How to Tell If an Image Was Made with ChatGPT: OpenAI’s Verify Page and the Legacy Gap

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.

Upload the original file to OpenAI Verify at openai.com/verify: a detected C2PA record or SynthID watermark means it was made with ChatGPT or the OpenAI API. But no signal proves nothing, images made before February 2024 carry no record, screenshots strip the metadata, and the tool can fail. A clean result starts the check; it is not the verdict.

This guide shows you how to tell if an image was made with ChatGPT, using the two provenance signals OpenAI actually attaches to its pictures, the three free tools that can read them, and the honest limits of each one. Everything below was checked on 12 November 2026, including a real three-file test in which two of the three tools misbehaved. That mess is itself part of the answer, because it shows how fragile provenance checks are in practice.

What “made with ChatGPT” means: an image generated or edited by ChatGPT, Codex or the OpenAI API, the three OpenAI tools that attach provenance signals to their output. It does not mean “made by any AI tool”, and it does not mean “made by DALL-E 2”, which predates the signals entirely. Keeping that definition tight matters, because every tool in this guide can only answer the narrow question, was an OpenAI tool involved?, and nothing broader.

What does “made with ChatGPT” actually mean?

ChatGPT makes pictures with OpenAI’s image models. The current generation is GPT Image, sold to developers as the GPT-Image models; the newest version, Images 2.6, arrived on 4 November 2026 with two API models, GPT-Image-2.6 Flare and Sunburst. Older pictures came from DALL-E 3, and older still from DALL-E 2. All of them can look identical on screen, and none of them carries a visible label by default. That is why the check has to look inside the file, not at the picture.

There are exactly two things to look for. The first is a C2PA Content Credential: a signed manifest tucked inside the file that names the tool and the organisation that signed it. The second is a SynthID watermark: an invisible pattern woven into the pixels themselves. OpenAI has attached Content Credentials to DALL-E 3 images since February 2024, according to OpenAI’s own provenance announcement, and it added the SynthID watermark to images from ChatGPT, Codex and the API on 19 May 2026, when it also joined the C2PA steering committee. One is a paper trail; the other is a fingerprint. Together they are the whole basis for answering the question.

That narrow scope is the first thing to understand. A general AI detector that says “probably AI” has not told you the image was made with ChatGPT, it has told you it looks synthetic, which is a different claim, and a probability score never names a tool. Likewise, asking ChatGPT or Gemini what a photo shows answers a question about content, not about origin. If you want the full picture on that distinction, our guide to telling whether any image is AI-generated covers the detector side; this guide covers the provenance side only.

Which signals does a ChatGPT image carry, and since when?

The answer depends entirely on when the image was made, because OpenAI switched the signals on in stages. This is the legacy gap at the heart of the article, so it is worth setting out plainly.

When the image was madeC2PA recordSynthID watermarkWhat a check can show
Before February 2024 (DALL-E 2 and earlier)NoNoNothing, no signal exists to find
February 2024 to May 2026 (DALL-E 3, early GPT Image)YesNoThe C2PA record only, if the file is original
From 19 May 2026 (ChatGPT, Codex, API)YesYesBoth signals, if the file is original
From 4 November 2026 (Images 2.6)YesYesBoth signals, per OpenAI’s launch announcement

OpenAI describes the two layers as deliberately complementary: C2PA carries detailed context about where a file came from, while SynthID survives transformations that strip metadata, such as screenshots. That design is the reason the company offers a single verification page that checks both at once, and it is also the reason a failed check is never a clean bill of health. The details above come from OpenAI’s provenance announcement of 19 May 2026, which also notes that SynthID was extended to supported AI-generated audio on 31 July 2026.

Edits complicate the picture further. An image generated in ChatGPT and then edited inside ChatGPT keeps its signals, updated to reflect the edit. An image downloaded from a chat, re-saved in another app, cropped, or re-uploaded to a social platform may keep the watermark but lose the metadata, or lose both. The rule of thumb is simple: the further a file travels from the original download, the less its signals can be trusted to still be there.

How to tell if an image was made with ChatGPT

OpenAI Verify is the company’s public verification page, and it is the only tool that checks both of OpenAI’s signals in one place. It is free, it works without signing in, and it accepts one file at a time. Use it like this:

  1. Get the original file. Ask the sender for the file as ChatGPT delivered it, a PNG, JPG or WebP. A screenshot of the picture, or a copy saved from a social feed, is a weaker starting point, because the metadata is usually already gone.
  2. Go to openai.com/verify. The page accepts images and audio; for pictures the supported formats are PNG, JPG, WebP and GIF.
  3. Upload the single file. Drag it in or click to browse. There is no batch mode and no URL paste option.
  4. Read the result. The page reports whether it detected C2PA metadata, a SynthID watermark, or no supported signal.
  5. Interpret it one way only. A detected signal means the file was generated with an OpenAI tool. No signal means the tool could not conclude anything, it does not mean the picture is a real photograph.

