Is it AI? How to tell if a photo or video was made with AI

No method can tell you for certain whether a photo or video was made with AI. What works best is a check in three steps, in this order: the context (who posted it, and whether reliable sources report the same thing), then any label or hidden marker that records where the file comes from, and only last the picture itself, where odd details are weak hints at most. A missing label proves nothing.

This site is written with AI tools, including Claude by Anthropic. This page uses only what makers, public authorities and published studies say, with links. Labels and tools change often; everything was checked on 31 October 2026.

Why can’t you just tell by looking?

Pictures, videos and voices made by generative AI, the kind of AI that creates them from a written request, are now hard to tell from real ones. Studies find that people do little better than guessing.

In a study published in the journal PNAS in February 2022, Sophie Nightingale and Hany Farid asked 315 people to sort real faces from computer-made ones. They got 48.2% right, about what tossing a coin gives. A second group, first shown which flaws to look for and told after each answer whether it was right, reached only 59%.
A study of 3,002 people in the United States, Germany and China, presented at the IEEE Symposium on Security and Privacy in 2024, found that most were “simply guessing” on pictures, voices and texts, and rated AI-made media as more human-like.
In an online quiz by Microsoft’s AI for Good Lab, about 287,000 answers from over 12,500 people were right 62% of the time. This is a preprint, not yet checked by other scientists, and its authors think the figure probably flatters people today.

The 2022 faces came from tools that are old today, but all three studies agree: a careful look is not enough, and learning the typical flaws helps little. That is why the look at the picture comes last.

Check in this order: context first, then labels and hidden markers, the picture itself last. An odd detail only sends you back to steps 1 and 2, and a missing label proves nothing.

Step 1: Who posted it, and does anyone else report it?

Who posted it first? A person or newspaper you know, or a page you have never heard of?
Do reliable sources show the same thing? A flood in a real city, or a famous person saying something shocking, will be reported by newspapers and broadcasters. If none does, wait.
Does it push you to act fast? Share now, donate now, buy now: hurry is a reason to slow down.

Then run a reverse image search, which finds the same picture elsewhere on the web: on google.com, click Search by image and upload the file, as Google’s help page explains, or in Chrome right-click a picture and choose Search with Google Lens. For a video, use a screenshot of a clear moment.

In some countries Google’s “About this image” also shows when Google may have first seen a picture. A photo that turns up years earlier, or with a different story, has been taken out of context, AI or not.

Step 2: Does it carry an AI label or Content Credentials?

The origin of a file can be recorded in three ways: a label added by the platform, Content Credentials stored inside the file, or an invisible watermark such as SynthID mixed into the picture itself.

Labels on YouTube, Facebook, Instagram and TikTok

YouTube requires creators to disclose realistic content made or meaningfully changed with AI, such as a real person appearing to say something they never said. The label appears in the expanded description, under “How this content was made”, and for realistic videos may also appear on the player. YouTube may add it itself when the video was made with its own AI tools, carries Content Credentials saying it was made with AI, or its systems detect AI. Some videos instead show “Captured with a camera”, meaning unedited footage from a camera that records its origin; YouTube says its absence does not mean a video was altered.
Facebook, Instagram and Threads: Meta says it adds an “AI info” label when it detects industry-standard signals that AI was used, or when the person posting says so. Since September 2024, for content only edited with AI tools, the label sits in the post’s menu instead of on the post.
TikTok: its Community Guidelines, in force since 24 September 2026, require creators to label AI-generated or significantly edited content showing realistic scenes or people; unlabeled content “may be removed, restricted, or labeled”.

Content Credentials

Content Credentials are information stored inside a photo, video or sound file about how it was made and edited. The standard comes from C2PA, a coalition that includes Adobe, Google, Meta, Microsoft, OpenAI and Sony and compares the credentials to “a nutrition label for digital content”. To check a file, upload it to the free Verify site of the Content Authenticity Initiative, which shows the record if there is one.

But the credentials are not meant to show whether a piece of content is “real”: they say how it was made, not whether what it shows happened. And stored information like this “can be removed deliberately or accidentally”: screenshots, photos of a screen and re-recordings remove it.

SynthID and the Gemini app

SynthID is an invisible watermark, built to survive cropping, filters and compression, that Google puts into images, video and audio made by its AI. Since 7 October 2026 Google’s SynthID Detector is open to everyone; Google says it checks images, video and audio for SynthID marks from Google and from partners including OpenAI, NVIDIA and Kakao.

In the Gemini app, signed in, you can upload a picture, a video shorter than 90 seconds or an audio file and ask “Is this AI-generated?”. Gemini looks for SynthID and, on gemini.google.com and in the Android app, for Content Credentials, with about 10 checks of each kind in any 24 hours as of 31 October 2026. Its help page still says: “Gemini can currently only recognize content created by Google AI tools.” Either way, finding no watermark does not mean no AI was used.

