Is this video real or AI-generated? How to check

How FactHeck analyses video for AI generation

When you submit a TikTok, Instagram Reel, or YouTube video, FactHeck runs three analysis dimensions:

1. Visual artefacts — sampled keyframes are sent to Claude (Sonnet) with a structured prompt covering eight artefact categories: anatomy and hands, garbled text, texture inconsistencies, geometry, lighting and shadows, background anomalies, AI style indicators, and compositing seams. Each category is scored for severity.
2. Temporal consistency — this is the dimension that most reliably separates real from AI video. FactHeck extracts dense keyframes and groups them into consecutive windows (typically 3 frames approximately 0.5 seconds apart), then passes those windows to Gemini (Flash) to look for object morphing, appearance and disappearance of elements, background warping, physics violations, body and face changes between frames, texture flickering, and abrupt lighting shifts. Normal camera shake and motion blur are explicitly excluded as false positives.
3. Scenario plausibility — separately, the pipeline asks whether the depicted scene could realistically exist. Impossible physics, pristine dreamlike environments, absent safety equipment in dangerous scenarios, no bystander reaction to emergencies, and stacked implausibilities are each scored. Three or more implausibilities together are a high-confidence indicator of AI generation.

Each dimension returns a verdict ("Likely AI", "Partially AI / AI Edited", or "Likely Real") and a confidence level. These are combined into the final result shown on your check page.

Why temporal consistency is the key signal

Still-image AI detectors have trained generators to produce better-looking individual frames. The temporal dimension is harder to fake: generating hundreds of frames that are photorealistic and physically consistent across time is a much harder problem. Common temporal tells in current AI video:

• A person's hand changes shape or finger count between frames 0.5 seconds apart
• A background element (tree, sign, person) appears or vanishes with no explanation
• The background warps during a pan in a way that doesn't match real-world parallax
• A falling object accelerates at the wrong rate or bounces with impossible energy
• Clothing texture or a person's face subtly changes between consecutive frames

Common scenario red flags in AI video

Many of the most-shared AI videos on TikTok and Instagram follow recognisable patterns that the plausibility dimension is designed to catch:

• People falling from dramatic heights (cliff waterslides, roof edges, canyon platforms)
• Impossible structures built into natural landscapes with no engineering rationale
• Apparent serious injury or death that is still live on a major platform — real footage of this is routinely removed by content moderators
• No bystander reaction to a horrifying event; calm, cinematic observation of something that would cause panic
• Pristine fantasy locations with zero mundane detail (no signage, no litter, no other visitors)

The presence of three or more of these signals in the same video is a high-confidence indicator of AI generation, regardless of how convincing individual frames look.

Honest accuracy caveats

• Platform re-compression strips per-frame artefacts. TikTok, Instagram, and YouTube re-encode every video they host. Fine-grain visual tells can be washed out by compression. The temporal-consistency dimension is more robust because it looks at changes between frames rather than within a single frame, but a heavily compressed clip can still return a lower-confidence result.
• Newer AI video models produce fewer tells. The temporal-consistency analysis is currently the most reliable dimension, but the best current models already produce more stable temporal output than first-generation tools. Expect this arms race to continue.
• Unusual real footage can trigger false positives. Shaky handheld video, extreme slow motion, and heavily colour-graded footage can resemble AI artefacts or implausible scenarios. Use context and source alongside any detector result.
• No audio-specific deepfake detection. FactHeck checks whether the audio is consistent with the visuals (for example, whether a synthetic-sounding voice matches the visible speaker), but does not run a dedicated audio-deepfake classifier. If your concern is specifically about synthetic speech, a specialist audio tool will give you a more focused signal.

For further background on the state of AI video detection, see MIT Technology Review's coverage of deepfake detection research and First Draft's guide to AI-generated video verification.

How to use the result

• A high-confidence "Likely AI" result with multiple dimensions flagged — especially temporal inconsistency — is a strong reason for caution. Don't share until you have investigated further.
• A low-confidence or "Partially AI / AI Edited" result is inconclusive. One weak signal is not enough to be certain either way.
• A "Likely Real" result does not certify authenticity. The video could still be real footage that is misleadingly captioned, taken out of context, or from a generator the pipeline hasn't been optimised against.

Frequently asked questions