A clip with millions of views and no traceable author has become the default shape of a viral video. Two stories on this site show the pattern at full speed: Fruit Love Island, a series that pulled in hundreds of millions of views from generated footage, and Tiffany Sloane, a stand-up act with six figures of followers and nobody behind the microphone.

Squinting at the frames no longer settles the question.

What does settle some questions is evidence stored inside the file, plus a few commands that read it. This workbench runs five checks in order. Each produces something you can write down, and a results table near the end explains what every outcome does and does not prove. Plan on fifteen minutes the first time, five after that.

Why Your Eyes Are the Weakest Instrument

Spotting guides still circulate, and some of their advice holds up. TechRadar’s checklist for Sora 2 footage tells viewers to inspect backgrounds for impossible proportions, test whether shadows and reflections obey physics, watch for puppet-like movement, and look for smudged or over-smoothed texture. The same piece concedes the limit: the newest generators have fixed most of the old giveaways, and even people who work in the field get fooled.

Treat visual tells as a tiebreaker, never as a verdict.

For a snapshot of how thoroughly one generated format has saturated feeds, this commentary video on the AI Love Island trend is a useful reference point.

“Love Island AI Slop Has Taken Over TikTok” (YouTube, Rebal D). Third-party commentary; not produced by x9m8.

Saturation on that scale is exactly why a repeatable method beats intuition.

What You Need Before You Start

  • c2patool: the open-source command-line reader from the Content Authenticity Initiative’s c2pa-rs project. Installation steps live in the project’s README.
  • FFmpeg (ffmpeg and ffprobe): stream details and frame extraction.
  • ExifTool: container metadata fields that ffprobe may not list.
  • A browser, for the validators in Step 4.
  • A terminal, and the video file itself.

One rule governs everything below. Work from the copy closest to the source, because a screen recording or a re-uploaded download is a new file, for reasons the first step explains.

Step 1: Get the Closest Copy to the Source

Content Credentials travel inside the file as a signed manifest. Re-encoding can strip it, and SoftwareSeni’s 2026 adoption review reports that social platforms often remove manifests during upload and transcoding. A platform can support the standard and still lose the signature before you ever download the clip.

Ask the poster for the original. Check whether the platform displays its own label. Open the account’s history before touching a command line: TechRadar advises caution with random viral pages built on sensational clips that never cite a source, and suggests searching for coverage by established outlets when footage claims to be news.

Write down where your copy came from. Every later result depends on it.

Step 2: Read the Container With ffprobe and ExifTool

This step does not detect AI. It records what the file claims about itself, which gives later findings context.

ffprobe -v quiet -print_format json -show_format -show_streams clip.mp4
exiftool -a -G1 -s clip.mp4

Note the encoder tag, creation time, codec, resolution, frame rate, and duration. An encoder string beginning with Lavf means the file passed through FFmpeg’s libavformat at some point, which is true of countless ordinary edits and platform exports. Missing camera make, model, or location data is just as unremarkable, since upload pipelines remove it. Neither observation points to synthesis. Their value is as a record: two clips from one account that share identical encoder settings and near-identical durations form a pattern worth logging.

Printed video stills on a steel table with red adhesive dots and a gloved hand holding a loupe

One more measurement belongs here. Scene-change timestamps show how a clip was assembled:

ffmpeg -i clip.mp4 -vf "select='gt(scene,0.4)',showinfo" -f null - 2>&1 | grep pts_time

Cuts that land at suspiciously even intervals are a pattern, not proof. Human editors working to a rhythm produce the same thing.

Step 3: Look for Content Credentials With c2patool

By default, c2patool prints an asset’s manifest as JSON, according to the project’s usage documentation. Four variants cover most needs:

c2patool clip.mp4            # print the manifest JSON
c2patool clip.mp4 --info     # manifest count, store size, validation status
c2patool clip.mp4 --tree     # assertions and ingredients as a tree
c2patool clip.mp4 -d         # detailed report of the internal format

The –info report is the quick triage. It states whether a manifest store exists, how many manifests it holds, and whether the manifest validated. If nothing turns up, jump to the results table. If a manifest does exist, read three things: which tool generated the claim, which actions are recorded (created, edited, transcoded), and whether any assertion declares a digital source type. In the IPTC vocabulary that C2PA draws on, the value trainedAlgorithmicMedia marks output from a generative model. Confirm exact field names against the current C2PA specification, since manifests vary between signers.

Validation is not trust. A manifest can be internally consistent and still come from an unknown signer. The trust subcommand checks the signing certificate against a list you supply:

c2patool clip.mp4 trust --trust_anchors='<URL of the official C2PA trust list>'

The project documentation says the official list needs only that one option, with no allowed list or trust config. Its worked examples use the older interim list, which SoftwareSeni says was replaced in mid-2025 and frozen on January 1, 2026; content signed under it stays valid. Take the current list URL from the C2PA’s own site rather than from a blog post, this one included.

