checker looks at the active manifest’s actions first, then at ingredients and earlier manifests, and reports the claim generator and software agent names alongside.

The signature tells you the claim was made by the holder of the signing certificate and that the file has not changed since; it does not tell you the claim is true, and the C2PA specification says its specifications “SHOULD NOT provide value judgments” C2PA Technical Specification 2.4, Apr 2026, §1.2. The checker verifies the signature and content hash with the site’s C2PA engine but does not check the signer against the C2PA trust list.

Which Stable Diffusion and ComfyUI traces does it find?

Local image-generation tools often save their settings inside the image so it can be reproduced. The checker looks for both common formats:

AUTOMATIC1111 (stable-diffusion-webui): its source code writes the generation info into a PNG text chunk whose key defaults to parameters, and for JPEG and WebP into the EXIF UserComment tag; GIFs get it as a comment AUTOMATIC1111 stable-diffusion-webui source, modules/images.py (read 10 Oct 2026). The text is the prompt, then a line such as Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: …, Size: …, Model hash: …; the checker matches that pattern wherever it appears.
ComfyUI: PNG output stores metadata in “PNG tEXt chunks” with the standard keys prompt and workflow, both JSON; animated WebP stores strings like workflow:{JSON} in EXIF tags. Saving can be turned off with --disable-metadata ComfyUI docs: Workflow metadata (read 10 Oct 2026). The checker recognises the node graph (KSampler, checkpoint loader, sampler settings) and also decompresses zTXt and compressed iTXt chunks.
Everything else: other text chunks, JPEG COM segments, GIF comments, EXIF Software, Make, Model, ImageDescription and XP tags, and IPTC captions are searched for the names of generators and AI front-ends. A match is shown with the surrounding text so you can see whether it records the tool or is just a caption mentioning AI.

These traces are easy to lose: ComfyUI can be told not to write them, and a screenshot, a re-encode or a platform upload that strips metadata removes them. Their absence means nothing.

Free AI image detectors compared

This page reads the file; the services below either look for a watermark from specific generators or run a classifier on the pixels. Classifiers can flag an AI image whose metadata is gone, which this page cannot do; their scores are the vendors’ models, not verified facts. Everything in the table is what each provider says on its own pages, read on 10 October 2026: vendor claims, not our tests. All of them require uploading your image to the provider.

Service	Type	What the provider says it does (vendor claim)	Free use, as stated	Source
SynthID Detector (Google)	Watermark detector	Checks “if an image, video, or audio file was made with AI from Google or our partners, including OpenAI, NVIDIA, Kakao, and soon, Apple”; available globally in English from 7 Oct 2026	Opened to “everyone”; price not stated in the announcement	Google blog, Pushmeet Kohli, 7 Oct 2026
OpenAI Verify	Watermark + C2PA check	Looks for SynthID or a trusted OpenAI C2PA manifest; “not designed to detect content generated by other AI services”	Not stated; files “are not stored unless legally required”	OpenAI: Verify OpenAI-generated content (read 10 Oct 2026) OpenAI Help
Content Credentials Verify	C2PA reader	“Content Credentials provide deeper transparency into how content was created or edited.” Reads C2PA only	Not stated on the page	Content Credentials Verify (read 10 Oct 2026)
Hive	Classifier	“Hive returns confidence scores for whether media is AI-generated and, when available, the likely model behind it.” No accuracy figure on the page	Lists “free AI detection tools”: Hive Detect (uploads and URLs) and a Chrome extension; limits not stated	Hive: AI-generated content detection (read 10 Oct 2026)
Sightengine	Classifier (API)	Claims “the highest accuracy among the tools tested” in a study by the University of Rochester and University of Kansas on 80,000 images; also says “no detector is 100% reliable”	Free tier of “2,000 operations / month (max 500/day)”; the pricing page does not say how many operations one AI-image check uses; paid from $29/month	Sightengine: Detect AI-generated images (read 10 Oct 2026) Sightengine pricing (read 10 Oct 2026)
AI or Not	Classifier	Claims “98.9% accuracy” for images, text, video and deepfakes; the page gives no test method	“20 AI image checks” and “$5 in free credits” (renewal of the image checks not stated); Pro from $5/month	AI or Not home page (read 10 Oct 2026) AI or Not pricing (read 10 Oct 2026)
Illuminarty	Classifier	“Find out the probability of AI generation for a given image”; paid plans add localized detection and model identification; no accuracy figures, and “we make no representations or warranties of any kind”	“We will always provide the basic AI detection functionalities for free”; Basic $10/month	Illuminarty home page (read 10 Oct 2026)

