How to Tell If an Image Was Made With Microsoft Copilot

0

๐Ÿ”Ž Reverse Image Search

Image search techniques made simple: upload a photo or paste a URL and search 6 engines at once. 100% free, nothing stored.

๐Ÿ”’ Your image is never stored on our servers longer than 24 hours and is never shared.

Microsoft Copilot can turn a typed sentence into a finished picture in seconds, and those pictures now travel everywhere: social feeds, blog posts, product listings and group chats. The trouble is that a Copilot picture looks like an ordinary photo once it leaves the app that made it. There is no file name that gives it away, and the image itself rarely says where it came from. This guide explains how to tell if an image was made with Microsoft Copilot, using the traces Microsoft leaves in the file.

That matters when you need to know the truth about a picture. You might be checking a viral image before you share it, deciding whether a photo is safe to use on your site, or working out whether a picture in a listing or a message is real. Guessing is not good enough for any of those jobs. The good news is that Microsoft leaves real traces in its images, and you can read them yourself in a minute or two.

Short answer: The most reliable way to tell if an image was made with Microsoft Copilot is to upload the file to the Content Credentials Verify tool and read its C2PA record. A Copilot or Designer image names Microsoft as the issuer and the tool that made it. If the file has no record, check the bottom-right corner for the Copilot watermark.

What Content Credentials are: C2PA Content Credentials are a signed record attached to a media file. Think of it as a label sewn into the file itself. For a Copilot image the label records that the image was created with an AI tool, which Microsoft tool made it, when it was made, and who signed the record. The record is protected by a digital signature, so it cannot be edited without breaking the seal. Our full guide to checking Content Credentials on a photo explains the standard in detail. This article stays on the Copilot part: what Microsoft writes into the record, and what to do when the record is missing.

How to tell if an image was made with Microsoft Copilot: the quickest way

Start with the file, not the picture. If you have the original file, upload it to the Content Credentials Verify tool. That is the fastest route because Microsoft says it adds content credentials to every image Copilot creates. Its Transparency Note for Copilot, updated 10 November 2026, puts it plainly: for all images created with Copilot, Microsoft has put in place content credentials, provenance based on the C2PA standard, to help people identify AI-generated content and establish its origin. The Designer help pages say the same for images made with Designer features.

Answer capsule: if the original file is available, the Verify check settles the question in under a minute.

Here is the order that wastes the least time.

  1. Keep the original file. Ask for it, download it directly from the source, or save the file as it was first shared. A screenshot of the image will not do, because a screenshot is a new picture of the old one, and the record does not travel with it.
  2. Upload the file to the Verify tool. The official reader is at contentcredentials.org/verify. It takes the file from your device and reads the record inside it. There is no sign-up and no payment for a basic check.
  3. Read the result. If the image is from Copilot or Designer, Verify says the image was generated with an AI tool and names Microsoft as the issuer. The worked example below shows exactly what that screen looks like.
  4. If there is no record, look for the watermark. Check the bottom-right corner for a small Copilot icon and the words “AI-generated”. The watermark is optional and easy to crop out, so its absence means little, but its presence is a clear answer.
  5. If both checks are blank, fall back to visual clues and a reverse search. This is the weakest evidence, and the last section of this article explains how far it can be trusted.

One warning before you start: do not ask Copilot itself whether it made the image. Showing a picture to Copilot and asking “did you make this?” is a different job entirely, and it cannot prove where a file came from. Our guide to asking Copilot about a photo covers that direction. Provenance is a property of the file, and only the file can answer it.

How do I check an image in Content Credentials Verify?

The Verify tool is the official reader for C2PA records, and it is free to use. Go to contentcredentials.org/verify, which takes you to the Verify site, and choose the file from your device. It accepts JPEG, PNG, WebP, AVIF, GIF and other common formats. There is no option to check an image by pasting a web address, so you need the file on your own device first. That is worth knowing before you start, because an image you can only see on a web page has to be downloaded first, and a download from a social platform may already have lost the record.

