How to Tell If an Image Was Made With AI, Using Content Credentials

A photo of a product defect arrives in a customer complaint. A striking chart image is circulating in a team chat. A vendor sends a "photo" of a finished installation. Increasingly, it is fair to ask whether an image came from a camera or a generator. Looking for extra fingers or warped text used to be the advice, but image models have improved, and visual tells are unreliable. A better starting point is the provenance information that more tools now attach to images: Content Credentials and invisible watermarks.

1. Content Credentials: a signed history attached to the file

Content Credentials come from an open standard developed by the Coalition for Content Provenance and Authenticity (C2PA). The C2PA's FAQ describes them as a cryptographically secure way to record a piece of content's history: how it was created, what tools were used, when and where it was made, and how it has changed.

Key points from the C2PA FAQ:

• Tamper-evident. Each credential is digitally signed and includes cryptographic hashes of the asset and its provenance data. Any modification breaks that link, which signals tampering.
• Not just for AI. OpenAI's help article notes that camera makers and news organizations also use the standard to certify where their media came from.
• Not DRM. The C2PA says Content Credentials record provenance without restricting how content is used.
• Not an identity tag. The core specification does not attribute content to individual people or organizations, to stay privacy-preserving.

The weakness is that this information travels as metadata, and metadata can be removed. OpenAI's article says Content Credentials can be lost through everyday actions such as taking a screenshot or converting a file. The C2PA FAQ describes "durable" credentials that use watermarking or fingerprinting to help rediscover stripped credentials, but not every file will have them.

2. Invisible watermarks: a signal inside the pixels

The second approach embeds a signal in the image itself. Google DeepMind's SynthID page describes an invisible watermark added at the moment content is created, designed to hold up against changes like cropping, adding filters and lossy compression. Google says SynthID is used across its generative AI consumer products.

OpenAI's article explains why it uses both methods for images. Metadata can carry detailed information but is easily removed; a watermark carries less context but, because it is part of the content, may survive some transformations that strip metadata. According to OpenAI, images generated with ChatGPT, Codex or the OpenAI API include both Content Credentials and SynthID watermarks.

Watermarks have limits too. OpenAI notes that heavy cropping or file compression can make a watermark undetectable.

3. How to check an image

Several tools can read these signals (descriptions as of October 2026, from the vendors' own pages):

• Google's tools. The SynthID page says you can upload an image, video or audio clip to Gemini, and ask whether it was created or altered by Google AI; Gemini, Search and Chrome will check for a SynthID watermark. Google also describes a SynthID Detector portal, now open to everyone, that checks for watermarks from Google and partners including OpenAI.
• OpenAI's verification page. OpenAI's article describes a page where you upload a supported image or audio file to check for OpenAI provenance signals, such as a SynthID watermark or a trusted C2PA manifest.
• Content Credentials viewers. Any tool that reads C2PA data can display the signed history if it is still attached.

For the most reliable result, OpenAI recommends checking one file at a time and using the original exported file where possible, without cropping or converting it first.

4. What a result does and does not tell you

This is the part most people get wrong.

A positive result is not a verdict on truth. OpenAI says its verification tool does not confirm that content is accurate, unedited, legally owned, or presented in the correct context, and does not identify who created it or why.

A negative result does not mean "made by a person." OpenAI lists several reasons an AI image might show no signal: it was created before signals were added, came from an unsupported product or file type, had its metadata stripped during upload or editing, or had its watermark degraded by compression, cropping or other changes. OpenAI also says its tools are not designed to detect content from other companies' AI models.

Provenance is not a disclosure label. OpenAI notes that watermarks and C2PA metadata are machine-readable information and do not replace visible labels or other notices that may be required.

5. A practical routine for work

1. Ask for the original file rather than a screenshot or a version forwarded through a chat app, which may have stripped the metadata.
2. Check for Content Credentials and, where available, a watermark, using the tools above.
3. Treat the result as one clue. Combine it with who sent the image, when, and whether the details match other evidence.
4. Verify through another channel when the stakes are high, for example by asking for a second photo from a different angle or confirming with the sender by phone.
5. Label your own AI images. If your team publishes generated images, keep the original files with their credentials intact and follow your organization's disclosure policy.

6. Key takeaways

• Content Credentials, based on the C2PA standard, are a signed, tamper-evident record of how an image was made and edited.
• Metadata can be stripped by screenshots or file conversion, so watermarks such as SynthID are used alongside it.
• OpenAI says its generated images carry both Content Credentials and SynthID watermarks.
• A detected signal does not prove an image is accurate or in context, and no signal does not prove a person made it.
• Ask for original files and treat provenance as one piece of evidence, not the whole answer.