The Labels YouTube Shows, and What Each Proves
YouTube has two kinds of signal, and they point in opposite directions. One says a video was made or changed with AI. The other says a video came straight off a camera. Neither appears on most videos, and that matters more than anything else in this guide.
| Signal | Where you see it | Who adds it | What it tells you | What it does not prove |
|---|---|---|---|---|
| Altered or synthetic content, on the player | On the video itself | The creator, or YouTube's systems | Realistic content in the video was made or meaningfully changed with AI | Which part, or how much |
| Altered or synthetic content, in the description | Expanded description | The creator, or YouTube's systems | AI content that is not photorealistic, or is animated | That the rest of the video is untouched |
| Captured with a camera | "How this content was made", in the expanded description | Added when the video was recorded with C2PA 2.1 tools and not edited in sound or picture | The footage came from a camera and was not altered after | That a video without it is fake |
| No label at all | Nowhere | Nobody | Nothing | That the video is real |
Sources: YouTube Help, disclosing altered or synthetic content, and "Captured with a camera" disclosure.
YouTube adds the AI label itself in three situations: when the video was made with YouTube's own generative tools, when the file carries C2PA metadata saying AI was used, and when its internal systems detect AI generated or altered content. Everything else depends on the creator telling the truth.
The camera signal is rarer still. It only appears when a creator chose to record with a device or app that writes C2PA 2.1 metadata, and any edit that breaks the chain of provenance removes it. YouTube is explicit that when it is missing, "it doesn't mean the content has modified audio or visuals."
To see both, open the full description on the watch page (tap the title on a phone, or "more" under the video on a computer) and scroll to the section called How this content was made.
What Creators Must Label, and What They Need Not
The rule is about realism, not about whether AI was involved at all. A creator has to disclose content that viewers could mistake for something real.
| Situation | Label needed? | Why |
|---|---|---|
| A real person made to appear to say or do something they did not | Yes | Realistic, and could mislead viewers |
| Footage of a real event or place that has been altered | Yes | Realistic, and could mislead viewers |
| A realistic scene that never actually happened | Yes | Realistic, and could mislead viewers |
| Beauty filters, colour and lighting adjustments | No | Minor edits |
| Cloning your own voice for a voice over or dub | No | Your own voice, used for narration |
| AI help with a script, outline, thumbnail, title or infographic | No | Production assistance |
| Clearly unrealistic content, like riding a unicorn through a fantasy world | No | Nobody would take it as real |
Source: YouTube Help. Creators who consistently fail to disclose can have a label added for them, videos removed, or lose Partner Program access.
Read the right hand column carefully, because it explains most of the confusion. A video can use a cloned voice for the whole narration, an AI written script and an AI thumbnail, and still carry no label at all, legitimately. The label is there to warn you about fake events and fake people, not about AI assisted production.
Seven Checks You Can Do in Ten Minutes
When a video has no label and something feels off, these checks, done in order, answer most cases. None of them needs special software.
- Read the whole description. Look for the AI label and for "How this content was made". Look for a named original source too. Genuine footage is usually credited to the person or outlet that filmed it.
- Look at the channel, not just the video. When was it created? Does it post the same kind of dramatic clip every day? Does the person on screen appear anywhere else, in interviews or on their own verified channel, sounding and looking the same?
- Slow it down. Pause, then press the full stop key to move one frame forward and the comma key to move one frame back. The less than key slows playback. Watch hands, teeth, earrings, hair against the background, text on signs and people in the background. These are where generated video most often breaks.
- Listen without the picture. Close your eyes, or cut the audio out and play it alone. Listen for breaths, room sound that matches the space, and words that land on the lips. A voice that never pauses for breath in a long speech is worth a second look.
- Find where it first appeared. Take a screenshot of a clear frame and search it with a reverse image search such as Google Lens. If the same footage appears earlier, with a different caption, a different date or a different place, you have your answer.
- Check the event itself. If the video shows something newsworthy, real events leave a trail: other angles, other uploads, reports from outlets that were there. A dramatic event that exists in one video only deserves doubt.
- Treat detection tools as a hint. Browser plugins such as InVID WeVerify break a video into keyframes for reverse search, which is genuinely useful. Automatic "AI or not" scores are not proof. They disagree with each other and get real footage wrong.
What Each Check Catches, and What It Misses
| Check | Time it takes | Good at catching | Can miss |
|---|---|---|---|
| Labels and description | Under a minute | Disclosed AI, and camera verified footage | Everything undisclosed |
| The channel | 2 minutes | Channels built to post fakes in bulk | A single fake on a normal channel |
| Frame by frame | 3 minutes | Warped hands, melting text, flickering detail | High quality generations and short clips |
| Audio alone | 2 minutes | Cloned voices with no breath or room sound | Well produced voice clones |
| Reverse image search | 2 minutes | Real footage reused with a false caption | New footage with no earlier copies |
| The event itself | 2 minutes | Invented events | Fakes of small, private moments |
| Detection tools | 1 minute | Giving you a lead to follow | Anything. Treat the score as a hint, never a verdict |
Notice that the most common fake is not generated at all. It is real footage with a false caption: an old video described as today's, or a clip from one country presented as another. Reverse image search catches those in minutes, and AI detectors cannot, because the footage is genuine.