Step 4: Cross-Check in a Browser Validator

A second implementation catches tool-specific quirks. Adobe’s Inspect tool is documented as a C2PA conforming validator for image, video, audio, and PDF files. Its documentation says you can upload a file or point the tool at a public address using the pattern inspect.cr?source=ASSET_URL. The asset cannot require a login, and the host must permit cross-origin requests. When no manifest exists, Inspect reports that no Content Credential is present; when the signing certificate is absent from the C2PA trust list, it marks the result as unrecognized.

The Content Authenticity Initiative also hosts a free verifier at verify.contentauthenticity.org. Per SoftwareSeni, it checks the manifest and certificate chain against the Trust List, and because certificates travel inside the manifest, no call to the original signer is needed.

Step 5: Run the Behavior Checks on Extracted Frames

Only now do visual tells earn a turn. Pull two frames per second and study them at full size instead of at scroll speed:

mkdir frames && ffmpeg -i clip.mp4 -vf fps=2 frames/f_%04d.png
  • Backgrounds: lines that fail to meet, shifting proportions, bystanders doing odd things.
  • Physics: shadows, reflections, and lighting that disagree within a single scene.
  • Motion: hair or fabric moving without wind, objects that wobble, blinking or smiling that feels mechanical.
  • Texture: grainy patches, melted detail, over-airbrushed skin.

Next, run the strongest frames through a reverse image search such as Google Lens or TinEye. A match to older footage means the clip was recycled. No match tells you little, because generated footage is new by definition.

Macro of a half-cracked wax seal on a paper slip beside an SSD and USB cable

Last, notice your own reaction. TechRadar notes that content engineered to provoke strong feelings makes people less likely to question it. A clip that makes you furious is the one to run through this workbench.

How to Read Your Results

ResultWhat it supportsWhat it does not prove
Valid manifest, trusted signer, declares generative AIStrong evidence the file is AI output, as declared by whoever signed itAnything about other clips from the same account
Valid manifest, signed at capture by a deviceThe file left a signing device and records its later editsThat the scene itself was not staged
Manifest present, signer unrecognizedTreat the claims as unverifiedThat the video is fake
No manifestNothing either wayThat the video is real or fake; per the Reuters Institute figure SoftwareSeni cites, under 1% of news images and videos carry C2PA data
Only visual tellsA reason to keep investigatingAnything on its own; the tells are unreliable now

Microsoft’s February 2026 report, as quoted by SoftwareSeni, is blunt about the ceiling on all of this:

Preventing every attack or stopping certain platforms from stripping provenance signals isn’t possible.

Jessica Young, Microsoft director of science and technology policy, via SoftwareSeni

The standard also records claims rather than detecting synthesis. SoftwareSeni stresses that C2PA does not detect AI-generated content; it reports what was declared at signing.

Where Platforms and Hardware Stand Right Now

Hardware support is real but uneven. SoftwareSeni lists the Google Pixel 10 as signing every photo by default with hardware-backed keys, while the Samsung Galaxy S25 adds credentials only to photos created or edited with generative AI. The Nikon Z6 III received C2PA through a firmware update in August 2025, then lost it when a signing vulnerability forced certificate revocations; the review says service had not returned by early 2026. For live video, C2PA 2.3 (December 2025) added signing at the segment level.

Platform policy is moving as well. C2PA Viewer, citing a YouTube post dated May 27, 2026, says YouTube now applies AI labels automatically when it detects significant photorealistic AI use that the creator did not disclose, and treats a signed fully generative manifest as a permanent label. I could not confirm the original YouTube post, and C2PA Viewer sells verification tools, so read that framing with care. The same page says the EU AI Act’s Article 50 transparency rules become enforceable on August 2, 2026, with details still being finalized as of May 31. Check primary sources before relying on either point.

The Five-Minute Version

  1. Get the original file, or note exactly where your copy came from.
  2. Run c2patool clip.mp4 --info and record the result.
  3. Open the file in a browser validator and compare the verdict.
  4. Read the encoder and scene-cut data from ffprobe and ffmpeg for context.
  5. Extract frames, run a reverse image search, and decide what the evidence supports, not what you suspect.

Common Questions

Does a missing watermark mean a video is real? No. TechRadar warns that watermarks can be cropped, blurred, or faked, and that removal often leaves clues such as odd cropping or black bars.

Gloved hands removing a foil label from a parcel on a conveyor belt

Can I verify a screen recording? Not for provenance. A recording is a new file with no manifest, so ask for the original.

Does Content Credentials detection prove a video is AI? Only when a trusted signer declared it. Absence of a signature proves nothing.

Sources