Free, on your device, no upload
AI image detector: is this image AI?
Drop an image to read what it records about itself: signed Content Credentials that name the AI tool, such as OpenAI's gpt-image or Adobe Firefly, AI labels, and the prompts and settings generators leave behind. Then check it for an invisible SynthID watermark.
- Content Credentials verified
- AI labels, prompts and settings
- Where to check SynthID
Checking
Check result
SynthID check for ChatGPT and Google images
This checker reads what the file records on your device. It cannot see SynthID, the invisible watermark some AI tools hide in the pixels. The watermark is built to survive cropping, filters and compression. Not sure where an image came from? To check the watermark, start with Google's SynthID Detector, which looks for the SynthID that Google, OpenAI, NVIDIA and Kakao add.
Made with ChatGPT or OpenAI?
Upload the file to OpenAI's verify page, which reads OpenAI's SynthID watermark and its Content Credentials. Google's SynthID Detector reads OpenAI's watermark too.
OpenAI's verify pageMade with Google AI?
Images from Gemini, Imagen and Nano Banana carry Google's SynthID. Google's SynthID Detector reads it and is open to anyone who accepts Google's terms, within a daily limit. Gemini and Google Lens also read it if you ask "Is this AI-generated?".
SynthID Detector Ask in GeminiOther AI tools
Midjourney, Stable Diffusion and FLUX are not SynthID partners. For their images, the labels, prompts and settings this checker reads are the signs to look for.
On Google's detector, "Watermark detected" means a SynthID-enabled model from Google or a partner likely made or edited the image, and "Watermark not detected" does not rule out AI. Read how SynthID works and how to read a result.
What no file check can see
- Invisible watermarks. Reading SynthID takes a watermark check, such as Google's SynthID Detector or OpenAI's own check, so this page hands that job to the checks above.
- Removed metadata. A screenshot, most social apps and many converters strip every label above. "No AI labels found" proves nothing either way.
- The pixels. We do not guess from how a picture looks, so you will never see a made-up "97% AI" score here.
How to check if an image is AI-generated
Five steps, starting with the evidence inside the file. Our guide on how to tell if an image is AI-generated explains each one, with sources.
Use the original file
Labels live inside the file as it was made or downloaded. Screenshots and most social apps strip them, so ask for the original when you can.
Drop it into the checker
It reads Content Credentials, AI labels, generator settings, software names and camera details in your browser. Nothing is uploaded.
Read the verdict and the evidence
Content Credentials from a signer on the public trust lists are the strongest evidence, AI labels and generator settings are strong hints, and software names, typed-in AI fields and camera details are only hints.
Check the invisible watermark
Google's SynthID Detector reads the SynthID watermark from Google and partners such as OpenAI. For ChatGPT or OpenAI images, you can also upload the file to OpenAI's verify page. For Google images, you can also ask Gemini.
If nothing turns up, check the source
Find where the image first appeared with a reverse image search, and look at who posted it and when.
Checking a ChatGPT or OpenAI image
OpenAI marks its images in two ways. Its help center says pictures generated with ChatGPT, Codex or the OpenAI API carry both Content Credentials and an invisible SynthID watermark, except API outputs from models it plans to retire soon, such as gpt-image-1, gpt-image-1.5 and sora-2. This checker reads the first mark on your device. OpenAI's own tools read both.
The October 2026 picture we checked names the model, not the app. We made it in ChatGPT, and its record was signed by OpenAI OpCo, LLC, written by "OpenAI Media Service API" and created with "gpt-image 2.0", which OpenAI's model page describes as its image generation model. From a record like that, the checker can tell you OpenAI made a picture, but not whether it came from ChatGPT or from an app using the OpenAI API, and it says so in the result. When a trusted record names ChatGPT itself, the checker shows that instead.
