Is This Image Real? A Simple Guide to AI Watermarks
TABLE OF CONTENTS
You scroll past a dramatic photo, stop for half a second, and wonder — did a person take this, or did an AI make it? That question just got easier to answer, but not as easy as you might hope.
In May 2026, OpenAI announced new provenance signals for AI-generated images, combining Content Credentials based on the C2PA open standard with an invisible SynthID watermark developed by Google DeepMind. OpenAI also published a public verification tool at openai.com/research/verify where anyone can upload an image and check for these signals.
That is real progress. But watermarks are useful signals — not magic proof. Here is what ordinary people actually need to know.
The News in 30 Seconds
- OpenAI is now embedding two layers of provenance into AI-generated images: C2PA metadata and a SynthID invisible watermark.
- SynthID, developed by Google DeepMind, has already been applied to over ten billion images and video frames across Google’s services.
- A public tool at openai.com/research/verify lets anyone upload an image to check whether it came from OpenAI tools.
- The conversation on Hacker News was lively: many readers asked whether the watermarks can be removed, and whether the detection tool requires uploading your image to a server (yes, for now).
The reason this matters for everyday users: you will increasingly see labels like “AI-generated” or “made with AI” attached to images on social media, news sites, and messaging apps. Understanding what those labels mean — and what they do not mean — helps you decide what to believe and what to share.
Three Kinds of AI Labels You May See
Not all AI image labels work the same way. There are three main types:
Visible watermarks
A logo, symbol, or corner tag added on top of the image. Easy to see, easy to crop off. Common on stock photo sites and some AI tools that add their branding to free-tier outputs. They tell you something about origin, but they are the least durable signal.
Invisible watermarks (SynthID)
A signal embedded directly into the pixels of an image by a neural network. It is invisible to the human eye and is designed to survive common edits: resizing, cropping, color adjustments, and lossy compression like saving as a low-quality JPEG.
SynthID does not produce a yes-or-no answer. The detector outputs one of three results: detected, not detected, or uncertain. That third category is important — not every image gets a confident verdict.
Metadata and Content Credentials (C2PA)
Information embedded in the file itself — not in the pixels but in the file’s metadata — that records who or what created the image, which tools were used, and what edits were made. Think of it as a nutrition label for a digital file.
The C2PA standard is backed by Adobe, Microsoft, Google, OpenAI, Sony, and major news agencies including AP and Reuters. The verification tool at verify.contentauthenticity.org can read these credentials for any compatible file.
| Label type | Can you see it? | Survives screenshots? | What it tells you |
|---|---|---|---|
| Visible watermark | Yes | Often no | That the image was processed or published by a specific platform |
| Invisible watermark (SynthID) | No | Designed to survive basic edits | That the image was likely generated by a specific AI model |
| Metadata / Content Credentials | No | Usually no — metadata is stripped by screenshots and social media re-uploads | Detailed origin info when present |
Why a Watermark Is Not Proof
OpenAI makes an important point in its announcement: provenance signals can show where an image came from, but they do not prove whether the image’s claim is true. A detected signal means the image originated from OpenAI tools — it says nothing about whether the content is accurate, taken in context, or being used honestly.
In plain words: a watermark tells you where something came from, not whether you should believe it.
Three reasons not to rely on watermarks alone:
No watermark does not mean real. Most AI-generated images online were not made with ChatGPT or Gemini. Open-source models, older tools, and images generated before watermarking was introduced carry no signal. Absence of a watermark is not evidence of a human photographer.
A watermark does not explain context. A real photograph can still be taken out of context, mislabeled with a false date or location, or cropped to mislead. The image is authentic; the story around it is not.
Screenshots strip metadata. If someone screenshots an image and shares it again, the C2PA metadata is typically gone. SynthID may survive, but the context — the caption, the source account, the date — may already be lost.
Think of it like a seatbelt. A seatbelt is genuinely useful and you should use one every time. But wearing a seatbelt does not mean you can drive recklessly. AI watermarks reduce risk; they do not eliminate it.
A Practical Trust Scale
Before you share or act on an image, run a quick mental check:
| Situation | Trust level | Suggested action |
|---|---|---|
| Official account with original photo, source traceable | Higher | Read context and date before sharing |
| Known news outlet, no AI label | Medium | Verify headline matches the image |
| Friend forwarding a screenshot with a strong claim | Low | Ask where they got it |
| Viral image with intense emotional appeal, unknown account | Very low | Do not share until verified |
| Image arrives with a request to send money or take urgent action | Very low | Stop. Verify through a completely separate channel. |
The pattern to watch for: the stronger the emotional pressure, the slower you should move. Alarm, outrage, and urgency are the conditions in which misinformation spreads fastest.
The uncomfortable part is that both sides are true: watermarks help, and they can still fail. The habits below work regardless of whether a watermark is present.
How to Check an Image Before You Believe It
You do not need specialized tools for most checks. Five habits cover the majority of cases:
1. Check the source. Who posted this originally? Does the account have a history? When was the account created? A new account with a single dramatic post is a warning sign.
2. Look for the original. Reverse image search with Google Lens or TinEye takes about thirty seconds. TinEye can show you when an image first appeared online — if a photo claimed to be from last week was indexed two years ago, something is wrong.
