Multimodal and generative AIMiddleLesson 2495 min read

Watermarks and content credentials

Invisible watermarks such as SynthID and signed content credentials (C2PA) help show where an image, video or sound came from. Learn how they work, and what they cannot prove.

✦ Ask the tutor🧪 Lab · 20 min📝 Lesson test
Lesson in motion

In 60 seconds

Watermarks and content credentials

Invisible watermarks such as SynthID and signed content credentials (C2PA) help show where an image, video or sound came from. Learn how they work, and what they cannot prove.

1/5
In simple words
A ₹500 note has a hidden watermark you can see against the light. AI pictures can carry hidden marks too, and some carry a signed "label" that says where they came from. But most pictures carry neither, so a missing mark tells you very little. You still need other checks.
As AI-made images, video and sound get better, looking closely is often not enough. Two families of tools try to help: watermarks hidden inside the content, and content credentials attached to the file. They work differently and fail differently.

Invisible watermarks

An invisible watermark is a pattern built into the pixels or sound itself, too subtle for people to notice, that a special detector can find. Google's SynthID is a well-known example, used to mark content made by Google's AI tools. The generator adds the watermark as it creates the content; the company's detector can later check for it.
  • Strength: it is in the content itself, so it can survive some common changes such as resizing or compression.
  • Limit: heavy edits, cropping, re-recording or deliberate attacks can weaken or remove it.
  • Limit: a detector usually only finds its own company's watermark. Content from a tool that adds no watermark shows nothing.

Content credentials (C2PA)

C2PA (Coalition for Content Provenance and Authenticity) is an open technical standard for content credentials. A content credential is a signed record attached to a file that can say who made it, with what tool, whether AI was used, and what edits were made. The signature is made with cryptography, so if someone changes the record or the content, a check can show that it no longer matches.
  1. 1

    Create

    A camera, editing app or AI generator that supports C2PA writes a record: made by, when, with what tool, AI used or not.
  2. 2

    Sign

    The record is signed digitally by the tool or organisation.
  3. 3

    Edit

    Supporting editors add new entries on top, keeping a history.
  4. 4

    Verify

    Anyone can check the file with a verify tool: is the signature valid, who signed it, what does the history say?

Watermark vs content credential

Invisible watermarkContent credential (C2PA)
Where it livesInside the pixels or soundIn metadata attached to the file
What it saysUsually just "made by this tool" (a yes or no signal)A detailed, signed history
Easy to remove?Harder, but possible with heavy edits or attacksEasy: a screenshot or many re-uploads can drop it
Who can checkUsually the company that added itAnyone with a C2PA verify tool
Because each has gaps, some systems combine them: a watermark can help find the content credential again after the metadata has been stripped.

What these tools cannot prove

Wrong conclusion
  • "No watermark found, so it is real"
  • "No content credentials, so it is fake"
  • "It has credentials, so everything in it is true"
Right conclusion
  • Many AI tools add no watermark, and watermarks can be removed
  • Most real photos today have no credentials, and they are easily stripped
  • Credentials show who signed and what they claimed, not that the scene is true or fair
Watch out
Absence proves nothing. A missing watermark or credential is the normal case for most content today. Always combine these checks with the source, context and trusted fact-checkers.
Note
Several governments, including India's, have discussed or introduced rules about labelling AI-made content, and platforms have their own labelling policies. These change, so check the current rules for your case.
Try this. Take a photo with your phone and send it to yourself on a messaging app. Does the received file still have its original details, such as the date and camera? This shows how easily metadata can be lost.

Watch and read more

C2PA: Coalition for Content Provenance and AuthenticityC2PA · reading

Lab

A short provenance report on three images, checked with a public content credentials verify tool.

~20 min

The problem

Search the web for a public "Content Credentials verify" tool. Several free ones exist. Read its privacy note before uploading anything, and use only images you have the right to share, with no private people or documents.

  1. Collect three images: (a) a photo straight from your own phone; (b) the same photo after sending it through a messaging app or taking a screenshot; (c) an image from a source that says it adds content credentials, such as a news or stock image page that shows a credentials label, or an image you made with an AI tool that says it adds them.
  2. Check each in the verify tool. Record: credentials found or not, who signed, what tool, whether AI use is stated, and the edit history.
  3. Write one line per image on what you can and cannot conclude.

You are done when

Hard questions

Try to answer before you reveal. If you can answer these, you understood the lesson.

Q1Why is it easier to prove content is genuine than to prove content is fake, with these tools?Reveal

A valid signature from a trusted source is positive evidence: someone accountable stands behind the file's history. But fake content simply leaves no mark, or has its marks removed. You cannot prove a negative from a missing signal. So provenance tools work best as a way for honest creators to show their work is genuine, with other checks used for everything else.

Q2A political party posts a video with valid content credentials signed by itself, stating it is unedited. What should a careful viewer still check?Reveal

The credential proves that the party signed that claim and the file has not changed since. It does not prove the scene was not staged, that it was not cut from a longer clip before signing, or that the context given is fair. Check who else has reported the event, whether other footage exists, and what independent fact-checkers say.

Please sign in to continue.

Questions people ask

If content credentials are so easy to strip, why use them?

Because when they are present and valid, they give a strong, checkable history. News agencies, camera makers and editing tools can use them to show that genuine content is genuine. They help most as more tools and platforms keep and display them.

Can I check for a SynthID watermark myself?

Google has offered ways to check for SynthID in some of its products and tools, but access and coverage change. Look for Google's current official information. Remember that a negative result does not mean the content is real.

Do content credentials reveal my personal data?

They can include details such as the creator's name, the device, or edit history, depending on the tool and settings. Check what your camera or app adds before sharing sensitive photos, just as you would check location data.

Can someone fake content credentials?

They cannot easily forge a valid signature from a trusted signer. But someone can sign misleading content with their own key, or take a real photo out of context. That is why you check who signed it and whether you trust them.

Lesson test

5 questions. Get 3 right (60%) to pass and complete this lesson.

Sign in to take the test and save your progress