What does it mean if there is no label?

Nothing, either way. A label appears only when a creator declares AI or when the file still carries a marker the platform can read. People who want to deceive do not declare, not every AI tool adds a marker, and markers are lost in screenshots and re-recordings. Google says the same of SynthID: when no watermark is found, the file “could have been created by other AI systems”.

A label can also mislead. Meta renamed “Made with AI” to “AI info” in July 2024 because merely retouched photos were getting it, and Google warns that the details in About this image “can be modified with image generation and editing tools”. A label is a fact to weigh, not a verdict.

Step 3: Which details are worth a second look?

Google’s Gemini help page lists flaws common in AI pictures and videos:

writing in the picture that is garbled or inconsistent;
mistakes in small details such as hands, teeth or objects in the background;
shadows and light that do not match;
patterns that repeat in an unnatural way.

Use them as hints that send you back to steps 1 and 2, never as proof: a real photo can have odd shadows, and a good fake may have none of these flaws. Voices are no easier: in the 2024 study most people were guessing on audio too. The US Federal Trade Commission’s advice fits voice messages as well: “Don’t trust the voice.”

What should you do if you think it is fake?

Do not share it, not even to ask friends whether it is real.
Report it on the platform. On YouTube, tap More under the video, then Report, and pick the reason. YouTube says reports are anonymous and reviewed, not taken down automatically.
If the fake shows you, YouTube’s privacy complaint process lets you ask for removal of AI content that realistically looks or sounds like you; YouTube weighs labeling, parody or satire, and public interest.
If a voice or video of someone you know asks for money, stop. The FTC advises calling the person back on a number you know is theirs, or reaching them through family or friends. If you have already paid, see what to do in the first 30 minutes.

What does EU law say about deepfakes?

The EU AI Act defines a deepfake as AI-made or AI-changed image, audio or video resembling real people, places or events that would falsely appear authentic. Its Article 50, which applies from 2 August 2026, sets two duties:

Companies whose AI makes images, video, sound or text must mark it in a way machines can read, so it can be detected as artificial. Tools already on the market before 2 August 2026 have until 2 December 2026.
Whoever uses AI to make a deepfake must say clearly that it is artificial, at the latest when people first see it. For evidently artistic, satirical or fictional work, a mention that does not spoil the work is enough. This duty covers people and organizations using AI in their work: the Act leaves out purely personal, non-professional use.

The law does not stop people who break it, so the three steps still apply. For your own case, ask a lawyer.

Where to go next

Generative AI: how AI makes new pictures, videos and texts.
EU AI Act: the whole law and its dates.
What is an AI agent?: AI that clicks and types for you, where checking before you trust matters too.

Common questions

How can I tell if a video on YouTube is AI?

Open the full description of the video and look for the section "How this content was made"; for realistic content a label may also appear on the player. YouTube requires creators to disclose realistic content made or changed with AI, and adds the label itself in some cases. A video without a label is not proof that it is real.

How do I know if a photo on Facebook or Instagram was made with AI?

Meta adds an "AI info" label when it detects industry signals that AI was used or when the person posting says so. For pictures only edited with AI, the label sits in the post's menu rather than on the post. Many AI pictures carry no such signal, so check who posted it and search for the image elsewhere too.

Is there a free tool that tells if an image is AI?

There are free tools that check specific markers, but none catches every AI picture. Google's SynthID Detector and the Gemini app look for SynthID, the invisible mark Google puts on what its AI makes and some other companies now use too; the Verify site of the Content Authenticity Initiative reads Content Credentials. If they find nothing, the file may still come from another AI tool or have lost its markers.

Does counting fingers still work to spot AI pictures?

Only as a weak hint. Google's own help page still lists mistakes in hands, teeth, writing and shadows as things to look for, but studies find that people mostly guess, and in one study people taught which flaws to look for scored 59%, against 48% for people who were not. Check who posted the picture and look for labels first.

If a photo or video has no AI label, is it real?

No. Labels appear only when the creator declares AI or when the file still carries a marker the platform can read, and markers are often lost when a file is saved again, screenshotted or re-recorded. A missing label tells you nothing either way.

Do deepfakes have to be labeled in the EU?

Since 2 August 2026 the EU AI Act requires anyone who uses AI to make a deepfake for professional purposes to say that the content is artificial, with a lighter duty for clearly artistic, satirical or fictional work. Companies whose AI makes images, video, sound or text must also mark it so that it can be detected; tools already on sale before that date have until 2 December 2026. The law does not stop people who break it, so the checks on this page still apply.

Topics: AI images, AI video, deepfakes, Content Credentials, SynthID, fact-checking