Accuracy & Results
SeeThrough shows promising results in controlled tests but no detection system is perfect. Accuracy varies based on video quality, length, content type, and the sophistication of the AI generator. Use SeeThrough as one tool among many -- not as the final word.
False positives can happen for several reasons: heavily edited videos, unusual motion patterns, artificial visual effects, heavily compressed footage, or unusual lighting. Our system isn't perfect -- treat it as a screening tool, not definitive proof.
We provide two metrics: an Authenticity Score (0-100%) and Model Confidence (0-100%). Authenticity 70-100% = likely real; 30-70% = uncertain; 0-30% = likely AI-generated. Confidence tells you how sure the model is. These are probabilistic estimates -- always consider the full context.
Perfect detection is mathematically impossible. AI generators are trained to mimic reality, real videos can share artifacts with AI ones, and new generation techniques constantly emerge. Every detector trades off false positives against false negatives -- these are fundamental limits.
Technology & Coverage
SeeThrough uses a multi-signal approach combining deep visual features with temporal frequency analysis. We examine frame-level semantics, temporal consistency, and motion coherence across the whole video -- not just individual frames -- to spot patterns AI generators struggle to maintain consistently.
No. SeeThrough is optimized for diffusion-based generators like Sora 2, Runway, and Pika. Videos made with GANs, face-swap deepfakes, or traditional editing may not be reliably detected. As generation technology evolves, detection must keep pace.
Yes -- it's an ongoing arms race. As generators improve, evasion becomes easier. This is why we frame SeeThrough as a research tool and screening aid, not an infallible solution. We update models as new generators emerge, but a gap will always exist.
Key limitations: (1) Rapidly evolving generators create unseen artifacts. (2) Heavily compressed or low-quality video causes false positives. (3) Clips under 3 seconds provide limited temporal data. (4) Hybrid content mixing real and AI can confuse detectors. (5) Adversarial techniques can deliberately evade detection.
Usage
SeeThrough performs best with videos longer than 5 seconds, 720p or higher resolution, content with movement and scene changes, and footage that hasn't been heavily re-compressed. Challenging cases include very short clips, extremely low quality video, static scenes, and heavily edited content.
Most videos under 30 seconds are analyzed in 10-30 seconds. Longer or higher-resolution videos may take 1-2 minutes. Our system processes videos in chunks to analyze temporal patterns, which takes more time than simple frame-by-frame checks.
No single tool should be the sole basis for important decisions. SeeThrough is a screening tool. For high-stakes situations -- legal proceedings, journalism -- combine our analysis with metadata inspection, expert review, source verification, and digital forensics professionals.
Privacy
Uploaded videos are processed in real-time and temporarily stored only during analysis. We don't permanently store your videos or share them with third parties. For sensitive content, note that any cloud analysis service involves uploading data -- don't upload confidential videos to the demo.
── TRY IT ──
See it in action.
Upload a TikTok video and get a verdict with confidence scores in seconds.
Still have questions? Visit the about page to learn more.