Here is what you may see when you drop in a picture you think came from ChatGPT:
| What the checker shows | What it means |
|---|---|
| Made with AI: OpenAI gpt-image 2.0, signature trusted, signed by OpenAI OpCo, LLC | The file still carries OpenAI's signed record and has not changed since. This is what our ChatGPT test picture showed in October 2026, along with a note that OpenAI added an invisible watermark. |
| Record claims AI (image changed after signing): OpenAI gpt-image 2.0 | OpenAI signed the record, but the picture changed afterwards, so the record describes the original rather than this file. When we changed a single byte of image data in our ChatGPT test picture, the checker showed exactly this. |
| Record claims AI (unverified signer), with a certificate that names OpenAI | The certificate is not on the public C2PA trust lists. Anyone can put OpenAI's name in a certificate of their own, as we did in a test file, so treat the record as unconfirmed and check the file on OpenAI's verify page. |
| No AI labels found | Not a clean bill of health. When we saved the same ChatGPT picture again as a JPG, the record was gone and the checker showed this. OpenAI lists stripped metadata, a degraded watermark, an unsupported model or export path, and pictures made before these signals existed as reasons its own images can test negative. Check the file on OpenAI's verify page next. |
OpenAI's verify page takes PNG, JPG and WebP images and looks for its Content Credentials and its SynthID watermark. OpenAI advises against cropping or converting an image before you check it, so upload the file as you got it. If your copy is HEIC, AVIF, GIF, TIFF or BMP, use Google's SynthID Detector instead, which accepts those formats. The same check is available as OpenAI's Content Provenance API, which OpenAI says is currently free within usage limits. OpenAI is clear about the limits: its tools are not designed to detect pictures from other AI models, a positive result does not say how much of the picture the model made, and no result shows the prompt or who wrote it.
Sources: Provenance signals in OpenAI-generated content (OpenAI Help Center), Verify OpenAI-generated content (OpenAI), GPT Image 2 model page (OpenAI), SynthID Detector (Google).
What the results mean
| Result | What the file says | How much it proves |
|---|---|---|
| Made with AI | Trusted Content Credentials, an IPTC AI label, a generator's prompt and settings, or China's AIGC label with the value 1 say an AI tool created the image. | Trusted Content Credentials are close to proof. A label or settings are strong evidence, but they can be copied into another file. |
| Edited with AI | Generative AI changed or added part of an existing image. | As above, for the part that was changed. |
| Made with AI, signed by a company | Trusted Content Credentials say generative AI made the image but do not name the tool, so the checker names the company that signed them. | As strong as the signature. |
| Likely made with AI | Only a software field, such as EXIF Software or XMP CreatorTool, or the IPTC "AI system used" field names an AI tool, and no AI label backs it up. | A hint. People and apps type these fields in, and nothing signs them. |
| Possibly made with AI | China's AIGC label marks the file as possibly (2) or suspected (3) AI-generated, rather than AI-generated (1). | A hint. The label itself says it is not sure. |
| Record claims AI, unverified signer | Content Credentials say AI made the image, but the certificate that signed them is not on the public C2PA trust lists. | A claim, not proof. Anyone can make a certificate with any company name in it, so the checker does not say who signed it, only whose name the certificate carries. |
| Record claims AI, fails its checks | Content Credentials say AI made the image, but the file or the record changed after signing, or the certificate is not valid. | A claim, not proof. The summary says which check failed. |
| Captured by a camera | Trusted Content Credentials from a camera or phone say it captured the image, and record no AI changes. From a signer we cannot verify, you see "Record claims a camera capture" instead. | Strong for the capture. Edits made later in apps that do not write credentials are not covered. |
| Content Credentials, no AI recorded | Someone signed a history for the file that has no AI step. | Depends on the signer. Trusted means a company on the public C2PA trust lists signed it. |
| Content Credentials are stored online | The file only links to a record kept on a server, which the C2PA standard allows. | Nothing yet. We do not fetch the record, so read it with a C2PA tool that looks records up online. |
| No AI labels found | Nothing in the file says either way. | Not evidence that the photo is real. Screenshots and most apps remove labels, so check the invisible watermark and the source too. |
What "trusted" and "valid" mean, and what neither proves: valid vs trusted Content Credentials.
