Think of these percentage scores like a risk score, the lower the % the less risk of a deepfake so the closer the score is to “1%” the more confident we are the media is real. The higher the percentage (e.g. a score at or close to 99%) the more confident we are the media file is fake.
An important concept to understand is Probabilistic vs. Deterministic, our models use probabilistic inference so we never know the ground truth at 100% certainly. Thus we live in a world from 1 - 99%.
Audio Detection:
Authentic Range (0% - 55%): "Real Human Voice"
"When we analyze audio and get a score between 0-55%, our models are telling you 'This sounds like a real person speaking.' The lower the number, the more confident we are it's genuine human speech."
"Think of 10% as 'definitely human' and 55% as 'probably human, but we're keeping an eye on it.'"
Suspicious Range (55% - 65%): "Proceed with Caution"
"Scores in this narrow band mean 'Something doesn't sound quite right.' It's like when you hear someone on the phone and think 'That doesn't sound like them.' We can't definitively call it fake yet, but we're raising a yellow flag."
Artificial Range (65% - 100%): "AI-Generated Voice"
"Once we hit 65% and above, we're confident this is AI-generated speech. At 80%, we're very sure. At 95%, we'd bet money on it being deepfake audio."
Video Detection:
Authentic Range (0% - 45%): "Real Human Video"
"For video, our 'real person' threshold is a bit tighter. 0-45% means we're seeing natural human movements, expressions, and visual cues that are consistent with genuine footage."
Suspicious Range (45% - 55%): "Needs Investigation"
"This is our 'something's off' zone. Maybe the lip-sync isn't perfect, or facial movements seem slightly unnatural. We're not calling it fake yet, but we recommend human in the loop review."
Artificial Range (55% - 100%): "AI-Generated Video"
"Above 55%, we're detecting clear signs of AI manipulation. This could be face swaps, synthetic expressions, or entirely user generated. The higher the score, the more obvious the manipulation."
Image Detection:
Authentic Range (0% - 45%): "Real Photograph"
"For still images, 0-45% indicates an authentic image. We're looking at lighting, shadows, pixel layer patterns, and facial consistency that match real-world images."
Suspicious Range (45% - 55%): "Questionable Content"
"Images in this range have some inconsistencies. Maybe the lighting doesn't match, or there are subtle synthetic artifacts. It warrants a closer look."
Artificial Range (55% - 100%): "AI-Generated Image"
"Above 55%, we're detecting clear signs of AI generation - whether that's an AI-created face, manipulated features, or a completely user generated synthetic image."
The confidence or “risk” score is our way of translating complex AI analysis into a simple percentage you can act on.
"Remember: Higher percentages mean higher risk of deepfake. When you see scores above our 'suspicious' thresholds, that's your cue to implement additional verification steps and or takedown procedures."