Does SynthID Detect AI Videos Made With Runway and Pika in 2026?

The question “does SynthID detect AI videos made with Runway and Pika in 2026” comes up constantly from creators, journalists, and moderators trying to verify suspicious clips. Short answer: no — Runway and Pika do not use SynthID watermarks, so their videos come back clean on every SynthID check. That single fact makes per-tool coverage the most misunderstood part of AI video detection, because almost no guide explains which generators are covered and which are not. Below is the coverage matrix the internet is missing, an honest guide to reading each verdict, and a test protocol you can run yourself.

What SynthID Actually Is

SynthID is Google DeepMind’s invisible watermarking system for AI-generated content. It embeds an undetectable signal directly into the pixels, audio, or text a model produces — one designed to survive common edits like cropping and compression. Anyone can later scan a file and determine whether a participating AI model generated it.

Two facts shape everything that follows: SynthID Detector opened to the public on October 7, 2026 (verification was previously limited to partners), and Google reports more than 180 billion files watermarked so far (Google’s figure — re-verify on the official SynthID page, since adoption keeps changing). Crucially, SynthID is opt-in: it only fires when a tool’s maker chose to adopt it. That is exactly why Runway and Pika videos slip through.

The Coverage Matrix: Who Watermarks, Who Doesn’t

This is the table most detection guides never publish. Coverage can shift as partnerships change, so re-verify before relying on it — but as of October 2026:

Two implications follow. A positive SynthID match is strong evidence — it identifies content from a watermarked model, often which model. A negative result proves nothing: the clip could be real footage, or a Runway/Pika video that was never watermarked. Every guide that skips this distinction misleads its readers.

What “No Watermark Found” Really Means

Treat a negative result the way a doctor treats a negative test for one specific illness — it rules out one thing, not everything. “No watermark found” means exactly one of these:

1. The file was never watermarked (human-made, or made with a non-participating generator like Runway, Pika, or Midjourney).
2. The file was watermarked but the signal was destroyed — aggressive re-encoding, heavy filters, or face-swapping on top can degrade invisible marks.
3. The file carries a watermark from a system this detector doesn’t check for.

Never report a clean SynthID result as “this video is real.” It only means “no surviving SynthID watermark.” Conversely, if you publish with Runway or Pika and want audiences to know the content is AI-generated, don’t rely on SynthID — add a visible disclosure.

Reading a SynthID Detector Verdict

When you upload a video to the public SynthID Detector, you generally get one of three outcomes:

• Watermark detected. The content came from a participating model. Cross-check the stated source against the claimed provenance — if someone posted it as “my original footage,” you have your answer.
• No watermark detected. Inconclusive. This is where most mistakes happen — especially with Runway and Pika content.
• Partial or uncertain. The signal is weak, possibly from heavy editing or multi-generational re-posting. A prompt to use other signals, not a verdict.

A Hands-On Test Protocol

Don’t take the matrix on faith — run this controlled test in under an hour:

1. Generate your own test clips. Make a 5-second clip in Runway, one in Pika, and — if you have access — one in Google Veo. Known provenance, no guessing.
2. Run each through the SynthID Detector (current link on Google DeepMind’s SynthID page) and record every verdict.
3. Test robustness. Screen-record the Veo clip that flags positive, re-upload the recording, and re-test. Note whether the watermark survives — this shows why detection degrades on reposted content.
4. Test evasion. Compress one positive clip at very low bitrate and re-test. Record the result.
5. Compare. Your table should show positives only on watermarked-model outputs and negatives on Runway/Pika — confirming that per-tool coverage, not “AI video detection” in general, decides the outcome.

What Actually Catches Runway and Pika Videos

Since SynthID can’t help with these two generators, your verification stack needs other layers: C2PA content credentials (check with a validator, though metadata is easily stripped); platform labels (YouTube, Instagram, and TikTok flag what their classifiers catch — coverage is partial and labels vanish on re-uploads); multi-signal detector tools that analyze visual artifacts (treat every score as a probability, not proof); and manual visual inspection — see the full checklist on https://aivideodetector.site/ and run suspect clips through it before you share. No single layer is reliable alone. Combine at least two.

Frequently Asked Questions

Does SynthID detect AI videos made with Runway in 2026?

No. Runway does not embed SynthID watermarks, so Runway-generated videos return “no watermark found.” A clean result says nothing about whether the video is real.

Does SynthID detect AI videos made with Pika in 2026?

No. Pika is not a SynthID partner either — its clips pass through watermark checks undetected, just like Runway’s.

Can I trust a “no watermark found” result?

Only as a negative test for participating models. It rules out a surviving SynthID watermark — it never proves a video is human-made, because non-participating generators never carried the mark at all.

Which AI video tools are covered by SynthID?

Google’s own models (Veo for video, Gemini for image/audio) watermark by default, and Google lists OpenAI, NVIDIA, and Kakao as announced partners. Verify the current partner list on the official SynthID page, since adoption changes.

Does the public SynthID Detector work on re-uploaded videos?

Sometimes. SynthID is designed to survive compression and cropping, but heavy editing, screen-recording, and face-swapping can degrade the signal. Test with your own re-uploads to calibrate expectations.

The coverage gap is the whole story: SynthID catches what its partners make and is blind to everything else. Whenever provenance matters, run a watermark check and at least one other layer — start with https://aivideodetector.site/ before you trust, share, or publish a suspect video.