7 Free Ways to Check If That "Too-Real" Video Was Actually Made by AI

Last month I spent 40 minutes arguing in a group chat that a video of a goat riding a skateboard was "obviously real" because the lighting looked natural. It was AI. The goat was not real. My dignity was not real.

If you work in content, marketing, or social — congrats, you will eat this exact humble pie at least once a quarter. The good news: you don't need a forensic lab or a six-figure verification suite. You need about seven free things, a little paranoia, and the willingness to look stupid for ten seconds before you share.

Before you lose sleep verifying everyone else's clips, make sure yours are unmistakably yours — we'll come back to that at the end. 👇

Table of contents

• What "AI detection" even means (and why most of it is a lie)
• The 7 free checks
• Turn this list into a 60-second habit
• 4 tips that save you from looking dumb
• How to choose: a 3-step triage
• FAQ — just remember TRACE

What "AI detection" even means (and why most of it is a lie)

There are three completely different things people call "AI detection," and mixing them up is how smart teams get burned:

• Camp A — Provenance / watermark verification (SynthID, C2PA): reads the birth certificate the generator left behind in the file. Extremely reliable for the models that use it.
• Camp B — Pixel-guess classifiers (Hive, Winston AI, TruthScan): stares at pixels and guesses "vibe says fake." Sometimes right, often vengeful.
• Camp C — Source / context checks (reverse image search, your own brain): asks where did this come from and does the story hold up.

Here is the sentence that should be tattooed on every content calendar:

A clean result is NOT proof of real.

"SynthID detected" means an AI system signed this. "SynthID not detected" means no partner signal was found — which could just as easily be Meta, an open-weight model running on someone's laptop, or a watermark that got stripped. Google itself says the tool "is not a general-purpose AI detector." Treat it as one layer, not the whole answer.

The 7 free checks

1. SynthID Detector — the bouncer that only checks IDs from partner nightclubs

• Link: synthid.com
• What it's like: a bouncer who only recognizes IDs from a short list of partner clubs — Google, OpenAI, NVIDIA, Kakao (Apple "soon"). Show up from anywhere else and he shrugs.
• Best for: checking whether an image, video, or audio file carries a SynthID watermark from a supported AI system.
• Who should use it: literally anyone with a Google / OpenAI / Apple login. Free, but capped at roughly 10 checks per day.

2. C2PA Content Credentials Verify — the opposite end of the same stick

• Link: contentcredentials.org/verify
• What it's like: instead of tagging the fakes, it signs the real thing at the moment of capture — a passport stamped at the camera, not a lie-detector for everything else.
• Best for: checking whether a photo or video carries a signed provenance trail (who made it, with what, when).
• Who should use it: photographers, Pixel owners, and anyone receiving "original" media they need to trust.

3. Google Lens / Reverse Image Search — grab it by the scruff

• Link: lens.google.com
• What it's like: grabs the image by the scruff of its neck and drags it back to the first place it ever appeared online.
• Best for: finding the original source, earlier posts, and the context that blows the "breaking" story wide open.
• Who should use it: everyone. Zero skill required, works on your phone in five seconds.

4. InVID & WeVerify — a magnifying glass taped to a metadata stomach-pump

• Link: invid-project.eu
• What it's like: a magnifying glass duct-taped to a metadata stomach-pump — it pulls video frames apart and reads what the file is trying to hide.
• Best for: keyframe extraction, reverse-searching individual video frames, and spotting edits.
• Who should use it: journalists, fact-checkers, and the chronically suspicious.

5. EXIF / metadata viewers — the fingerprint reader

• Link: exif.regex.info (or your OS file inspector)
• What it's like: the digital fingerprint reader — reads what the camera left behind before someone wiped it clean.
• Best for: checking capture metadata and spotting files where the metadata is missing or clearly scrubbed.
• Who should use it: investigators and anyone who's ever received a "leaked" image with suspiciously empty properties.

6. Pixel-guess detectors (Hive / Winston AI / TruthScan) — the vibe-based gut-check

• Link: hivemoderation.com, winston.ai
• What it's like: a vibe-based gut-check — useful in a pinch, wrong often enough that you should never bet the farm on it.
• Best for: a second opinion when no watermark exists and you just need a directional read.
• Who should use it: community moderators filtering user-generated content at scale.

7. The human sniff test — the oldest detector in the world

• Link: (not a tool — a habit)
• What it's like: your own "wait, that's weird" reflex, finally weaponized.
• Best for: catching the stuff no tool catches — a story that doesn't add up, a source with a motive, a timeline that's off by a day.
• Who should use it: you, in the ten seconds before you hit share.

