OpenAI now lets you check whether an image came from its AI tools

OpenAI is making image verification a little more practical for everyday users. The company has previewed a public tool at openai.com/verify that lets people upload an image and check whether it contains signals associated with OpenAI-generated media.

The tool is not a universal truth machine. It does not prove that a picture is real, fake, accurate, misleading, stolen, or legally usable. Its job is narrower: look for supported provenance signals that indicate an image was generated with OpenAI tools such as ChatGPT, Codex, or the OpenAI API.

That limitation matters, but the launch is still a meaningful shift. Until recently, AI image detection usually meant guessing from visual artifacts, running unreliable classifiers, or trusting whatever a platform label happened to say. OpenAI is now leaning into a more structured approach: signed metadata plus invisible watermarking.

In simple terms, the company wants AI images to carry receipts.

What OpenAI is launching

OpenAI's new verification page is a research preview for checking images. You upload a PNG, JPG, or WEBP file, and the service scans for two categories of signals: C2PA Content Credentials and SynthID watermarks.

If one of those supported signals is detected, the result can indicate that the image likely originated from OpenAI's image tools. According to OpenAI, detected signals are designed to be reliable, with rare false positives.

The tool is aimed at a real problem: synthetic images are now good enough that many people cannot spot them by eye. A dramatic news-like photo, a fake product shot, a staged public figure image, or a viral screenshot can travel faster than anyone can verify it.

OpenAI is not saying this preview solves that problem by itself. It is saying that the provenance layer should be easier for ordinary users to inspect.

That is the useful part. Verification should not require a forensic lab or a command-line metadata tool.

OpenAI announced the change on X, saying it is adding new ways to identify AI-generated images and understand where they came from, including SynthID watermarks alongside C2PA Content Credentials.

How C2PA and SynthID split the work

The first layer is C2PA, short for the Coalition for Content Provenance and Authenticity. C2PA is an open technical standard for attaching provenance metadata to media. It can describe where a file came from, how it was created or edited, and which organization signed that information.

OpenAI says it has added Content Credentials to images since 2024 and has now become a C2PA Conforming Generator Product. That should make OpenAI provenance data easier for other platforms and tools to read consistently.

The second layer is SynthID, Google's invisible watermarking technology from Google DeepMind. Unlike metadata, SynthID is embedded into the media itself. The signal is meant to survive some common transformations such as resizing, compression, filters, or screenshots.

Those two systems solve different weaknesses. Metadata can carry more context, but it is easy to strip accidentally or deliberately. A watermark carries less explanatory detail, but it may survive when metadata disappears.

Together, the two signals make the chain of origin harder to break than either approach alone.

How openai.com/verify works

The public flow is intentionally simple. You upload one image, wait for the scan, and review whether the tool found C2PA metadata, a SynthID watermark, both, or no supported signal.

OpenAI recommends cropping screenshots closely around the image and avoiding uploads that contain multiple images in one file. That is sensible because a screenshot with several pictures, UI chrome, text, and compression artifacts makes provenance detection more ambiguous.

The tool currently focuses on OpenAI-generated images. If an image was made by another model, the OpenAI verifier may not detect it even if the image is entirely AI-generated.

That is one reason the result needs careful wording. A positive signal is informative. A negative result is not a clean bill of authenticity.

For privacy, OpenAI says uploaded images are processed to check supported signals, are not used to train its models, and are not stored unless legally required.

The important limitation

The most important line in this story is the one that sounds least exciting: no signal does not mean no AI.

An image could fail the check for several reasons. The metadata might have been removed by a social platform. The watermark could have been degraded by editing. The image could have been created before these signals were available. It could also come from a different AI system that does not use OpenAI's provenance stack.

That makes the tool best understood as a confirmation mechanism, not a universal detector.

If it finds an OpenAI-linked signal, that is useful evidence about origin. If it finds nothing, you still need ordinary verification habits: search for the earliest source, check context, compare with trusted reporting, inspect surrounding claims, and avoid treating a viral image as proof by itself.

This is less glamorous than the phrase "AI detector", but it is much safer.

Why this matters for platforms and journalists

For journalists, fact-checkers, educators and platform trust teams, provenance is becoming part of the basic media workflow.

The old internet assumption was that a photo could still lie, but it usually started as a camera capture. The generative AI era breaks that assumption. A plausible image can now be created from text, edited conversationally, exported, screenshotted, reposted and detached from its origin in minutes.

C2PA gives professional users a way to inspect a file's declared history. SynthID gives them another signal when metadata has gone missing.

Neither replaces editorial judgment. A genuine photo can still be used in a false context. An AI-generated illustration can be harmless if it is clearly labeled. A screenshot can hide important surrounding details.

Still, a reliable origin signal helps platforms make better integrity decisions and helps newsrooms avoid treating every suspicious image as a purely visual guessing game.

Why the Google partnership is bigger than one tool

The OpenAI update matters partly because of the Google piece. OpenAI is adding Google DeepMind's SynthID to images generated through ChatGPT, Codex, and the OpenAI API.

Google is also expanding its own verification and provenance work across products such as Search, Gemini, Chrome, Pixel and Cloud. At the same time, companies including OpenAI, Kakao and ElevenLabs are adopting SynthID for more generated media.

That makes this less like a one-company feature and more like an early attempt at shared infrastructure.

The web does not need twenty incompatible labels that only work inside one platform's garden. It needs provenance signals that can travel, be read by multiple tools, and survive common resharing patterns.

The open question is whether enough AI companies will participate. If only a few major labs label their output, bad actors can simply move to tools that do not.

What users should do before trusting an image

For ordinary users, the new OpenAI verifier is worth bookmarking, but it should sit inside a broader verification routine.

First, check provenance if the image claims to show something consequential: a public event, a product announcement, a scandal, a disaster, a market-moving screenshot, or a person doing something surprising.

Second, treat "no signal found" as inconclusive. It means the tool did not detect supported OpenAI signals, not that the image is real.

Third, compare the image with other sources. Reverse image search, original publisher checks, known account history and timestamps still matter.

Fourth, watch the context. A real image from one event can be reposted as if it came from another. Provenance can help with creation history, but it cannot judge every claim made around the image.

The practical habit is simple: use the verifier as one piece of evidence, not the final verdict.

The trust layer is becoming infrastructure

OpenAI's verification preview is not the end of AI image confusion. It is one useful brick in a much larger trust layer.

The internet is moving toward a world where media needs machine-readable context. Cameras, creative tools, AI generators, social networks, browsers and search engines may all have to participate if provenance is going to work at web scale.

That will not stop deception. People can still crop, re-render, screenshot, forge context, or use unlabeled models. But it can raise the cost of deception and give honest creators, publishers and users better tools.

The best version of this system is not a magic detector that claims certainty. It is a set of interoperable signals that make origin easier to inspect.

OpenAI's new page is a small public doorway into that future. If enough of the industry follows through, checking where an image came from may become as normal as checking a link before clicking it.

(Photo: MJ Duford / Unsplash, license.)