What it found in real test files
We ran the checker in Chrome on public sample files from the Content Authenticity Initiative, the C2PA and the IPTC, on a picture we made in ChatGPT, and on files we built ourselves with known labels, including records signed with a certificate we made up. These are the results it gave in October 2026.
| Test file | Where it came from | What the checker said |
|---|---|---|
| ChatGPT picture, as downloaded | Made by us in ChatGPT | Made with AI: OpenAI gpt-image 2.0. Signature trusted, signed by OpenAI OpCo, LLC, with an invisible watermark recorded. |
| The same picture with one byte changed | Made by us | Record claims AI (image changed after signing): OpenAI gpt-image 2.0. Still signed by OpenAI OpCo, LLC, but the image no longer matches the record. |
| The same picture saved again as a JPG | Made by us | No AI labels found, with the reminder that this proves nothing. |
| Adobe Firefly text-to-image picture from 2023 | Content Authenticity Initiative test files | Made with AI: Adobe Firefly 1.0.0. Signature trusted, signed by Adobe Inc. This early record has no IPTC AI label, so the checker reads Adobe's own generative AI entry instead. |
| Photo edited with Generative Fill in Photoshop | Content Authenticity Initiative test files | Edited with AI: Adobe Firefly. Signature trusted, signed by Adobe Inc., written by Adobe Photoshop 24.7.0. |
| Firefly picture, then edited again with Firefly | Content Authenticity Initiative test files | Made with AI: Adobe Firefly, with the later AI edit listed too. |
| Test record that claims a Firefly edit, signed with a test certificate | Content Authenticity Initiative test files | Record claims AI edits (unverified signer): Adobe Firefly. The summary says to read it as a claim, not proof. |
| PNG whose record names OpenAI and gpt-image 2.0, signed with a certificate we made that says "OpenAI OpCo, LLC" | Made by us | Record claims AI (unverified signer): OpenAI gpt-image 2.0. The checker says only that the certificate names OpenAI OpCo, LLC, and that anyone can make such a certificate. |
| Nikon Z 9 photo from a 2022 demo | C2PA public test files | Content Credentials found, no AI recorded. The Nikon certificate had expired and is not on the trust lists, so the record cannot be confirmed. |
| Test picture changed after it was signed | C2PA public test files | Content Credentials found, no AI recorded, with the note that the image was changed after it was signed. |
| Photoshop picture whose credentials were saved to Adobe's cloud | Content Authenticity Initiative test files | Content Credentials are stored online, with the link. We do not fetch it. |
| PNG with Stable Diffusion settings | Made by us | Made with AI: Stable Diffusion, with the prompt, negative prompt, sampler, seed and model shown. |
| PNG with a ComfyUI workflow | Made by us | Made with AI: ComfyUI, with the prompt, negative prompt and model file. |
| JPG, WebP and AVIF with only the IPTC AI label | Made by us | Made with AI in all three formats, with the credit line we typed in. |
| JPG with the IPTC 2025.1 AI system and prompt fields | Made by us | Made with AI, naming the AI system we typed in, with its version and the prompt. |
| JPG whose only clue is "Adobe Firefly" in XMP CreatorTool | Made by us | Likely made with AI: Adobe Firefly, marked as a hint because anyone can edit that field. |
| IPTC's 2025.1 reference photo, with the AI system fields filled in and a "rendered by software" source label | IPTC | No AI labels found. The AI system and prompt are shown, but not counted as AI, because the file's own label says otherwise. |
| PNG with China's AIGC label set to 2, "possibly AI-generated" | Made by us | Possibly made with AI, with the AI service and platform codes from the label. |
| PNG whose "parameters" text holds camera settings, not a prompt | Made by us | No AI labels found. The text is shown, but not counted as AI. |
An AI image detector that reads evidence, not pixels
There are two kinds of AI image detector. One kind looks at the pixels and gives a score, such as "87% likely AI". The other kind, which this page is, reads what the image file itself records about how it was made. The first kind always answers, and is sometimes wrong. The second kind answers only when there is something to read, and then it can show you what it found.
So the results here fall into three groups. Made or edited with AI: the file carries a signed record, a label or generator settings that say so, and the page names the tool where the file does. A camera or an editor signed it: the file has Content Credentials that show another origin. Nothing found: the file holds no record either way, which is common after a screenshot or a share through a social app, and proves nothing. In that case the page sends you to the SynthID checks and to a reverse image search instead of guessing.
Use it to check a photo before you share it, to see whether a picture in the news or in a listing is a real photograph, to confirm that your own AI images are labeled as they should be, or to read the prompt a generator saved in a file.
How accurate is an AI image detector?
