What We Are Building

Video detection is experimental

~76% accuracy. Short clips and low-resolution video are less reliable. Do not use for high-stakes decisions alone.

Short audio clips are unreliable

Clips under 5 seconds do not give the model enough signal. Results on very short audio should be treated as indicative only.

Non-native English may trigger false positives

Text detection models were primarily trained on English. Non-native phrasing patterns can sometimes appear machine-like to the model.

Human-edited AI content is harder to detect

When AI output has been manually edited or paraphrased, detection accuracy drops. There is no perfect detector.

Results are probabilistic, not definitive

All verdicts come with a confidence score. Use human judgment alongside Aiscern results for any consequential decision.