Save the Evidence Before It Changes
If a video matters, because it names you, misleads people about something important, or you plan to report it, save it before you do anything else. Uploads are edited, relabelled, made private and deleted, and the version you saw may not exist tomorrow.
| What to save | How | Why it matters |
|---|---|---|
| The full link and a timestamp link | Right click the video, "Copy video URL at current time" | Points anyone straight to the exact second |
| Channel name, upload date and view count | A screenshot of the watch page, with the date visible | Shows where it was and how far it spread |
| The description as it was | A screenshot of the expanded description | Labels and captions can be edited later |
| The exact seconds in question | Cut and download just that range as MP4 | Evidence that survives the upload being removed |
| Your own note | Date, time and what you noticed | Memory fades faster than you expect |
Cutting only the seconds that matter keeps the evidence small, easy to share with a platform, a journalist or a lawyer, and focused on the moment itself rather than a three hour upload. Our guide to YouTube timestamp links covers the link side of this in detail.
The EU Rule That Started on 2 August 2026
Platform labels are now backed by law in Europe. Article 50 of the EU AI Act applies from 2 August 2026, and it covers deepfakes directly: AI generated or manipulated content that resembles real people, places or events and would falsely appear authentic.
| What | Detail |
|---|---|
| Applies from | 2 August 2026 |
| Who must disclose | Anyone who publishes a deepfake must disclose that it was artificially generated or manipulated |
| Who must mark | Providers of generative AI systems must mark their output in a machine readable way |
| Grace period | Systems already on the market before 2 August 2026 have until 2 December 2026 for machine readable marking |
| Creative exception | Evidently artistic, satirical or fictional work needs only a disclosure that does not spoil the work |
| Maximum fine | EUR 15 million, or 3% of worldwide annual turnover if higher (Article 99) |
Sources: EU AI Act, Articles 50 and 99. This is a summary, not legal advice.
For viewers, the practical effect is that a professional deepfake published in the EU without disclosure is now a legal problem for whoever published it, not only a breach of platform rules. That makes it worth saving evidence and reporting, rather than just scrolling past.
When You Think a Video Is a Harmful Fake
Use Report in the three dot menu under the video and choose the reason that fits, such as misinformation or impersonation. Attach your saved timestamp in the report so the reviewer lands on the right moment.
If the fake shows you, YouTube also runs likeness detection. It launched in 2025 for roughly 4 million creators in the YouTube Partner Program, and in 2026 YouTube began piloting it with government officials, political candidates and journalists. It finds AI generated content that uses a person's face and lets them request its removal.
Frequently Asked Questions
Does YouTube label AI videos automatically?
In three cases: when the video was made with YouTube's own generative AI tools, when the file carries C2PA metadata showing AI was used, and when YouTube's systems detect AI generated or altered content. Otherwise the label depends on the creator disclosing it.
If a video has no AI label, is it real?
No. A missing label proves nothing. Creators must only label realistic content, many do not, and most videos carry neither an AI label nor the camera label.
What does "Captured with a camera" mean on YouTube?
It appears under "How this content was made" when the video was recorded with tools that support C2PA version 2.1 or higher and was not edited in sound or picture. YouTube says its absence does not mean a video was altered.
Can an AI detector tell me for sure?
No. Automatic detectors disagree with each other and misjudge real footage. Use them as a lead, then confirm with reverse image search, the source and the event itself.
Do I have to label AI in my own YouTube uploads?
Only for realistic content: a real person saying or doing something they did not, altered footage of a real event or place, or a realistic scene that never happened. Filters, your own cloned voice for a voice over, and AI help with scripts or thumbnails need no label.
What does the EU AI Act require for deepfakes?
Since 2 August 2026, anyone publishing a deepfake must disclose that it was artificially generated or manipulated, with a lighter rule for evidently artistic or satirical work. Fines can reach EUR 15 million or 3% of worldwide turnover.
How do I keep proof of a video that might be deleted?
Save the link with a timestamp, screenshots of the watch page and description with the date visible, and a clip of the exact seconds as an MP4. Add a note of when you saw it.
The Bottom Line
YouTube's labels help when they are there, but most videos carry none, and the absence of a label tells you nothing. The reliable method is still the old one: slow it down, find the source, check the event, and do not trust a single detector score.
When a video matters, save it before you judge it. A timestamp, a screenshot and a clip of the exact seconds will outlast the upload, and since August 2026 they matter to a law as well as to a platform.
This guide describes YouTube's published policies and the EU AI Act as of September 2026. It is not legal advice.