Responsible AI Disclosure

AI & Model Limitations Disclosure

Updated: October 11, 2026 • Statutory Disclosure

1. Probabilistic Nature of Acoustic Inference

VoiceShield inference engines output a calibrated probability score [0.0 - 1.0] representing acoustic confidence that an audio stream contains synthetic artifacts. A risk score is not an absolute legal determination of criminal fraud.

2. Telephony Degradation & Cellular Codecs

Narrowband cellular telephony (such as AMR 8kHz or VoIP jitter buffers) can strip natural high-frequency voice harmonics. While VoiceShield employs Kalman state-space estimation to compensate, severely degraded audio may exhibit elevated uncertainty intervals.

3. Mandatory Human-in-the-Loop Protocol

In accordance with international responsible AI directives, VoiceShield is designed as a decision-support copilot for qualified security analysts. Critical actions (e.g. freezing bank accounts or denying emergency services) must incorporate human review.

4. Model Drift & Retraining

As adversarial generative vocoders advance, acoustic baselines require periodic model fine-tuning. We release model updates and publish corresponding changelog entries to preserve high detection sensitivity.