One honest warning: the tool can simply fail. On 12 November 2026 I uploaded three genuine files to OpenAI Verify and every upload returned the same red error: “Something went wrong. Please try again. If the issue persists, contact support.” The newest file was tried five times, including after a page reload and a minute’s wait, and it failed every time. OpenAI’s own FAQ for the page says that if the tool is not working you should try again in a moment, because some errors, such as rate limits or temporary issues, can occur. So if you meet the same error, wait and retry later; do not treat an error as a result, and do not assume a failed upload means the picture is real. The tool’s own help page confirms it is usable without signing in, although a “Log in” button is shown.

What does the SynthID Detector check, and why does it ask you to sign in?

Google’s SynthID Detector, at synthid.com, is the dedicated reader for the SynthID watermark. Google opened it to everyone on 7 October 2026, and it checks images, video and audio for the watermark. It is a narrower tool than OpenAI Verify: it reads the watermark only, not the C2PA record, and its own page states plainly that it is not a general AI detector, it can only detect media from the companies that have adopted SynthID, which it lists as NVIDIA, OpenAI, Google and kakao.

The catch is the sign-in. When I opened synthid.com on 12 November 2026, the page showed “Not signed in”, and clicking the upload area opened a sign-in dialog reading “Please sign in before detection”, offering exactly three buttons: sign in with Google, sign in with Apple, or sign in with ChatGPT. There is no guest or anonymous mode, and no way to run a check without an account. I did not sign in, because this site does not log in to services with the owner’s accounts, so I cannot report a SynthID Detector result of my own. If you have one of those accounts, you can use it yourself; if you do not, OpenAI Verify and C2PA Verify are the two sign-in-free routes.

Other generators have their own marks, and they must not be confused with ChatGPT’s. Grok images, for example, carry a visible corner mark rather than a provenance record, our guide to telling whether an image was made with Grok explains that one. A SynthID hit tells you the file came from one of the four listed companies; only the C2PA record or OpenAI Verify can narrow that down to OpenAI specifically.

The worked check: three files, three tools, honest results

To see how the three tools behave on real files, I ran the same three pictures through OpenAI Verify and C2PA Verify on 12 November 2026, and I checked the SynthID Detector’s sign-in requirement the same day. The files were a GPT Image 2.5 picture made on 21 September 2026, an Oktoberfest scene in the style of a Rembrandt painting; a ChatGPT image from 1 September 2025; and an original DALL-E picture from January 2021, made from the prompt “an illustration of a baby daikon radish in a tutu walking a dog”. Nothing below is a pre-written outcome; it is what the tools actually said, including where they said nothing at all.

FileOpenAI VerifyC2PA VerifySynthID Detector
GPT Image 2.5 picture, 21 Sep 2026No result, “Something went wrong” on all 5 attemptsValid credentials: “Issued by OpenAI OpCo, LLC”, “This image was generated with an AI tool”Not run, sign-in required
ChatGPT image, 1 Sep 2025No result, “Something went wrong”Invalid: “This file may have been tampered with. Its Content Credentials can’t be verified or viewed.”Not run, sign-in required
Original DALL-E picture, Jan 2021No result, “Something went wrong”No verdict shown after 3 attempts, cannot claim “no credentials found”Not run, sign-in required

Only one of the nine possible checks produced a clean positive, and it was the newest file. The C2PA Verify result for the September 2026 picture was unambiguous: the panel read “This asset has content credentials”, named the software as “OpenAI Media Service” and the organisation as “OpenAI OpCo, LLC”, showed a “Conformant” badge, and listed two actions in the file’s history, “Created” and “Format converted”. No creation date was displayed anywhere in the Verify interface, so none can be quoted.

The September 2025 result is the most instructive of the three. That file does contain a C2PA box in its bytes, I checked, yet C2PA Verify refused it as invalid, with the red banner saying the file may have been tampered with and its credentials cannot be verified or viewed. A manifest that is present but broken proves nothing, and a reader who only looked for the presence of a record would have been misled. The January 2021 DALL-E picture, which predates every signal, produced no verdict at all: the upload attached three times but the app stayed on its landing page. I am not going to dress that up as a “no credentials found” result, because the check never ran.

Reading the C2PA record yourself

If you only need the metadata half of the check, C2PA Verify needs no account: upload the original file and look for the issuer line. A genuine ChatGPT record names “OpenAI OpCo, LLC” as the issuer and “OpenAI Media Service” as the software that made it. Our full walkthrough of checking a photo’s Content Credentials covers the portal step by step, including what a valid, invalid and absent record each look like, this article does not repeat those steps, because the portal is the same whichever picture you are checking. If the photo is yours, you can sign it yourself: our guide to adding Content Credentials to your own photos covers the three routes, and what each one actually proves.

Why does “no signal” never prove an image is real?

Because the check is one-directional by design. A detected signal is positive evidence: the file carries a signed record or a watermark that only an OpenAI tool would have put there. An empty result is the absence of evidence, and there are at least five ordinary reasons a genuine ChatGPT picture can come back empty.