Answer capsule: Verify reads the signed record inside the original file and tells you, in plain words, whether the image was generated with an AI tool and which Microsoft tool made it.

Here is a real test, done on a real Copilot image, so you can see what a positive result looks like.

The worked example: a viral birthday image

The test image is a picture that circulated widely as engagement bait: a soldier in a wheelchair holding a cardboard sign that reads “TODAY’S MY BIRTHDAY, PLEASE LIKE”, with a United States flag behind him. It is published on Wikimedia Commons as a public-domain file, listed as created with Microsoft Copilot on 4 October 2024, under the category of images generated by Microsoft Copilot. The description on Commons says it was made from a suggested prompt at an “AI slop” clickbait seminar. The prompt named an American soldier veteran holding a cardboard sign with that birthday wording, injured in battle, with an American flag.

I could not generate a fresh image inside Copilot for this test, because Copilot asks you to sign in with a Microsoft account and no signed-in account was available. So instead of a new generation, this test uses a genuine Copilot-made file that is already public. The check itself was performed live on that exact file, and every result below is what the tool actually showed.

The file was uploaded to Verify on 20 November 2026. The result was a clear yes. The header read “This asset has content credentials”, and the summary said “This image was generated with an AI tool.” The record gave these details.

CheckWhere to lookWhat this Copilot image showedWhat it means
C2PA recordVerify result header“This asset has content credentials”The file still carries its signed provenance record.
Tool that made it“AI tool used” in Verify“Image Creator from Designer”The image was made with Microsoft’s Designer image tool.
App or device“App or device used” in Verify“Microsoft Designer 1.0”The app named in the record is Microsoft Designer, version 1.0.
Action“Actions” in Verify“Created”The record says the asset was created, not edited.
Issuer“Issued by” in Verify“Microsoft Corporation”Microsoft signed the record, which is why it can be trusted.
Date“Issued on” in Verify“Oct 4, 2024 at 5:46 PM GMT+5”The record was issued minutes after the image was made.
Trust badgeBadge beside the issuer“Legacy trust”Microsoft’s certificate sits on the frozen interim trust list, not the newer conformant list.
Visible watermarkBottom-right corner of the imageNone. No Copilot icon and no “AI-generated” text.This 2024 image was made before the watermark setting existed, so the corner is blank.
Original fileSecond node in the Verify chainA JPEG named asset-c45900f5-d360-421e-a701-12df48b73e8f.jpg, issued 5:45 PM the same dayThe PNG on Commons is a copy of a JPEG original, and both records survive.

Three things in that result deserve attention. First, Verify names no AI model. It does not say DALL-E 3, GPT Image or anything else. The tool that made the picture is named, the model is not. Second, Verify shows no prompt and no user name. The record says what made the image, not who asked for it or what they typed. Third, the trust badge reads “Legacy trust”, not “Conformant”. That does not mean the record is fake. It means Microsoft’s certificate is on the frozen interim trust list, and the C2PA conformance programme has not evaluated that product. The record is still signed by Microsoft, and that signature is what matters for this check.

The file itself backs up the Verify screen. Reading the raw record inside the PNG shows the software agent as “Image Creator from Designer”, the creation time as 4 October 2024 at 12:45:33 UTC, the action as “c2pa.created”, and the claim generator as “Microsoft Responsible AI Image Provenance 1.0”. The ingredient entry points back to the JPEG original. So the plain-language Verify result and the raw record agree with each other, which is exactly what a genuine signed record should do.

Other checkers can read the same record. Google’s AI Overview for this question names tools such as Sightengine and WasItAI, and they work by reading the same metadata. They are useful if you already use them, but Verify is the direct route, it is the official C2PA reader, and it shows the record without an account. Start there.

If Verify says it cannot find a record, do not stop there. A missing record is the most common result, and the next sections explain why it happens so often.

What does the Copilot watermark look like, and how do I switch it on?