It depends on what the checker looks at. Some evidence can prove where an image came from, and some can only guess. This checker reads the first kind and tells you plainly when a file holds none of it.
| Kind of check | What it can prove | Where it falls short |
|---|---|---|
| Signed Content Credentials | Who signed the file's history and which tool made it. A valid signature from a trusted company is close to proof for that file. | Easy to lose: OpenAI notes that a screenshot or a file conversion can remove them. A missing record proves nothing. |
| Invisible watermark (SynthID) | That an AI from Google or a partner made or edited the image, when Google's or OpenAI's check finds the watermark. | Silent for tools that do not add it, such as Midjourney, Stable Diffusion and FLUX. Heavy editing can weaken it, and no file check can read it. |
| AI labels and generator settings | Strong hints, such as an IPTC "digital source type" label or a Stable Diffusion prompt saved in the file. | Not signed, so they can be copied into another file or stripped out. |
| Pixel-guessing detectors | A probability, not proof. OpenAI's own classifier found about 98% of DALL-E 3 images and wrongly tagged under 0.5% of real ones in its internal tests. | The same classifier flagged only 5 to 10% of images from other AI models. Detectors like this can be wrong both ways and can be evaded. |
| Visual clues | Hints only. Google's Gemini help lists garbled text, errors in hands or teeth, odd shadows and repeating patterns. | Real photos can show odd details too, so a clue is never proof on its own. |
That is why this checker shows no percentage. A score such as "97% AI" from a pixel detector is a guess that can be wrong in either direction, while a signed record or a watermark found by a SynthID check is evidence you can point to.
Sources: Understanding the source of what we see and hear online (OpenAI, May 2024), Provenance signals in OpenAI-generated content (OpenAI Help Center), Verify AI-generated images (Gemini Apps Help), SynthID Detector (Google), Robustness of AI-Image Detectors: Fundamental Limits and Practical Attacks (Saberi and others, arXiv).
Which AI tools leave which marks
| AI tool | Content Credentials | Invisible watermark | Visible mark or label |
|---|---|---|---|
| ChatGPT, Codex and the OpenAI API | Yes, signed by OpenAI | SynthID, since May 2026 | None |
| Google Gemini and Nano Banana | Yes, on Nano Banana Pro and newer models | SynthID | Gemini sparkle on free and AI Pro plans |
| Adobe Firefly | Yes, when every pixel comes from Firefly | Not documented | None |
| Microsoft Designer | Yes, on images made or edited with its AI features | Not documented | None documented |
| Bing Image Creator | Announced by Microsoft in 2023 | Not documented | Optional: a setting adds a visible watermark |
| Samsung Galaxy AI, Galaxy S25 series | Yes, for content created and edited with generative AI | Not documented | A visible watermark on Generative Edit results |
| Meta AI | Not documented | Meta's own, no public checker | "Imagined with AI" on photorealistic images |
| Stable Diffusion apps and ComfyUI | No | None you can check | None, but the prompt and settings often stay in the file |
| AI uncrop on uncrop.si | No | None you can check | The IPTC "composite with trained algorithmic media" label in the file's XMP metadata, and "ai-uncropped" in the file name |
From each company's own documentation, checked October 2026: OpenAI, Microsoft Designer, Microsoft via IPTC (2023), Bing Image Creator help, Samsung and Samsung's Generative Edit note. Google, Adobe and Meta sources are in how to tell if an image is AI-generated.
What this checker reads
Everything below is read from the file on your device, without any AI model. The checker names the tool or company only when the file itself names it.
