# How You Actually Verify Footage in 2026

Last month a friend sent me a clip from a protest. It looked real. The lighting matched the weather report. The audio had the right kind of crowd noise. Someone in our group chat asked the question everyone asks now: "How do we know this wasn't generated?"

That question has a better answer in 2026 than it did two years ago. It is still not a single yes or no. If you work in news, legal, or just want to share something responsibly, you need a small stack of checks, not one magic detector.

Here is what actually works today, what breaks on social platforms, and why we built POV the way we did.

## Start with provenance, not detection

The most durable shift in 2026 is provenance: cryptographically signed metadata that travels with a file and says who created it, when, and whether it has been edited since.

The open standard behind this is [C2PA Content Credentials](https://spec.c2pa.org/specifications/specifications/2.4/explainer/Explainer.html), maintained by the Coalition for Content Provenance and Authenticity. A C2PA manifest is a signed record attached to an image or video. A validator can confirm the signature, read the capture timestamp, see edit history, and check whether the bytes match what was signed.

Hardware is finally showing up. Leica shipped C2PA-native cameras starting with the M11-P in 2023. Sony's PXW-Z300 became one of the first professional video bodies with native signing for broadcast workflows. And in 2025, [Google made Content Credentials default on Pixel 10 photos](https://blog.google/security/pixel-android-trusted-images-c2pa-content-credentials/), using hardware-backed keys through the Titan M2 chip and reaching C2PA Assurance Level 2 on Android.

That matters because signing at capture is harder to fake than adding a badge after upload. It is also rare. Most footage you see on TikTok, X, or Telegram still has no credentials at all.

C2PA tells you about the file's history. It does not tell you whether the scene itself is truthful. A perfectly signed clip can still mislead if it is decontextualized. Provenance answers "was this file tampered with since capture?" not "did the camera operator tell the whole story?"

## Where the chain breaks (usually before you see the clip)

Even signed files lose their credentials fast. Social platforms re-encode video for delivery. Many strip embedded metadata entirely. A clip that was verifiable at the source can arrive in your feed as an anonymous MP4 with no manifest left to inspect.

C2PA 2.3 and 2.4 added live-streaming models (segment-level signing for CMAF/DASH), which helps for broadcast workflows. Consumer repost chains are still the weak link. If your verification workflow starts at the viral repost, you are often already too late.

That is why serious desks still do contextual verification: landmarks, shadows, weather, traffic patterns, language in background audio, and cross-checks against other sources. The craft has not gone away. The tooling has just added one more layer when you can get the original file.

## AI detectors: useful, not sufficient

When provenance is missing, teams reach for AI-generated video detectors. In 2026 those tools are better than the 2022 era of obvious face-swap glitches, but the honest research picture is sobering.

Benchmark work summarized by outlets like [PressVerified](https://pressverified.com/blog/video-deepfake-detectors-2026-compression-problem) shows lab accuracy collapsing once video goes through social-style recompression. Detectors trained on clean files often lean on compression artifacts or watermark cues rather than universal synthetic signatures. [Recent challenge results](https://couldthisbetrue.com/blog/deepfake-video-detection-2026/) note strong performance on unmodified synthetic clips, then sharp drops after x264/x265 passes, camera re-capture, or watermark removal.

False positives are real too: heavy filters, CGI, and unusual lenses can flag as synthetic. False negatives are worse: short clips, screen-recorded fakes, and brand-new generators outpace retraining cycles.

My rule: treat a detector score as one input. Never auto-publish or auto-ban on a threshold alone. Pair it with source tracing and, when possible, provenance validation on the earliest copy you can find.

## Context still wins for eyewitness video

For breaking news, the verification stack I trust looks like this:

1. **Trace the original uploader**, not the tenth repost.
2. **Check place and time** against independent signals (maps, satellite, local reports).
3. **Inspect edit history** if credentials exist; if not, look for splices, loop points, or audio mismatch.
4. **Run detection only on the best available source file**, and record the uncertainty.
5. **Document what you verified** so the next editor inherits the chain.

Newsrooms have done versions of this for years. What changed is the volume of synthetic video and the speed of reposting. The labor is the same. The failure modes are faster.

## Why we built POV around capture context

I started POV because I kept seeing the same gap: by the time footage goes viral, the original capture context is gone. Who filmed it? Were they actually there? Has the file been swapped?

POV is not a C2PA camera. We do not sign pixels inside your phone's secure enclave today. I want to be precise about that because the board and our users deserve honesty, not marketing fog.

What POV does today is bind footage to **capture context inside our app**:

- **Geo-fenced, time-bound bounties.** A requester pins a location and window. Creators record in-app, and we store a GPS track during capture tied to that bounty.
- **Continuous in-app recording.** Submissions are expected to be single-take clips captured through POV, not uploaded files from the camera roll (for standard bounty flows).
- **Metadata preserved with the submission.** Location history, timestamps, and orientation travel with the clip in our database.
- **Integrity hashing on licenses.** When footage is licensed, we attach a content hash and verification record buyers can inspect in the agreement.

Delivery runs through Bunny.net Stream; the file you license is the file we stored from that submission pipeline.

That is a **platform chain of custody**, not hardware provenance. It raises the cost of swapping in a generated clip after the fact because the video, the uploader account, the bounty context, and the location trace are linked at submission time.

It is also not perfect. GPS spoofing on mobile devices is a known industry problem, and we treat stronger device attestation as direction we want to harden, not a claim we make today. We are watching C2PA adoption closely and exploring how POV's capture pipeline can align with open provenance standards over time.

The mission is straightforward: make it easier to reward people who were actually on scene, and give buyers metadata they can audit. We are building toward hardware-backed provenance. We will not pretend we are already there.

## A practical checklist you can use tonight

If someone sends you a clip and asks "is this real?", run this:

- **Ask for the earliest file**, not a screen recording of a screen recording.
- **Look for Content Credentials** if the source is a Pixel 10 photo, a C2PA-enabled camera, or a platform that preserves manifests.
- **Verify context** (place, time, weather, shadows, language) even when credentials check out.
- **Use AI detectors as a hint**, not a verdict, especially on compressed social video.
- **Write down what you checked** so the next person does not restart from zero.

If you are a creator, the best thing you can do is preserve your original capture and capture notes (location, time, what you saw). Provenance tools are spreading, but contextual detail still saves you when metadata gets stripped.

## Where POV fits (soft CTA)

If you are a journalist, researcher, or local org that needs footage from a specific place and moment, POV lets you post a bounty, fund it, and receive in-app captures with submission metadata attached. You still apply your own editorial judgment. We give you a cleaner starting point than a random DM with a watermark cropped out.

If you are on scene and willing to record, the app is free to use. Show up, press record, submit. Your clip keeps its context.

That is the bet we are making: verified context at capture beats detective work on a repost chain every time.