The visible watermark is the second sign to look for. On an image that has it, you will see a small Copilot icon and the words “AI-generated” in the bottom-right corner. It is small and quiet, and it is easy to miss if you are not looking for it. On the birthday image in the worked example there is nothing in any corner: the top corners are flag, the bottom corners are a dark wall and a wheelchair. That is normal for an image from 2024, because the watermark setting did not exist then.

Answer capsule: the watermark is a small Copilot icon plus “AI-generated” in the bottom-right corner, and it only appears if the person who made the image switched it on.

For a personal Microsoft account, the setting lives on the Microsoft privacy dashboard. The path is account.microsoft.com/privacy, then Privacy, then under “Empower your productivity” select Copilot. The setting is headed “Include a watermark when content from Microsoft 365 is AI-generated”. Microsoft’s own support page for this setting gives the steps.

Turning it on adds the icon and the text to the bottom-right corner of any new image generated with Microsoft 365 apps, or of an existing image that you alter with them. Three limits matter. The setting is per account, so it follows the account to every device signed in to it. It does not apply to images created in Copilot Chat, so a picture made in a Copilot chat will not carry it. And it is not retroactive: images made before the setting was switched on do not gain a watermark afterwards.

For a work or school account the route is different. The setting is reached at myaccount.microsoft.com, then My Account, then Data and Privacy, then Data options, where the toggle is called “Show watermark”. Microsoft’s Learn page for administrators, updated 4 November 2026, describes the same visible mark: the Copilot icon and the words “AI-generated” in the lower-right corner of the image. The same page is honest about the weakness: the watermark can be removed by cropping the image, by downloading it in a different file type, or by taking a screenshot of it.

There are two quirks that catch people out. The first comes from Copilot Queen, who tested the setting in April 2026 and found that the watermark does not show while you are creating or editing inside Copilot Create. It appears only on the downloaded image, and on a copy pasted into Word. An image made in Word with the Designer agent, she noted, did not seem to get the watermark at all. The second quirk is about timing. According to Office Watch, images made in Microsoft 365 apps have been watermarked since 22 July 2026, and adding a Copilot image into a Microsoft 365 app makes a watermarked copy of it. So the same picture can exist in two forms: the original without a watermark, and the copy with one.

That is why the watermark is a useful clue but a poor test on its own. A watermark in the corner is strong evidence that the image came from a Microsoft tool. A blank corner proves nothing at all, because the maker may have had the setting off, may have made the image in Copilot Chat, may have made it before the setting existed, or may have cropped the corner away.

Which model made the image: DALL-E, GPT Image or something else?

People often want a model name, and Copilot makes that hard, because it has changed engines more than once and Microsoft does not always say which one is running. The honest answer is that the model depends on when and where the image was made, and the file itself will not tell you.

Answer capsule: older Designer images came from a DALL-E 3 pipeline, while current Microsoft 365 Copilot images come from the GPT Image family, but the C2PA record names the tool, not the model.

The timeline looks like this. Designer and the old Bing Image Creator ran on a DALL-E 3 based pipeline. That is the engine behind the October 2024 birthday image in the worked example, which is why Wikimedia Commons files it under a DALL-E category as well as under images generated by Microsoft Copilot. Then the engine changed. Microsoft announced in April 2026 that ChatGPT Images 2.0, built on OpenAI’s GPT Image 2.0, was available in Copilot in PowerPoint and was coming soon to Copilot Chat. By September 2026 Microsoft was showing GPT Image 2.5 in PowerPoint. In November 2026, OpenAI released GPT Image 2.6 and Microsoft showed its own MAI Image 2.7 in its Foundry platform at Ignite.

So if someone tells you a Copilot image was definitely made by GPT-Image-1.5, be careful. That name appears in third-party guides as the engine that replaced the older DALL-E pipeline at some point, but no official Microsoft page names it as the current Copilot backend, and the consumer pages name no model at all. Neither the Copilot Transparency Note of 10 November 2026 nor the Designer FAQ names a model. A guide that states a single model name for all Copilot images is guessing.