| Signal | Where it lives in the file | What you see |
|---|---|---|
| Content Credentials (C2PA) | A signed record inside the image file | The AI tool and its version, the app that wrote the record, the signer and the date, every recorded edit, and whether the signature is trusted, valid or broken. The signature is checked against the public C2PA trust lists with the open-source reader from the Content Authenticity Initiative. |
| Content Credentials kept online | A link in the XMP metadata (dcterms:provenance) | The address of the record. The C2PA standard lets a file carry only this link, and the Content Authenticity Initiative notes that only Adobe runs such a cloud service so far. |
| IPTC digital source type | XMP metadata | trainedAlgorithmicMedia ("Created using Generative AI") or compositeWithTrainedAlgorithmicMedia ("Edited using Generative AI"), plus camera and screenshot types. Next to an AI label you also see the credit line, creator and creator tool, when the file has them. |
| IPTC AI fields, added in 2025 | XMP: AI System Used, AI System Version Used, AI Prompt Information and AI Prompt Writer Name | The AI system and version, the prompt, and who wrote it. IPTC notes that the prompt writer should not be considered the image's creator. |
| Generator settings | PNG text chunks ("parameters", and ComfyUI's "prompt" and "workflow") and the EXIF user comment | The prompt, negative prompt, sampler, seed and model that Stable Diffusion apps, ComfyUI, NovelAI and InvokeAI save. |
| Software names | EXIF Software, XMP CreatorTool and the XMP edit history | The app that last saved the file. When it is an AI image tool, you see "Likely made with AI", because these fields are easy to edit. |
| China's AIGC label | A metadata field named AIGC, in XMP, a PNG text chunk or the EXIF comment | Its Label value, where China's national standard GB 45438-2025 uses 1 for AI-generated, 2 for possibly and 3 for suspected AI-generated, and the AI service it names. China's labeling measures require AI services to put such a label in the file's metadata from 1 September 2025. |
| Camera details | EXIF Make, Model, lens and date | A hint towards a real photo, never proof, because they are easy to copy. |
Sources: IPTC Photo Metadata Standard 2025.1, IPTC Digital Source Type vocabulary, C2PA Technical Specification 2.4, Working with manifests (Content Authenticity Initiative), Measures for Labeling AI-Generated Synthetic Content (Cyberspace Administration of China, in Chinese), GB 45438-2025, annex E (standard text in Chinese, copy published by Sidley Austin).
Questions
Is this an AI image detector or an AI image checker?
Both words describe it. It detects AI images from the evidence a file carries: signed Content Credentials, AI labels, generator settings and, through the linked checks, the invisible SynthID watermark. It does not guess from the pixels, so it gives no percentage score, and it tells you plainly when a file holds no evidence either way.
How does this AI photo detector differ from detectors that give a percentage?
A percentage comes from a model that looks at the pixels and guesses, and such guesses can be wrong in both directions: real photos get flagged, and AI images pass. This detector reports only what can be shown: who signed the file, which AI tool it names, and which label or prompt is stored in it. When it finds something, you can point to it.
Is this AI image checker free?
Yes. Everything that runs in your browser is free, with no sign-up and no daily limit. The SynthID checks we link to are run by Google and OpenAI on their own sites: Google's detector has a daily limit, and OpenAI says its check is currently free within usage limits.
Is this image AI generated? How do I find out for free?
Drop it into the checker above. If the file carries Content Credentials, an IPTC AI label or a generator's settings, you see which AI tool made it, at no cost and without an upload. If it carries none, run the SynthID checks from Google and OpenAI, then look for where the image first appeared with a reverse image search. No free tool can prove a picture is real.
Can this tell whether any image is AI?
No tool can do that reliably, and tools that guess from the pixels often accuse real photos. This checker reads what the file itself records: signed Content Credentials, AI labels and generator settings, which most big AI tools now add. When those are missing, it says so instead of guessing, and points you to Google's and OpenAI's own checks for the invisible SynthID watermark.
Does ChatGPT add a watermark to images?
Not a visible one. OpenAI says images generated with ChatGPT, Codex or the OpenAI API carry both Content Credentials and an invisible SynthID watermark. The ChatGPT images we checked in October 2026 had trusted Content Credentials from OpenAI OpCo, LLC, and the credentials themselves record that a watermark was added.
How do I check an image made with Gemini or Nano Banana?
Google puts its invisible SynthID watermark in images made with its AI. Upload the image to Google's SynthID Detector, which is open to everyone with a daily limit, ask the Gemini app "Was this image created or edited by Google AI?", or ask Google Lens "Is this made with AI?". Drop it here as well: if Content Credentials or an AI label survived, this checker shows them.
Why does an image I know is AI show no labels?
Screenshots never carry labels, and most social apps and messengers strip them when you post or forward an image. Ask for the original file, or try the SynthID checks, which survive many of those steps. Midjourney, Stable Diffusion and FLUX images only show up when their settings or labels are still in the file.
Is this a SynthID checker?
