Photos and video
AI can make fake photos and video. Here are three ways to check a picture.
A person has checked every link on this page.
Search with the picture
A reverse image search uses a picture as your search. Google Lens is one way to do it, and TinEye is another. It shows where else the same or a similar picture appears.
This can catch a common trick: an old photo passed off as new. If the same picture was online years ago, it isn't from today's news.
It isn't proof, though. It only finds what the search engine has seen. If it finds no match, that doesn't mean the picture is new. Different tools find different things, so try more than one.
Check where and when
Look closely at the picture. Are there signs, buildings, roads or hills? Do they match a map or a satellite picture of the place it claims to show? Do the weather and shadows fit that day? When was it first posted, and by who?
Files can also hold hidden details like the date, but those can be missing or wrong. And sometimes you just can't know for sure. It's fine to say "I can't tell."
Look for content credentials
Some photos and videos carry a label called content credentials. Think of a food label, but for a picture.
A content credential can show who made a file and how it was changed.
Where this comes from (2 sources)
This comes from a news report
Microsoft's research argues AI media authentication doesn't work reliably
Provenance data records who made a file, with what tool, and what edits were made.
recording who created it, what tool was used, and what edits were made
Archived copy:Wayback Machine
A person has checked this link.
This comes from the company's own report or page
How it works (Content Authenticity Initiative)
Content credentials are built to show the source and history of media.
certifying the source and history (or provenance) of media content
Archived copy:Wayback Machine
A person has checked this link.
One fact below comes from one test. The Washington Post made it. Two of the links under that fact are not two tests. One copies the Post's report and the other reports on it. The third link is from the group behind content credentials. It says the label can be removed.
Know the limits:
The label can be removed: Apps often take content credentials off. In one news test, a video with a content credential was posted on eight apps, and all eight took it off.
Where this comes from (3 sources)
This comes from a news report
Washington Post test of AI video labels (AI Incident Database copy)
This reproduces the same work as The Washington Post. It is not a separate check.
The Post posted one clip with a content credential on eight platforms and none kept it.
No platform that The Post tested kept the Content Credentials data on the video
Archived copy:Wayback Machine
A person has checked this link.
This comes from a news report
Remedies to identify AI video and why they have failed so far
This reports on the same work as The Washington Post. It is not a separate check.
In the Washington Post test, none of eight platforms kept the content credential.
None kept the C2PA metadata intact or made it visible to users
Archived copy:Wayback Machine
A person has checked this link.
This is an original document from the time
The group behind content credentials says the label can be removed from a picture or video. The eight-app test is a separate detail. It comes from one Washington Post test.
Yes it can. That is why the C2PA specification includes the concept of durable Content Credentials
Archived copy:Wayback Machine
A person has checked this link.
- Adding a label is a choice. No label doesn't prove a file is fake.
A content credential shows a file's history. It does not prove that what the file shows is true.
Where this comes from (2 sources)
This comes from a news report
Microsoft's research argues AI media authentication doesn't work reliably
Microsoft's report says a valid content credential does not prove the content is true.
even validated provenance data doesn't prove that content is true
Archived copy:Wayback Machine
A person has checked this link.
This is an original document from the time
The C2PA says content credentials do not judge if content is true.
Content Credentials do not provide value judgments about whether a given set of provenance data
Archived copy:Wayback Machine
A person has checked this link.
Some researchers warn that content credentials are not strong enough yet to rely on when a lot is at stake.
Where this comes from (2 sources)
This comes from a research or documentation centre
Verifying Provenance of Digital Media: Why the C2PA Specifications Fall Short
The researchers say content credentials should not yet be relied on when a lot is at stake.
it should not yet be relied upon for high-stakes uses such as financial disclosures, journalism, or legal evidence
Archived copy:Wayback Machine
A person has checked this link.
This comes from a news report
Microsoft's research argues AI media authentication doesn't work reliably
Microsoft's report says no single method for proving media is real is reliable on its own.
no single method is reliable on its own, and even combined approaches have limits
Archived copy:Wayback Machine
A person has checked this link.
More to read
The facts above have their sources under "Where this comes from." These guides go further.
- Reverse image search. Google Search Help and Bellingcat.
- Where and when. First Draft photo guide, Verification Handbook and Bellingcat on shadows.
- Content credentials, from the groups behind them. C2PA FAQs, C2PA explainer and Content Authenticity Initiative.
- Content credentials, from outside those groups. AI Incident Database and arXiv 2604.24890.