  • It is too old. Anything made before February 2024 predates the C2PA record, and anything made before 19 May 2026 predates the SynthID watermark. A DALL-E 2 picture has nothing to find, however good the tool is.
  • The metadata was stripped. Screenshots, re-saves, format changes and uploads to social platforms routinely drop the C2PA manifest. OpenAI says so itself: metadata can be stripped, lost through uploads and downloads, or broken by transformations such as format changes, resizing or screenshots.
  • The watermark degraded. SynthID is built to survive screenshots and re-encoding better than metadata, but it is still a signal that can degrade, and only two tools can read it, one of which needs a sign-in.
  • It came from another generator. Midjourney, Adobe Firefly, Grok and open-source models carry no OpenAI signal at all. An empty OpenAI Verify result says nothing about them.
  • The tool itself failed. As the worked check showed, OpenAI Verify can return a bare error instead of a verdict, and a verification app can stall without displaying anything. An error is not a result.

OpenAI puts the same point in its own announcement: no detection method is foolproof, and when no metadata or watermark is detected the tool will not make a definitive conclusion about whether the image was generated with OpenAI tools, because provenance signals can in some cases be stripped. That caution is the correct way to read every blank result when you are trying to tell if an image was made with ChatGPT. If you need certainty and the file shows no signal, the honest answer is “unproven”, not “real”.

One more boundary deserves stating. AI detectors such as Hive answer a different question, “does this look AI-made?”, and a probability score never names ChatGPT. Keep them subordinate: useful as a nudge, useless as proof of origin. And asking an AI assistant about a picture is a different task again, it tells you what the picture shows, not where it came from. Our guide to identifying a photo with ChatGPT and Gemini covers that side of the question. Perplexity is the assistant that answers with numbered sources, and our guide to asking Perplexity about a photo shows how its image upload works and how to check its claims.

What should I do when I need a reliable answer?

Follow the checks in order, and stop as soon as one of them gives you a positive signal. A detected C2PA record or SynthID watermark ends the question; everything after that is corroboration, not doubt.

  1. Get the original file. Ask whoever shared the picture for the file as it was downloaded from ChatGPT. If they can only send a screenshot, say so in your notes, the check is already weaker.
  2. Run OpenAI Verify. Upload the original at openai.com/verify and read the verdict. If it errors, wait and try again later; the tool was failing for every file on 12 November 2026, and its own help page blames temporary issues or rate limits.
  3. Run C2PA Verify. Upload the same file and look for the issuer line. “OpenAI OpCo, LLC” with the summary “This image was generated with an AI tool” is a positive result. “Invalid” means the record cannot be trusted, not that the picture is real.
  4. Try the SynthID Detector if you have an account. It reads the watermark that C2PA Verify cannot see, and it is the only way to catch a picture whose metadata was stripped but whose pixels still carry the mark.
  5. Reverse-search the image as well. A provenance check tells you how a file was made; a reverse image search tells you where it has appeared. Run the picture through our free reverse image search tool and look at the earliest appearance, a copy that predates ChatGPT’s image tools, or one sitting on a stock-photo site, changes the story entirely.
  6. Ask the sender. If the picture is being used as evidence, a listing, a profile, a news claim, ask for the original file and for the prompt or the chat it came from. A genuine creator can usually produce both in a minute.
  7. Write down “unproven”, not “real”. If every check comes back empty, that is the end of what the tools can tell you. Do not upgrade an absence of evidence into a verdict.

The short version is the one the tools themselves insist on: a signal found means ChatGPT made it; no signal found means you do not know. Everything else, the timeline, the three tools, the failed uploads in the worked check, is just the detail that makes that one sentence trustworthy. The same “no signal proves nothing” problem applies to searching itself: Google Images has no AI filter, so our guide to finding real photos, not AI images, in Google Images shows the workarounds, including the date limit and the one-minute spot-check, for getting real photos out of a search.

Can OpenAI Verify prove an image was not made with ChatGPT?

No. A detected signal proves an OpenAI tool was involved, but no signal proves nothing. Pictures made before February 2024 carry no record, screenshots and re-uploads strip the metadata, and other generators carry no signal. OpenAI says the tool makes no definitive conclusion when nothing is detected. Treat a blank result as unproven, never as proof of a real photograph.

Does a screenshot of a ChatGPT image still carry the signals?

Usually not the C2PA record. A screenshot re-encodes the picture, which normally drops the signed metadata, so C2PA Verify has nothing to read. The SynthID watermark is designed to survive screenshots, but only OpenAI Verify or the SynthID Detector can read it, and the detector needs a sign-in. Always ask for the original file before drawing any conclusion.

Can I use Google Lens or a reverse image search to tell if an image was made with ChatGPT?

No. A reverse search finds where a picture has appeared online, not how it was made, so a match never names the generator. It is still useful as corroboration: if the earliest copy predates ChatGPT’s image tools, or the same picture sits on a stock-photo site, that changes the story. Use it alongside a provenance check, never instead of one.

What is the difference between C2PA and SynthID?

C2PA is metadata inside the file: a manifest naming the tool and the organisation that signed it, rich in detail but easily stripped by screenshots and re-uploads. SynthID is a watermark woven into pixels: less detail, but it survives screenshots and re-encoding better, and only Google’s detector or OpenAI Verify can read it. ChatGPT images carry both since May 2026.

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