For your check, the model name matters less than it seems. The C2PA record names the tool, and the tool is what you need: “Image Creator from Designer” tells you the image came from Microsoft’s image tools. Whether the pixels were drawn by DALL-E 3 or by GPT Image 2.5 does not change the verdict. If the record says Microsoft made it with an AI tool, the image was made with Microsoft Copilot or Designer, whatever engine sat underneath on the day.

What is the hidden watermark that can point back to the person who made it?

Beyond the visible mark in the corner, Microsoft images can carry a second watermark that you cannot see at all. It is hidden in the file, and it is the part of this story that surprises most people. Two separate investigations found it, and they found two different versions of it, so it is worth keeping them apart.

Answer capsule: some Microsoft images carry an invisible identifier hidden in the file, but only Microsoft can match it to a person, and you cannot read it with any public tool.

The first finding comes from Office Watch, in an investigation published 27 August 2026. It reported that Microsoft quietly added a hidden watermark to Copilot images, and that the file carries a GUID, a long identifier, that is the same for every image made by one Copilot user, plus a timestamp. That means an image can be traced back to a user account, but only Microsoft can do the matching. The file metadata does not say which Microsoft app made the image, and it does not say what prompt was used. The GUID identifies a user, not a single image. This finding is Office Watch’s investigation, and it is reported here as their finding, not as a test performed for this article.

The second finding comes from software engineer Xusheng Li, who took apart recent versions of Paint and Photos and published his analysis in 2026. Paint’s Cocreator, on Copilot+ PCs, generates the picture locally on the device using a Stable Diffusion model, which sounds offline. It is not fully offline. Before Paint runs the local model, it sends the prompt and the chosen style to a Microsoft server for content moderation. The server answers with two identifiers: a prompt generation ID and a watermark ID. Paint hides the watermark ID, a 16-byte GUID, inside the pixels of the finished image, and it also attaches a server-signed C2PA record to the saved file. Paint only lets you save an AI image in formats that can carry that record: PNG, JPEG, GIF, or its own .paint format.

The Paint identifier is different from the Copilot one in an important way. The Paint watermark ID is issued per generation, so each picture gets its own value. The Office Watch Copilot finding is a GUID that stays the same across all of one user’s images. Do not mix the two up. Both are invisible, both are issued by a Microsoft server, and neither can be read by an ordinary user. The Verify tool does not show a GUID, and no public tool will turn one into a name.

Paint also has its own visible mark: a small Copilot logo in the bottom-right corner of the image. Microsoft’s Image Creator support page says plainly that images made with its tools “will contain C2PA manifest helping users identify that it is an AI generated image”. So a Paint Cocreator picture carries the same three layers as a Copilot picture: a visible logo, a signed C2PA record, and an invisible identifier. The checks in this article work the same way on it: read the C2PA record first, look at the corner second, and treat the invisible layer as something only Microsoft can use.

What if the image has no metadata and no watermark?

This is the situation you will meet most often, and it is the one where people make the most mistakes. An image shared on social media, sent through a messaging app, or saved as a screenshot usually arrives with no C2PA record at all. Microsoft’s own Designer FAQ warns that the metadata can be stripped by sharing on social media or by downloading the image, and that sharing the original file is needed to prove where it came from. Cropping, compression, converting to another format and re-uploading can all strip it too. So a blank result in Verify is not a clean bill of health. It only means the record is gone, and it says nothing about who made the picture.

Answer capsule: a missing record proves nothing, so fall back to visual clues and a reverse search, and treat the result as a likelihood, never as proof.

Start with the picture itself. Google’s AI Overview for this question lists the visual signs worth checking: hands and fingers that look wrong, small background text that reads wrong, reflections that do not match the scene, shadows that fall the wrong way, hair and edge transitions that look smeared, and 3D geometry that does not quite hold together. Microsoft Research tested how far the eye can be trusted: in its Real or Not quiz, more than 12,500 people judged about 287,000 images and got only 62% right, barely better than chance. The birthday image in the worked example is a good illustration of the text problem: the sign reads correctly, which is a point in its favour, but the lettering is painted on with a slightly uneven hand, and the bandage on the soldier’s leg has a wrapped, featureless look that a real photograph would rarely show. None of those details proves anything on its own. Together they raise a question, and the question is answered by the file, not by the eye.

