Sometimes, but only in one direction. If an image carries intact, signed provenance data that says an AI model made it, that is strong evidence. If an image has no metadata, or ordinary camera-style metadata, you have learned very little: missing metadata does not prove an image is AI-generated, and present metadata can be edited or copied.
A quick note on scope: this site checks text only, and we have no image tool. This guide explains the evidence that exists around images so you can weigh it sensibly.
Key takeaways
- Metadata is a label attached to a file. It describes the file; it does not analyse the pixels.
- Missing EXIF is normal for screenshots, re-saved files and anything uploaded to many platforms. It is not a sign of AI.
- Plain EXIF and XMP fields can be edited or removed with free tools, so their presence is not proof of anything either.
- Signed C2PA Content Credentials and the IPTC “digital source type” field can positively identify AI output when they survive. They record provenance, not truth.
- Invisible watermarks such as SynthID are vendor-specific. A result applies only to content from the vendor that built the watermark.
What is image metadata, and where does AI come into it?
Image files can carry information beside the picture itself. The best-known type is EXIF, a standard that cameras and phones use to record details such as make and model, date and time, exposure settings and, on devices with GPS, location. Other containers such as XMP and IPTC fields hold descriptive and editorial information: captions, credits, rights and the software or process behind the file.
A camera writes EXIF because a physical capture happened. A generative model has no lens or sensor, so it has no reason to write those fields. What a tool can write is a statement about itself: a software name, a text chunk with the prompt, or a formal provenance record. For example, the popular Stable Diffusion web interface saves generation parameters into the PNG as a text chunk. That is helpful when the file is untouched and useless once the chunk is gone.
So metadata can carry two kinds of signal: camera-style fields that suggest a capture, and tool-written fields that suggest generation or editing. Neither is attached to the image in a way that cannot be separated from it.
What does missing EXIF mean?
By itself, almost nothing. Plenty of ordinary routes leave a human-made photo with no EXIF:
- A screenshot is a new file made by the operating system. It records the screenshot, not the original image’s metadata.
- Many editing, export and “save for web” workflows drop metadata on purpose, to shrink files or protect location privacy.
- Messaging apps and social sites often remove embedded metadata on upload. The IPTC, the news-industry standards body, tested photo-metadata handling on social and sharing sites in 2013, 2016 and 2019 and reported that several services, including Instagram and Twitter, did not preserve it. Those tests are years old and platform behaviour changes, so treat them as evidence that stripping is common, not as a current list.
- OpenAI’s help center says Content Credentials “can be lost through everyday actions, such as taking a screenshot or converting a file”, which is why it pairs them with SynthID watermarks for images (OpenAI, read 5 October 2026). It adds that watermarks can also fail after heavy cropping, compression, or other extensive changes, and that provenance signals do not guarantee content is accurate or unedited.
Meta has also said in public that people can strip out the invisible markers its labels depend on.
The same logic runs the other way. An AI image that has been screenshotted or uploaded to a stripping platform also has no metadata. The empty field does not separate the two cases, so an image with no EXIF is simply unresolved.
Can present metadata be trusted?
Not on its own. Plain EXIF has no built-in tamper protection, and widely used tools such as ExifTool are made for reading, writing and editing it. Anyone can add a camera make, change a timestamp or delete a “Software” field. Wikipedia’s summary of the Exif format also notes that editors can corrupt or drop its data by accident.
Two practical points follow:
- Camera-style EXIF on a file is not proof of a real photo. It can be copied from a real photo onto any image.
- An AI-tool tag in EXIF or XMP is a claim by whoever wrote the file. Often that is the generator, which makes it useful. It is still unsigned text.
Editing matters too. Opening and re-saving a photo in software can add an editor’s name or remove fields. An AI-assisted edit, such as generative fill or background removal, can leave a real photo with an editor tag and no other change you can see. A software tag says something touched the file, not what the pixels contain.
What is the IPTC digital source type?
The IPTC maintains a controlled vocabulary called Digital Source Type, which lets a file declare how it was made. The values most relevant here, with the IPTC’s own definitions in its digital source type vocabulary, are:
| Value | IPTC definition (paraphrased) |
|---|---|
trainedAlgorithmicMedia |
Media created algorithmically by an AI model trained on captured content |
compositeWithTrainedAlgorithmicMedia |
A generative model was used to augment, correct or enhance, such as inpainting or outpainting |
compositeSynthetic |
A mix of several elements, at least one of which is generative AI |
algorithmicMedia |
Created purely by an algorithm with no sampled training data, such as a formula-based image |
When a generator or editor writes trainedAlgorithmicMedia, that is a clear, standardised statement. Meta says it reads the “AI generated” information in the IPTC and C2PA standards to label images. But it is still a metadata field. It reflects what the producing software chose to write, and it can be removed or, in an unsigned file, altered.
What are C2PA Content Credentials, and what do they prove?
C2PA (the Coalition for Content Provenance and Authenticity) defines a signed record, called a manifest, that can travel with a file. It lists how the asset was made and edited and is signed by the party that made the claim. Under the C2PA specification, a “hard binding” uses cryptographic hashes so a validator can tell whether the asset has been modified since signing. That is the key difference from plain EXIF: changes become detectable.
What a valid credential does not do matters as much. The C2PA FAQ frames Content Credentials as provenance, not truth. They record how content was created and changed so that people can judge how reliable it is for their purpose. The trust comes from the identity of whoever signed. So:
- A valid manifest saying “generated by an AI tool” is strong evidence for that image, from that signer.
