Can you tell whether a video was made by a person or by a machine? A few years ago the answer was usually obvious: warped faces, melting backgrounds, and physics that ignored gravity. Today the question is genuinely hard. Modern generative models produce clips with cinematic lighting, coherent motion, and details that hold up under scrutiny. The result is a practical problem for journalists, marketers, researchers, and anyone who consumes media online.
This guide walks through the detection techniques that still work, explains why many older methods are failing, and looks at the provenance systems that are becoming the real long-term answer. You will leave with a concrete checklist you can apply the next time you suspect a video is synthetic.
One caveat before we start: this field moves fast, and any specific list of tells will be partially outdated within a year. What ages well is the method: where to look, how to weigh evidence, and how to think about the problem. That is what this guide is really about.
Why Detection Became So Hard
Detection is hard because generation improved faster than detection. Early text-to-video systems were essentially animated image generators: each frame was produced independently, so characters flickered between frames, textures swam, and objects appeared and disappeared between cuts. Detectors could rely on pixel-level anomalies and repeating latent patterns, and they worked.
The current generation of models is different. They are trained with temporal consistency as a core objective, which means the model reasons about how a scene changes over time, not just how a single frame looks. Physical motion, occlusion, and lighting transitions are modeled rather than guessed. When Sora-style architectures and their competitors arrived, they brought long, coherent sequences that respect perspective and gravity most of the time. The old detector playbook stopped working almost overnight.
There is a second reason detection is hard: the gap keeps moving. Every time researchers publish a reliable detector, model builders see it and train the next version to be less detectable. This is an arms race, and it shows no sign of slowing.
The Visual Telltales That Still Work
Even the best models leave traces. You just need to know where to look. The most reliable tells are usually in the details that models are still not trained enough on: hands, teeth, ears, small text, and fine repetitive structures.
Before you start hunting for specific tells, set your expectations. No single feature proves a video is synthetic, and a clean frame does not prove it is real. What you are doing is accumulating evidence: each tell raises the probability, and the final judgment comes from the weight of the evidence rather than any one detail. The goal is calibrated suspicion, not paranoia.
Hands and Teeth
Hands remain the classic failure point, although modern models handle them far better than older ones. Look for extra fingers, fingers that merge, joints at impossible angles, or hands that change shape between shots. Teeth are similar: count them. Synthetic mouths often render teeth as a single white blur, or add a stray tooth where the gum line should be. These errors are less common than they were, but they still appear when hands or mouths are partially occluded or moving quickly.
Light and Reflections
Generative models learn lighting statistics, but they do not always respect physical optics. Check reflections: a reflection in an eye, a window, or a puddle should match the scene's light sources and move plausibly when the camera moves. Look for reflections that are slightly wrong, duplicated, or missing entirely. Also check shadows: a soft shadow that jumps, a shadow pointing away from the light source, or two objects casting inconsistent shadows is a strong signal.
Background Objects and Physics
Text is a gift to detectors. Models often render signage, book spines, or product labels as near-gibberish. If you can pause on a sign or a screen, zoom in, and the letters are wrong, that is a very strong tell. Physics is also useful: watch how liquids splash, how cloth falls, how dust settles. Minor physics errors, like a flag moving in two directions at once or a cup that tips without spilling, are common in synthetic clips.
Audio and Lip Sync as a Warning Zone
Many people focus on visuals, but audio often reveals synthesis faster. Synthetic speech has improved dramatically, yet it still shows signs: a slightly too-regular rhythm, unnatural breathing placement, or phonemes that are crisp but lack the messiness of real speech. The most useful check is synchronization. Real footage has natural audio-visual lag, room tone, and ambient noise that varies with camera position. AI clips often have perfectly synced audio that is too clean, or visibly unsynced lips on the other end.
A practical trick: mute the video and watch the speaker's mouth, then listen without watching. If either channel feels mechanical on its own, treat the clip with suspicion.
Behavioral and Semantic Signals
Beyond pixels and audio, there are semantic tells. Synthetic scenes often have an uncanny narrative logic: people walk into frame with no purpose, crowds move in subtly repetitive patterns, and background characters freeze or loop. Ask what the people are doing and why. Real footage has messy, goal-directed behavior; AI scenes tend to be visually rich but behaviorally shallow.
The Arms Race: Why Detectors Keep Losing
It is worth understanding why no detector will ever be a permanent fix. The core problem is that detectors and generators train against each other. Researchers build a detector that catches the artifacts of model X; model builders fine-tune model X+1 to erase those artifacts; the detector's accuracy collapses. Each cycle costs more compute and produces a shorter-lived advantage.
Adversarial Attacks
The problem is worse when someone is deliberately trying to hide. Adversarial techniques can add tiny perturbations invisible to the human eye that break a detector's classifier. Compression, re-encoding, cropping, and overlaying text or watermarks also degrade detection accuracy. This matters because the realistic threat is not the casual AI clip, it is the clip deliberately produced to mislead. And the adversarial side is not static: as detectors improve, so do the evasion methods, which keeps the whole field in permanent motion.
