How to Spot a Deepfake Video: What AI Detection Tools Can (and Can't) Tell You
Detection tools can flag signs that a video was manipulated. They can't hand you a verdict. Here's how the technology actually works, why a confidence score is more honest than a "real or fake" stamp, and what new EU rules mean for the clip in your feed.
A video crosses your feed. The audio sounds slightly off, or a face moves in a way that doesn't quite track. You run it through a detection tool and get a number back: "82% likely AI-generated." What do you do with that?
Most people treat it as a verdict. It isn't one. Detection tools measure statistical likelihood based on artifacts and patterns, not certainty. Understanding the difference between a signal and a verdict is the difference between using these tools well and being misled by them twice, once by the fake, and once by an overconfident scanner.
How Detection Tools Actually Work
Deepfake detectors don't "see" a fake the way a person might notice something is wrong. They look for statistical fingerprints left behind by the generation process. A few of the most common signals:
Physiological inconsistencies
Early detectors focused on things generators struggled to reproduce: natural blink rate, pulse-driven skin color changes, consistent eye reflections. Generation models have closed most of this gap, but inconsistencies still show up in edge cases like unusual lighting or side profiles.
Compression and frequency artifacts
Generated frames often carry subtle noise patterns or frequency-domain signatures that differ from a real camera sensor. This is invisible to the eye but detectable with the right models, at least until the footage is re-compressed or re-uploaded, which tends to wash the signal out.
Audio-visual sync
Lip movement that doesn't precisely match phoneme timing, or a voice that doesn't match the speaker's known cadence, is one of the more durable signals, particularly for face-swap and lip-sync style manipulations.
Provenance metadata
Some platforms and cameras now embed cryptographic content credentials at capture, following provenance standards like C2PA. Where present, this is stronger evidence than any after-the-fact visual analysis, because it traces a file's history rather than guessing from its pixels.
What These Tools Can Tell You
A well-built detector can tell you that a piece of content shows a statistically elevated likelihood of manipulation, based on specific, nameable signals. That's genuinely useful. It's a reason to slow down, look for corroborating evidence, and check the source before you share something.
The honest output of a detection tool is a confidence range, not a label. "62% probability of synthetic origin, driven primarily by frequency-domain artifacts" is a data point. "FAKE" in red text is a conclusion dressed up as a measurement, and it's usually the tool overstating what it actually knows.
What These Tools Can't Tell You
This is the part most coverage skips. A few limitations worth knowing before you treat any score as gospel:
There's no universal champion
No single detection tool holds a consistently verified accuracy edge across all manipulation types in 2026. A detector tuned for face-swaps can miss a voice clone. One trained on a specific generator's outputs can miss content from a newer model it's never seen.
Generation and detection are in an arms race
Every time a detection signal becomes well known, the next generation of generators is trained to remove it. Detectors published a year ago are measurably weaker against content made this year. Any tool's accuracy claims have a shelf life.
Real-world conditions degrade signal
Compressed video, low frame rates, re-uploads across platforms, and poor lighting all reduce the forensic signal a detector needs. Content that's been screenshotted, re-recorded, or run through a social platform's compression is harder to analyze than the original file, and most viral clips have been through exactly that process.
A clean result isn't proof of authenticity
A low manipulation score means the tool didn't find its known signals. It doesn't mean the content is genuine. Absence of detected evidence is not evidence of absence, especially against a generation method the detector wasn't built to catch.
Why confidence labels beat verdicts
A percentage with a named basis, for example "elevated frequency-domain artifacts, moderate confidence," gives you something to reason with. A binary "real" or "fake" label collapses that nuance into a single word and asks you to trust it blindly.
This is the same principle behind how Veritas Source handles AI-content signals generally: surface what the evidence shows, at the confidence level it actually supports, and let you weigh it alongside the rest of what you know about the source. Decide for yourself, with the full picture in front of you.
The EU AI Act Changes the Baseline
Starting in August 2026, the EU AI Act's transparency rules under Article 50 put new obligations on both sides of the deepfake equation.
| Who | Obligation |
|---|---|
| Providers of generative AI systems | Must mark outputs (image, audio, video) in a machine-readable format so they're detectable as AI-generated or manipulated |
| Deployers who create or share a deepfake | Must disclose that the content was artificially generated or manipulated, through a visible label, an opening disclaimer, or an audible notice |
There are carve-outs. Law enforcement use for detecting or investigating crime is exempt. Content that's evidently satirical, artistic, or fictional only needs to disclose that it's generated in a way that doesn't undercut the work itself, so a deepfake-based sketch or film doesn't need a disruptive on-screen watermark.
What this means practically: within the EU, a growing share of AI-generated video should carry some form of disclosure by design, rather than requiring after-the-fact detection at all. That shifts the burden partially upstream, from "can a viewer's tool catch this" to "was the platform or creator required to label it." Neither removes the need for detection tools, since plenty of synthetic content will originate outside EU jurisdiction or from bad actors who ignore the rule regardless of penalty.
How to Actually Use a Detection Score
- Read the basis, not just the number. A tool that names the specific signal it found (audio-sync mismatch, frequency artifacts, missing provenance metadata) is giving you something you can verify independently. A bare percentage with no explanation is not.
- Treat a low score as inconclusive, not clean. Especially for content that's been compressed, re-uploaded, or is otherwise low quality.
- Check for content credentials first. If a file carries C2PA or similar provenance metadata, that traces its actual history and is stronger evidence than any visual analysis run after the fact.
- Weigh the source alongside the content. A suspicious video from an account with a track record of fabricated content and no verifiable funding or ownership behind it is a different situation than the same video from an outlet with a transparent editorial history. The content signal and the source signal both matter, and neither is sufficient alone.
- Don't outsource the judgment call. A detection score is an input to your decision, not a replacement for it.
Deepfake detection is genuinely useful technology, and it's improving. But it works best as one layer of evidence among several, read with its confidence level and limitations in view, not as a machine that tells you what to believe. The tools give you a number. What you do with that number is still, and will remain, on you.
Check the Source, Not Just the Clip
Veritas Source is a free tool that surfaces a news source's ownership, funding, credibility history, and AI-content signals, no login required. See the evidence and draw your own conclusions.
Check a Source →FREE · NO LOGIN · NO PAYWALL