How to Tell if an Image Is AI Generated: What OpenAI’s New Watermarks Mean for Photographers

 8 min read

What OpenAI Changed and Why It Matters

Knowing how to tell if an image is AI generated has become a working requirement for photographers, not a curiosity. On May 19, 2026, OpenAI announced a change with direct consequences for your business. Every new image produced through ChatGPT, Codex, and the OpenAI API now carries two provenance signals: C2PA metadata and an invisible SynthID watermark. As a result, one of the largest AI image sources online became traceable overnight.

This matters because your competition is no longer only other photographers. Clients, photo editors, and contest judges now receive AI submissions alongside real work. For years, the question of whether AI will replace photographers stayed abstract. Now the practical task is narrower: proving a specific image is a genuine photograph. OpenAI’s move hands you a partial tool for the job.

The stakes are concrete. Judges disqualified the winner of Tokina’s monthly photo contest after finding an invisible SynthID watermark in the submitted image. Stock agencies screen uploads for AI content. Editorial clients ask for proof of authenticity before publication. Each of these checkpoints now leans on the same provenance signals OpenAI adopted.

This article explains how the new system works, how to check an image yourself, and where the technology breaks down. You will also see what the shift means for contests, licensing, and client deliverables. Adobe’s generative AI tools and other AI models stay outside OpenAI’s system, so the limitations section deserves close attention.

OpenAI’s Announcement at a Glance

The table below breaks the announcement into its core parts. Each row reflects what OpenAI confirmed in the update.

How to Tell if an Image Is AI Generated Right Now

Verifying an AI image takes four practical methods: OpenAI’s Verify tool, the C2PA Content Credentials, a SynthID watermark check, and a manual inspection for missing capture data. Still, none is conclusive on its own. Used together, however, these checks give you a reliable verdict on most images circulating today.

Run It Through OpenAI’s Verify Tool

OpenAI’s Verify tool is the fastest first step. Visit openai.com/research/verify, upload the image, and the site reports whether the file shows signs of an OpenAI origin. It looks for an OpenAI SynthID watermark or a trusted C2PA manifest from OpenAI. If it finds either signal, the image came from an OpenAI tool. OpenAI states errors happen but stay rare. Because the check takes seconds, run it before you trust any image submitted as a real photograph.

Check the Content Credentials

C2PA metadata travels with the file as Content Credentials. To read it, open the image at contentcredentials.org and use the free Verify page, or inspect the file in Photoshop or Lightroom. A genuine AI image from OpenAI shows a manifest naming the tool and the generation date. A camera original shows capture data instead, such as the device model and exposure settings. When the Content Credentials panel comes back empty, treat the result as inconclusive rather than proof of a real photo.

Look for a SynthID Watermark

SynthID hides a signal inside the image pixels, so it survives edits and screenshots. Google offers a SynthID detector, and the Gemini app reports whether an image carries the watermark. Ask Gemini directly whether an image is AI generated, and it checks for the SynthID signal. This method matters most when someone has stripped the metadata, because the watermark stays put. Yet the detector only recognizes watermarks from participating tools.

Inspect the Image Yourself

Manual inspection used to catch AI images through warped hands, mangled text, or impossible lighting. Current models fixed most of those tells, so visual inspection now ranks last. Look instead for context clues: a reverse image search with no camera-original source, a creator with no shooting history, or EXIF data with no capture settings. These signals raise suspicion, though none confirms an AI origin alone. Treat manual review as a prompt to run the three checks above.

C2PA vs. SynthID: Which Protects Your Work?

C2PA and SynthID solve the same problem from opposite directions. C2PA attaches a provenance manifest to the file. This manifest records how the image originated and which edits followed, all backed by cryptographic verification. SynthID, by contrast, embeds a signal directly into the pixels. One travels beside the image; the other sits inside it.

Durability is where the two systems split. Metadata is fragile: a screenshot, a re-upload, or a quick export strips C2PA entirely, and bad actors remove it on purpose. SynthID resists more abuse, since the watermark stays embedded after screenshots and many edits. As a result, the watermark is the stronger signal for catching a disguised AI image. C2PA, by contrast, carries far more detail for documenting a real photograph.

Neither replaces the other. OpenAI uses both for a reason, and the company put it plainly: watermarking survives transformations like screenshots, while metadata provides more information than a watermark alone. For your work, the takeaway is simple. Rely on SynthID to catch fakes, and rely on C2PA to prove provenance.

Where Knowing How to Tell if an Image Is AI Generated Breaks Down

OpenAI’s system closes one gap and leaves several open. First, it covers OpenAI tools only. An image from Midjourney, Google’s models, Adobe Firefly, or an open-source generator carries no OpenAI signal, so Verify returns nothing useful for those files. The internet runs on dozens of AI image sources, and this update tags one of them.