Model Noise Signatures
The one detection family that has held up longest is frequency-domain analysis. Generative architectures leave a distinctive noise fingerprint in the frequency spectrum, because diffusion models and GANs imprint their own statistical structure on the image. These fingerprints are hard to remove without visibly degrading quality. However, they are model-specific: a detector trained on model A's signature often misses model B. As the number of capable models grows, maintaining a complete signature library becomes expensive.
Provenance as the Real Solution
Detection answers the question "was this generated?" after the fact. Provenance answers a better question: "where did this come from?" The industry is converging on provenance as the primary defense, with detection as a fallback.
Invisible Watermarks
The first layer is in-process watermarking: the generator embeds an invisible, machine-readable signal at the moment of creation, before the clip leaves the system. Because the watermark is applied during generation rather than after, it is extremely hard to strip without destroying quality. Think of it as a birth certificate stamped into the pixels.
Cryptographic Signatures and Ledgers
The second layer is cryptographic. A clip is signed at creation with a key tied to its origin, and the signature is recorded in a tamper-evident ledger. Anyone can later verify the signature and the chain of custody. This does not prevent synthesis, but it makes provenance verifiable and creates accountability: a model provider cannot plausibly deny that a clip came from its system, and a platform cannot claim ignorance when harmful content circulates.
The Role of AI Directors in Quality Control
There is a less obvious benefit of provenance: production quality. When content is produced through structured pipelines with strong consistency controls, the artifacts that make detection easy are also reduced. Tools that fuse multiple reference images and lock character identities across shots do not just make better content, they make content that is harder to distinguish from human work, which raises the stakes for provenance even more.
What Detection Tools Actually Do Today
Commercial detection products have converged on a small set of approaches, and it helps to know what you are actually buying. The first category is artifact analysis: the tool scans frames and audio for known synthetic signatures, such as frequency fingerprints, interpolation patterns, and lip-sync anomalies. These tools are fast and cheap, and they are useful for triage, but their accuracy depends on how recently they were trained relative to the newest models.
The second category is provenance verification. The tool does not try to judge the pixels at all; it checks for embedded watermarks and cryptographic signatures, then consults the registry to confirm the clip's origin. This category is slower to adopt because it requires cooperation from generation platforms, but it is far more reliable for the content it can verify.
The third category is hybrid review, where automated analysis produces a risk score and a human analyst makes the final call. For newsrooms and trust and safety teams, this is the pattern that scales: machines do the sifting, people do the deciding. When evaluating a detection tool, ask three questions: which models was it trained against, how does it handle re-encoded or compressed video, and does it integrate provenance checking or only pixel analysis?
What Individuals and Companies Should Do
If you are a content consumer: build a verification habit. Pause on details, check the audio, check the physics, and ask whether the clip has verifiable provenance. If a video is suspicious and important, do not share it until you have checked.
If you are a business: add synthetic-content policies to your workflow. This includes labeling AI-generated marketing and internal assets, training staff on the tells above, and preferring vendors that embed provenance signals at generation time.
If you are a creator: treat provenance as a feature. Generating with tools that watermark and sign output protects you from false accusations and makes your work easier to verify.
The Road Ahead: GAN-Enhanced Detectors and Beyond
The next phase of detection research is adversarial training on both sides of the table. New detectors are being trained in direct competition with generators, using adversarial setups that force the detector to learn the current generation's failure modes and the generator to learn the detector's blind spots. The result is a faster cycle, but also better detectors at any given moment.
Expect to see three shifts in the next few years. First, provenance becomes default infrastructure rather than a premium feature, driven by regulation and platform policy. Second, detection moves from a point-in-time tool to a continuous service, updated constantly as new models ship. Third, human judgment becomes more important, not less: automated detectors will flag and triage, but decisions about what is trustworthy will increasingly be made by informed people. Expect the human side of verification to become a formal job rather than an afterthought: the teams that handle synthetic media well will combine technical screening with editorial judgment, and the tools that succeed will be the ones that support that combination instead of trying to replace it.
Frequently Asked Questions
Are AI-generated videos always detectable?
No. With the current generation of models, a well-produced clip can pass casual inspection, and adversarial processing can defeat even good detectors. Detection is a probabilistic tool, not a guarantee.
What is the single most reliable sign?
For most viewers, small text and signage is the most reliable tell, because models still struggle to render readable text consistently. For automated systems, frequency-domain analysis and provenance checks are the strongest signals.
Will watermarks solve the problem?
Watermarks and signatures dramatically improve accountability, but they only work when the generator cooperates. Open-source and offline models can strip or omit provenance. That is why provenance and detection will coexist rather than replace each other.
Should I stop trusting video entirely?
No. The goal is calibrated trust, not blanket distrust. Apply a suspicion gradient: a low-stakes entertainment clip does not need a forensic audit, but a video claiming to show a new product, a politician, or a disaster should be checked before it shapes your opinion or gets shared.