Where they beat this page: a classifier can say something about a screenshot or a stripped re-upload; a watermark detector can confirm an image from a participating generator even after metadata is gone. Where this page helps: it shows you the declaration itself (who signed it, which tool, which prompt), works offline on files you can’t upload, and never turns a guess into a number. Accuracy figures from vendors are measured on their chosen test sets; NIST’s review warns that cross-generator error rates are substantial NIST AI 100-4, Reducing Risks Posed by Synthetic Content, Nov 2024, p. 28.

Visual tell-tales, and why not to trust them

Looking closely is still worth doing, as long as you treat what you see as a reason to dig further, not a conclusion. NIST lists perceptible cues such as “reflections or shadows that do not match lighting” and “visual inconsistencies (e.g., earrings, eye colors, or reflections in eyes)” NIST AI 100-4, Reducing Risks Posed by Synthetic Content, Nov 2024, p. 24. Things to look at:

Shadows and reflections: do they agree with one light source, and do mirrors, windows and eyes reflect the scene?
Pairs and repeats: earrings, eyes, glasses, buttons, fence posts and windows that should match.
Text and signs: is lettering legible, spelled consistently and attached to real surfaces?
Hands, teeth, hair edges and where objects touch or overlap.
Backgrounds: people, buildings or patterns that blur into each other.
Context: does the place, weather, clothing and date fit the claim? The Image Verification Workbench walks through this.

Why not to trust them: NIST notes that “humans may find detection more difficult as synthetic content generation continues to increase in sophistication”, and cites a 2024 study where people performed near chance NIST AI 100-4, Reducing Risks Posed by Synthetic Content, Nov 2024, p. 23. Real photos have odd hands, motion blur and lens artefacts too. A clean-looking image is not evidence of a real one.

How was this tool tested?

On 10 October 2026, files were generated for each case and loaded into the page in headless Chromium (Playwright) through the file input, with all non-local network requests blocked to confirm nothing is sent:

Signed C2PA, AI declared: a JPEG signed with the open-source c2pa-python library (0.38.0) and a throwaway test certificate, with a c2pa.created action declaring trainedAlgorithmicMedia. Result: signature valid, content hash matches, “Signed Content Credentials declare this image AI-generated”. The same file with two bytes of image data changed: “Content Credentials declare AI, but an integrity check failed”.
Signed C2PA, capture declared: a camera-tagged JPEG signed with digitalCapture. Result: capture declaration reported as weak because the signer is a test certificate; no AI evidence.
Public C2PA test images (C2PA SDK fixtures with algorithmicMedia, a two-manifest file, a remote-manifest file): reported as info, “not the generative-AI code”, and “stored remotely (not fetched)”.
Stable Diffusion and ComfyUI: a PNG with an AUTOMATIC1111-style parameters tEXt chunk, a JPEG with the same text in EXIF UserComment (UTF-16), and a PNG with ComfyUI prompt (tEXt) and workflow (compressed zTXt) chunks. All three: strong, “made with generative AI”.
XMP: a JPEG with Iptc4xmpExt:DigitalSourceType = trainedAlgorithmicMedia (strong AI), a WebP with compositeWithTrainedAlgorithmicMedia (strong, AI editing) and a JPEG with digitalCapture (weak capture declaration).
Camera and bare files: a JPEG with Canon make, model, lens and exposure tags (weak, points to a capture; “no AI evidence found”), and JPEG and PNG files with no metadata at all (only the weak “no camera metadata” signal). HEIC and AVIF samples were read without errors.
Safety: a JPEG comment containing an HTML img tag with an onerror handler was displayed as text, not executed; the JSON report downloaded with a sanitised file name; the drop zone opens the file picker with Enter; error level analysis rendered locally; no request left the page during any check; no horizontal scrolling at 375 px.