3. Read the text inside the image carefully. AI image generators frequently produce distorted, nonsensical, or inconsistent text in signs, posters, t-shirts, and labels. If the writing in the image looks garbled, that is a meaningful clue.
4. Check the details. Hands with the wrong number of fingers, shadows falling in different directions, reflections that do not match, backgrounds with repeating patterns that merge unnaturally — these errors are becoming rarer as AI improves, but still appear in many images.
5. Use the verification tools. Upload an image to openai.com/research/verify to check for OpenAI provenance signals, or verify.contentauthenticity.org for C2PA credentials from any supported creator or tool.
Be Careful Before Uploading Images to AI Detectors
The most common question on Hacker News about OpenAI’s new tool: “Is there no way to do this without uploading it?” Currently, there is not — for SynthID, the check happens on OpenAI’s servers.
This is worth thinking about before you upload:
Generally lower risk to upload: Public news images, social media posts from public accounts, product photos, images you found on a public website.
Think carefully before uploading:
- Screenshots containing private conversations
- Photos of people who did not consent to share them online
- Images showing your address, location, or daily routine
Do not upload:
- Passport or government-issued ID
- Medical records or health documents
- Children’s photos
- Bank statements or financial documents
- Confidential work materials
If you want to check a private image, cover or crop out any sensitive details before uploading. The goal is to verify the image itself, not to share your personal information with a third party.
When the Image Contains Foreign-Language Text
Many misleading images circulate with text in a language you cannot read. A poster, a screenshot, a notice, or a social media post — if you cannot read the words, you cannot evaluate the claim.
Translating the text inside an image is a separate step from verifying whether the image is authentic. OpenL can extract and translate text from images and screenshots across more than 100 languages, so you can understand what the image says before deciding whether to trust or share it.
This matters because translation and verification are different tasks. Translation tells you what the image says. Verification tells you whether the image should be believed. Both steps are worth doing, in that order.
For a practical walkthrough of image translation tools, see how to translate everyday images and how to translate text from images and photos.
FAQs
Can an AI watermark prove that an image is fake?
No. An AI watermark — if detected — indicates that an image was generated by a specific AI tool. It says something about origin. It does not mean the image depicts something false, and it says nothing about the caption, the context, or how the image is being used.
Does no watermark mean the image is real?
No. Many AI-generated images carry no watermark at all: they were made with open-source models, generated before watermarking was introduced, or had their signals stripped through screenshots or re-uploads. Absence of a watermark is not evidence of human photography.
Can AI watermarks be removed?
Researchers and developers are actively studying the limits of watermarking systems. SynthID is designed to be more durable than metadata-based approaches, but no watermarking system is guaranteed to be permanent. The more important point for everyday users: even if a watermark is present, it does not prove an image is trustworthy; and if a watermark is absent, it does not prove the image is human-made.
How can I tell if an image was made by AI?
Start with the source, not the pixels. Look for the original post, check whether the account is trustworthy, search for older copies of the image, and read any text inside the image carefully. Visual clues like strange hands, broken text, mismatched shadows, or impossible reflections can help, but modern AI images may not show obvious mistakes.
How can I translate text inside an AI image?
Use an image translation or OCR tool to extract and translate the text. OpenL’s image translator supports over 100 languages and works with screenshots, photos, and scanned documents. Once you can read the text, you can evaluate whether the claim it makes holds up.
Should I trust a photo just because it has a “made with AI” label?
A label tells you the image was generated by AI. It does not tell you whether the claim it is being used to support is true or false. Treat the label as useful context, not as a verdict.
Conclusion: Trust Slowly
AI watermarks are genuine progress. For the first time, a growing share of AI-generated images carry signals that ordinary people and platforms can check. That is worth acknowledging.
But watermarks are one layer of a much larger problem. Real photos get misused. Context gets stripped. Screenshots lose metadata. Old images get recycled with new captions.
The most reliable habit is simple: slow down before you share. Check the source. Check when the image first appeared. Read the text inside it. If the image is in a language you cannot read, translate it first. Then decide whether it deserves your trust.
Sources
- OpenAI: Advancing content provenance for a safer, more transparent AI ecosystem — Official announcement of C2PA and SynthID adoption, and the public verification tool
- OpenAI Verification Tool — Public tool to check whether an image was generated by OpenAI tools
- Google DeepMind: SynthID — Overview of SynthID technology, covering images, audio, text, and video
- Content Authenticity Initiative: verify.contentauthenticity.org — Public C2PA verification tool for images with Content Credentials
- Hacker News discussion: OpenAI Adopts Google’s SynthID Watermark for AI Images with Verification Tool — Community discussion on watermark durability, removal research, privacy of upload-based verification, and limits of closed-source detection
- SynthID-Image paper (arXiv 2510.09263) — Technical paper on the SynthID-Image deep learning watermarking system
- C2PA: Coalition for Content Provenance and Authenticity — Open technical standard for content credentials used by Adobe, Microsoft, Google, OpenAI, and others
- TinEye Reverse Image Search — Tool for tracing the original source and earliest appearance of an image