One more quick check on the file itself: open the image’s properties and look for camera details. A real photograph usually lists a camera make and model and a focal length. An AI image has none of those fields, because no camera took it, so an empty properties panel is another small sign, not proof on its own. Then run a reverse image search. Upload the picture to Google Lens, TinEye and Yandex and see where it first appeared. If the earliest appearance is a Copilot gallery, a Designer showcase, or a social post that says it was made with Copilot, that is useful context. Our guide to how to reverse image search walks through the steps on each engine. A reverse search can also find the original file, which may still carry its C2PA record even though the copy you have does not.

Detector tools are a third option, but treat them with care. A detector gives a probability score, not a source name. No detector can tell you an image came from Copilot specifically, and scores can mislead badly: in our own testing for other articles, a proven face swap scored 0% AI on one detector while a real photograph scored 99.9%. One test on one image is a hint, never a verdict. Our guide to using the Pangram image detector explains how to run a test properly and how to read the result. The same caution applies to the general signs of AI imagery, which our guide to how to tell if an image is AI-generated covers in full.

Finally, remember the comparison that beats every other guide on this topic. The strongest competitor page we found, a twelve-generator comparison table, lists Microsoft Designer but has no row for Copilot at all. The other pages either explain watermarks without C2PA, or explain C2PA without Copilot. This article brings the two together, with a real file tested live, because that is the only combination that answers the question properly: the record says Microsoft, the tool is named, and the watermark and the hidden identifier are explained honestly, including their limits.

FAQs

Can I tell if an image was made with Copilot just by looking at it?

Not reliably. Some Copilot images have a Copilot icon and the words AI-generated in the bottom-right corner, but that watermark is optional and easy to crop out. Visual flaws such as wrong text or odd hands are only hints. The reliable check is the file itself: upload the original to the Content Credentials Verify tool and read its signed record.

Does every Copilot image have C2PA metadata?

Microsoft says it adds content credentials to all images created with Copilot, and the same applies to Designer. But the record only survives in the original file. Screenshots, social media uploads, cropping and format changes can strip it away. So a missing record does not prove an image is real. It only proves the record is gone.

Where do I turn on the Copilot image watermark?

For a personal account, go to the Microsoft privacy dashboard at account.microsoft.com/privacy, select Privacy, then under Empower your productivity select Copilot, and switch on Include a watermark when content from Microsoft 365 is AI-generated. For a work or school account, the toggle is called Show watermark and sits under Data and Privacy in your Microsoft account.

Does the watermark prove who made the image?

No. The visible watermark is the same on every image, so it says nothing about the maker. The hidden identifier is different: Office Watch reported that Copilot images carry a GUID shared by all images from one user, but only Microsoft can match that GUID to an account. The Verify tool shows no name, no prompt and no user details.

Can a Copilot image have no watermark at all?

Yes, easily. The watermark only appears if the maker switched it on, and it does not apply to images made in Copilot Chat or to images made before the setting existed. Images in Microsoft 365 apps have been watermarked since 22 July 2026, but a cropped copy, a screenshot or a re-upload can lose the mark entirely.

Is a detector score enough to say an image is from Copilot?

No. A detector score is a probability that an image is AI-made, not a statement about which tool made it. No detector can name Copilot as the source, and scores can mislead: a fake has scored 0% AI while a real photo scored 99.9%. Use the C2PA record first, the watermark second, and a detector score only as a hint.

Ali Butt is the lead writer and SEO specialist at Image Search Techniques, where he covers reverse image search, visual search, and online image verification. With over four years of experience in SEO and content, he writes practical, hands-on guides that help readers trace a photo to its source, identify people, products, plants and brands from a picture, protect their own images online, and spot catfish, fakes and AI-generated images across Google, Yandex, Bing, TinEye and more.

Leave A Reply