Not on its own. The checker reads what the file records on your device and cannot see SynthID. For ChatGPT and OpenAI API images, OpenAI's verify page reads OpenAI's SynthID watermark and its Content Credentials. For an image whose source you do not know, use Google's SynthID Detector: it checks for the watermark that Google, OpenAI, NVIDIA and Kakao add, and Google says it is not a general AI detector. Midjourney, Stable Diffusion and FLUX are not SynthID partners, so no SynthID check says anything about their images.
How do I check if an image was made by ChatGPT?
First, drop the original file here: ChatGPT images carry Content Credentials signed by OpenAI, which this checker verifies on your device. Second, upload the file to OpenAI's verify page: it also reads OpenAI's SynthID watermark, which survives more editing than the credentials do. Google's SynthID Detector reads OpenAI's watermark too.
What is SynthID?
An invisible watermark woven into the pixels by Google DeepMind's models and, since 2026, by OpenAI's. Google says it is designed to survive cropping, filters and lossy compression, though many edits can still weaken it. Google's SynthID Detector reads it for Google and partners such as OpenAI, the Gemini app reads Google's, and OpenAI's verify page reads OpenAI's. Read how SynthID works and how to check it.
What are Content Credentials?
A signed record, defined by the C2PA standard, that AI tools such as ChatGPT and Adobe Firefly and some cameras attach to images. It says what made the image and how it was edited. We check the signature against the public C2PA trust lists: "trusted" means a known company signed it and the image has not changed since. Read Content Credentials explained.
Can it read the prompt from an AI image?
Yes, when the generator saved it in the file. Stable Diffusion apps such as AUTOMATIC1111, and ComfyUI, write the prompt and settings into the image, and some tools fill in the IPTC "AI prompt information" field added in 2025. The checker shows them. ChatGPT does not store your prompt in the image, and OpenAI says its provenance signals hold no details about the user or the prompt.
Can AI image detectors be fooled?
Yes. Detectors that guess from the pixels give a probability that can be wrong in both directions, and a screenshot, filter or resize can change the answer. Researchers have shown that both watermark checks and pixel classifiers can be evaded, and that real images can be made to look watermarked. Signed Content Credentials are the hardest to fake, but they are easy to strip, so a missing record proves nothing.
Does this checker work without watermarks or metadata?
It can only report what the file holds. With no Content Credentials, AI label, generator settings or invisible watermark, nothing proves the image either way, and the checker says so instead of guessing. Next, run a reverse image search to find where the picture first appeared, ask the sender for the original file, and look for clues such as garbled text or odd hands.
Is my image uploaded?
No. The checks run in your browser, and the checker does not fetch Content Credentials that a file keeps online. OpenAI's verify page and Google's SynthID Detector, which we link to, are uploads. OpenAI says files sent to its verify page or API are not stored unless legally required and are not used to train its models. Google says its detector deletes your file once the result is back, and keeps a digital signature of it for 24 hours to enforce daily limits and defend against automated attacks.
Does a photo extended with AI uncrop show up as AI?
Yes. A photo you download after AI uncrop on uncrop.si carries the IPTC "composite with trained algorithmic media" label in its XMP metadata, and this checker reports it as edited with AI. So does a photo extended with Smart fill, the on-device AI fill we offered until November 2026. Like any metadata the label is not signed and can be stripped, so a missing label proves nothing.
Can SynthID or Content Credentials be removed?
We do not offer that. SynthID is designed to survive editing, and stripping provenance so an AI image passes as a real photo is deceptive. Google's terms for its SynthID Detector forbid using it to attack the watermark or to tune edits that break, bypass or weaken it. Rules in China and the EU require AI-generated images to be labeled. If one of your own photos is flagged wrongly, the original file and its Content Credentials are the way to show its history.
Related tools
Guides for this
AI images
How to tell if an image is AI-generated
Five steps that work: get the original file, read its labels, check for SynthID, look for visual clues and trace the source. What each test can prove.
Read the guideAI images
What is SynthID, and how do you check for it?
SynthID is the invisible AI watermark in Google and, since May 2026, OpenAI images. How it works, how to check it for free, and what a result proves.
Read the guideAI images
Content Credentials (C2PA) explained: how to verify where an image came from
C2PA Content Credentials are a signed record of the app, AI or camera that made an image. What they show, valid vs trusted, and how to verify them free.
Read the guide