Metadata removal also stays trivial. Anyone able to take a screenshot defeats C2PA in one step. SynthID resists more abuse, yet no AI watermark survives a determined, well-resourced attacker. Treat a clean result as weak evidence, not a guarantee.

False results happen as well. OpenAI admits errors occur, even if rarely. For example, a real photograph wrongly flagged as AI damages a photographer. An AI image wrongly cleared misleads a client. Because the verdict is probabilistic, pair it with human judgment on anything with money or reputation attached.

Meanwhile, detection keeps advancing on the platform side. YouTube’s AI likeness detection now scans uploads for synthetic faces. Each platform builds its own checks, so a single universal answer does not exist yet. For now, your safest approach combines several signals and stays skeptical of any image with a missing trail.

What This Means for Contests, Licensing, and Clients

Photo contests changed first. Judges now run AI image detection on finalists, and disqualifications follow. The Tokina contest and the Hasselblad Masters 2026 competition both removed AI entries this spring. If you enter contests, expect provenance checks, and keep your raw files and Content Credentials ready as proof.

Stock and licensing markets tightened as well. Agencies screen submissions for AI content, and buyers ask for provenance before licensing editorial images. Agencies and judges increasingly need to know how to tell if a photo is AI generated before they trust it. C2PA Content Credentials give you a clean way to document a genuine capture. Major photo contests like Wiki Loves Monuments already publish detailed entry rules, and provenance requirements slot naturally into rules of this kind.

Similarly, client deliverables face the same shift. Commercial and editorial clients increasingly want assurance an image is real, especially for journalism, real estate, and product work. Provenance metadata turns the assurance into something verifiable. Big tech’s moves in photo editing show the whole industry pricing authenticity into its tools. Photographers who document provenance now will answer client questions faster than those who wait.

What Working Photographers Should Do Now

OpenAI’s update helps you most as a verification tool. The biggest gain is speed: a free public check on one of the largest AI image sources, ready in seconds. For any image arriving as a real photograph, run it through Verify and read its Content Credentials before you act on it. This habit costs almost nothing and catches the easiest fakes immediately.

Still, do not over-trust the result. The system tags OpenAI images only, metadata strips off easily, and false readings happen. A clean Verify result is not a certificate of authenticity. Where money, credit, or reputation is on the line, combine the automated check with raw files, capture data, and your own judgment.

The larger value is defensive. Provenance signals protect honest photographers as much as they expose fakes. When you attach Content Credentials to your own images, you give clients, agencies, and contest judges a verifiable record of a real capture. As more AI images carry watermarks, an unmarked genuine photograph with full provenance becomes easier to trust, not harder.

Start with three steps this week. First, bookmark OpenAI’s Verify tool and the Content Credentials Verify page. Second, turn on Content Credentials in Photoshop or Lightroom so your exports carry provenance automatically. Third, save raw files and capture data for any work headed to a contest or a paying client. Detection technology will keep changing, so treat these habits as a baseline rather than a finished solution.

Frequently Asked Questions

How accurate is AI image detection?

Accuracy varies by method and source. OpenAI reports its Verify tool rarely returns errors for OpenAI-generated images, since it reads a known watermark and metadata. Detection drops sharply for images from other AI tools or for files with stripped metadata. Treat a result as strong evidence, not absolute proof.

How to Tell if a Photo Is AI Generated Without Tools?

Manual checks still help, though they are weaker than before. Look for missing capture data in the EXIF fields, a reverse image search with no original source, or a creator with no shooting history. These checks teach you how to tell if a photo is AI generated when no watermark exists. Modern models fixed older visual tells like distorted hands, so context now matters more than the pixels themselves.

Does OpenAI’s Verify tool detect images from Midjourney or other AI generators?

No. Verify recognizes signals from OpenAI tools only. It looks for an OpenAI SynthID watermark or a C2PA manifest issued by OpenAI. An image from Midjourney, Adobe Firefly, or another generator returns no useful result, so you need separate checks for non-OpenAI sources.

Is a SynthID watermark possible to remove?

SynthID resists removal far better than metadata. The watermark survives screenshots, cropping, and many edits, since it sits inside the image pixels. Determined attackers with the right tools still degrade it, so no AI watermark is fully tamper-proof. For everyday verification, though, SynthID holds up well.

How do I add C2PA Content Credentials to my own photos?

Adobe builds Content Credentials into Photoshop and Lightroom. Turn the feature on in preferences, and your exports carry a provenance manifest automatically. The manifest records your capture and edit history. Clients and agencies then read it through the free Verify page at contentcredentials.org.

What does AI image detection mean for photo contests?

AI image detection now shapes contest judging directly. Several competitions disqualified winners in 2026 after provenance checks revealed AI origins. If you enter contests, read the rules for AI policies, keep your raw files, and attach Content Credentials to your submission. Proof of a real capture protects your entry.