- A valid manifest from a camera or editing tool says what that signer attests. It does not certify that the scene was real or the claim true.
- A manifest with a signer you do not recognise tells you less than one from a signer you can place.
- No manifest tells you almost nothing. The FAQ states that manifests, though typically embedded, can be separated from the asset.
C2PA’s answer to separation is a “soft binding”: an invisible watermark or a content fingerprint that can help rediscover the credential even after it has been stripped from the file. That helps when it is deployed, but it does not turn absence into evidence. Where it is not deployed, or the match fails, the reader is back to an unresolved image.
What is SynthID, and what does it detect?
SynthID is Google’s watermarking technology. According to Google DeepMind, it embeds imperceptible watermarks into AI-generated images, audio, text and video, and is designed to withstand modifications such as cropping, filters and lossy compression. Google’s SynthID Detector announcement says it covers content from its Gemini, Imagen, Lyria and Veo models and was initially rolled out to early testers on a waitlist, and its November 2025 post on the Gemini app says users can upload an image and ask whether it was generated or edited with Google AI.
Read those statements narrowly:
- It is vendor-specific. SynthID checks for Google’s own watermark. A negative result does not say an image is not AI-generated, only that this check did not find Google’s mark. Other generators are outside its scope. OpenAI says it also embeds SynthID watermarks in its own images and checks them with its own verifier at openai.com/verify, which, like Google’s, looks only for OpenAI’s signals and is “not designed to detect content generated by other AI models” (OpenAI).
- Robustness claims are the vendor’s. Google describes the design target. Its research paper, SynthID-Image, claims state-of-the-art robustness to common image perturbations, and says it documents threat models and challenges. We did not find independent, public test results for SynthID specifically, so we cannot give you a detection or removal rate.
- Watermarks in general are not unbreakable. A 2026 study, Vanishing Watermarks, found that diffusion-based editing nearly erased the watermarks it tested (StegaStamp, TrustMark and VINE) while keeping the image looking good. It did not test SynthID, so it does not show SynthID can be removed. It does show why “robust” depends on which edits are tried.
- Google says it plans to add C2PA support to its verification, so checks may cover non-Google content in time. That is a stated plan, not something we verified as live.
How reliable is each metadata signal?
| Signal | What it can tell you | What it cannot tell you | Survives screenshot or social upload? |
|---|---|---|---|
| Camera EXIF (make, model, date, GPS) | A device may have recorded it | That the image is real; fields are editable and copyable | Often removed |
| No EXIF at all | Metadata was never written or was removed | Anything about AI use | Not applicable |
| Software or prompt tag (EXIF, XMP, PNG text) | The producing tool may have declared itself | That it is unaltered; it is unsigned text | Often removed |
| IPTC digital source type | A declared origin such as trainedAlgorithmicMedia |
That the declaration is true or intact | Often removed |
| C2PA Content Credentials | A signer’s tamper-evident record of origin and edits | Whether the scene is true; anything if the manifest is absent | Can be separated from the file |
| Invisible watermark (e.g. SynthID) | Content from that vendor’s models | Anything about other generators; “not found” is not “human” | Designed to survive some edits; not guaranteed |
“Often removed” reflects documented platform behaviour from the sources above, not a guarantee for any specific site today.
A practical way to use metadata
- Look for signed provenance first. Check for Content Credentials with a verifier from a source you trust. A valid record from a recognisable signer is the strongest metadata evidence you will get.
- Read declared source fields as claims. An IPTC value or a generator tag is useful when it matches other evidence and weak when it stands alone.
- Treat empty metadata as a prompt to look elsewhere. Ask where the image came from and whether it passed through a screenshot, messaging app or platform upload.
- Find the original. The earliest upload, the photographer’s account or a reverse image search often says more than any field in the file.
- Use a vendor watermark check only for that vendor. A hit is informative; a miss is not.
- Never rely on one signal. How to tell if an image is AI-generated covers the checks that go beyond metadata, and what AI image detectors can and cannot tell you covers classifiers that look at the pixels.
Where this fits for text
A pasted paragraph has no EXIF and, in practice, almost never arrives with a provenance record. Text detectors, including the one on this site, work from the writing itself, and they share the core limit described here: a result is evidence to weigh, not a verdict. If you came here from a question about writing, how to tell if text is AI-generated and can AI detectors be wrong are the right next reads.
FAQ
Does a photo with no EXIF data mean it is AI-generated?
No. Screenshots, messaging apps, social uploads and many export settings remove EXIF from real photos, and AI images that have been re-saved or uploaded look the same. Missing EXIF is a reason to look for the original source, not a finding about how the image was made.
Can EXIF be faked?
Yes. EXIF fields are not signed, and tools such as ExifTool are built to write and edit them. A real camera’s make, model and timestamp can be copied onto any image. Present EXIF is a claim in the file, and it needs support from other evidence.
Do all AI images include C2PA Content Credentials?
We could not find a source saying so, and you should not assume it. Some companies say they add C2PA data. OpenAI’s help center states that images from ChatGPT, Codex and its API carry both Content Credentials and SynthID watermarks, and Google has said images from one of its image models carry C2PA metadata. Coverage differs by tool, model and export path, and the data can be removed or never be written by tools outside these programmes (OpenAI).
Can SynthID detect images from any AI generator?
No. Google describes SynthID as detecting its own watermark in content made with Google’s AI models. A negative result tells you the check found no Google SynthID watermark. It does not tell you whether another tool made the image.