Turn this list into a 60-second habit

Don't memorize the tools — memorize the order. A repeatable verification SOP looks like this:

1. Sniff first. Does the claim, source, or timing feel off? If yes, slow down.
2. Watermark check. Run the file through SynthID (and C2PA verify if it claims to be "original").
3. Trace it back. Reverse image / frame search to find where it actually came from.
4. Cross-check. If one method says "clean," don't stop — a clean result is not proof of real.

Do that four times and it becomes muscle memory. The goal isn't to be a forensic expert; it's to not be the person who shared the fake goat.

4 tips that save you from looking dumb

Use the real checker, not the chatbot. Asking Gemini "is this AI?" sometimes works and sometimes the model just makes up an answer. Forcing it with the @Verify AI command makes it actually call the verifier. If you want the truth, use the dedicated tool, not the chatty middleman.

Upload the highest-quality file you have. Google explicitly recommends it: compression and heavy edits weaken the invisible signal. A screenshot of a screenshot is the fastest way to get a useless "not detected."

Run it twice / in a fresh session. Practitioners have reported that a stale chat session can reuse the first image's result for later checks. New file, new session, clean answer.

Stack signals, don't bet on one. Watermark + reverse search + a ten-second context check beats any single "100% FAKE" badge. The tools that shout confidence are usually the ones you should trust least.

How to choose: a 3-step triage

• Got a suspicious image or video? → Start with SynthID Detector + reverse image search. Fast, free, and catches the watermarked majority.
• Need to prove something is a real original? → Use C2PA Verify. A signature beats a guess.
• Found nothing? → That is not proof of human. Document your process and move on — don't declare victory.

FAQ — just remember TRACE

Q: Does "no watermark" mean the content is real? No. It means no SynthID-partner signal was found. The generator could be Meta, an open-weight model, or the watermark could have been removed.

Q: Can SynthID be fooled or stripped? Heavy edits can weaken it, and community-built removal tools exist. Treat a positive as a strong signal, not an absolute law.

Q: Why did my own camera photo get flagged? False positives happen — practitioners have documented real, unedited photos getting flagged. The dedicated detector is more reliable than asking a chatbot to check.

Q: Do I have to log in, and is there a limit? Yes — a Google, OpenAI, or Apple account, with a daily quota of roughly ten checks across image, video, and audio.

Q: Does it check text too? The public tool covers image, video, and audio. Text isn't in the current public release.

The part creators actually need to hear

Verification tools only exist because the flood is real. If you're making AI video — and especially if you want people to trust it — generate on a stack that leaves provenance behind by default. textideo runs on generators (Google's Veo among them) that embed SynthID at creation, so when a viewer runs your clip through the detector, the answer is "yes, and here's the receipt" instead of "who knows."

Provenance isn't a tax on creativity. It's the thing that keeps "made with AI" from meaning "don't trust it."

---

Sources

• Google rolls out improved SynthID AI content detector, now available globally — launch details, partner list, daily quota, and Google's own "not a general-purpose detector" disclaimer. (https://arstechnica.com/ai/2026/10/google-rolls-out-improved-synthid-ai-content-detector-now-available-globally)

• Google's AI detection website is now available — supported formats, login requirement, and the can't-distinguish-created-vs-edited limitation. (https://www.engadget.com/2279565/google-synth-id-detector-ai-detection-website-is-now-available/)

• Google just launched a free AI detector for music, videos, and images — plain-English explainer of what SynthID is and why "not detected" ≠ authentic. (https://www.tomsguide.com/ai/google-just-launched-a-free-ai-detector-for-music-videos-and-images-heres-how-synthid-works-and-what-it-can-and-cant-tell-you)

• I tested Google's new SynthID detector on AI slop. There's a big catch — hands-on test showing a "no watermark" result does not prove a video is real. (https://www.pcworld.com/article/3255083/i-tested-googles-new-synthid-detector-on-ai-slop-theres-a-big-catch.html)

• SynthID Detector Goes Public: Google's Tool to Spot AI-Generated Content — scale stats (180B+ watermarked assets) and the "not a general-purpose AI detector" framing. (https://www.it-connect.tech/synthid-detector-goes-public-googles-tool-to-spot-ai-generated-content)

• Reddit — "Google SynthID is not perfect, flagged many of my photos" — community discussion confirming false positives on real, unedited photos. (https://reddit.com/r/isthisAI/comments/1uu3y77/google_synthid_is_not_perfect_flagged_many_of_my/)

• Reddit — "Proof that AI writing detectors are BS" — practitioner skepticism about the reliability of AI detectors broadly, useful context for the "stack, don't bet" point. (https://reddit.com/r/antiai/comments/1v4woi4/proof_that_ai_writing_detectors_are_bs/)