Sources

Frequently asked questions

Is there a free AI image detector that is reliable?

No detector is reliable on every image. NIST reported in November 2024 that detectors are often tied to specific generators and show substantial error rates on images from other generators, especially after compression or resizing. Even vendors say so: Sightengine states that no detector is 100% reliable. The most dependable evidence is what the file itself records, such as a signed C2PA Content Credential or an IPTC label saying the image was created with generative AI. This free checker reads exactly that, and links to classifiers for a second opinion.

How can I tell if an image is AI-generated?

Combine several checks. First read the file: Content Credentials, IPTC digital source type, Stable Diffusion or ComfyUI parameters and camera metadata, which this page does locally. Then check for invisible watermarks with Google's SynthID Detector or OpenAI's Verify tool, run a reverse image search to find the earliest copy and who posted it, and only then look at a classifier score and visual details. One signal alone rarely settles it.

Does this AI image detector upload my photo?

No. The file is read inside your browser tab and every check runs there; nothing is sent to Max Intel or anyone else. The links to SynthID Detector, OpenAI Verify, Hive, Sightengine and other services only open those sites, and you would upload the image there yourself under their terms.

Can Google tell if an image is AI-generated?

Google can tell whether an image carries its SynthID watermark. On 7 October 2026 Google opened SynthID Detector to everyone, globally in English, to check images, video and audio made with AI from Google or its partners, which Google names as OpenAI, NVIDIA, Kakao and, soon, Apple. You can also upload an image to the Gemini app, Search or Chrome and ask whether Google AI made it. A missing watermark does not prove an image is real, because other generators do not use SynthID.

Can ChatGPT or OpenAI tell if an image is AI-generated?

OpenAI's Verify page at openai.com/verify checks an uploaded image for a SynthID watermark or a trusted OpenAI C2PA manifest, so it can confirm images made with ChatGPT, Codex or the OpenAI API that still carry those signals. OpenAI says the tool is not designed to detect content generated by other AI services, and that finding no signal does not prove the image was not made with OpenAI tools.

Why does the checker say no AI evidence found when the image looks AI-generated?

Because most AI images online have lost their metadata. Social networks, messaging apps, screenshots and format conversions re-encode the file and drop Content Credentials, IPTC labels and generator parameters. NIST notes that many internet platforms strip at least some metadata from uploaded files. When the file carries nothing, the remaining options are watermark detectors, classifiers and finding the original through a reverse image search.

What does trainedAlgorithmicMedia mean in image metadata?

It is an IPTC digital source type code, written as http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia, whose label is Created using Generative AI. The related code compositeWithTrainedAlgorithmicMedia means Edited using Generative AI, for example inpainting or outpainting. Tools write these codes in XMP metadata or in the actions of a C2PA Content Credential.

Does missing EXIF data mean an image is AI-generated?

No. A generator has no camera, so it has no real camera values to record, but screenshots and many apps and platforms that re-encode uploads also leave none; NIST notes that many internet platforms strip at least some metadata from uploaded files. Missing EXIF is a weak hint at most. Camera tags that are present are weak evidence the other way, since EXIF can be copied or typed in with free tools.

Can AI image detection be fooled?

Yes, every method has limits. Metadata and Content Credentials can be removed by re-saving or screenshotting the image. NIST found that most current watermarking schemes have been consistently found vulnerable to removal, and that classifiers lose accuracy after post-processing and on new generators. That is why this page reports each piece of evidence and its weakness instead of a single score.