Ethics, Privacy, And Audio Notice

Why the release boundary is intentionally conservative

CantHarm contains harmful lexical content and waveform audio. This page states the ethics and privacy boundary directly instead of leaving it implicit.

Harmful-content notice

Content boundary

The benchmark includes harmful, offensive, or otherwise sensitive lexical material. Reviewers should expect explicit insulting, illicit, sexual-obscene, political, and extremist categories in the released label inventory.

Audio privacy notice

Waveform-specific caution

The audio layer consists of controlled canonical recordings from 11 pseudonymous speaker ids. It should not be repurposed for speaker recognition, biometric modeling, voice cloning, or any re-identification task.

Research-use framing

What the benchmark is for

  • Benchmark comparison under a documented protocol
  • Analysis of ambiguity, source disagreement, and speech grounding difficulty
  • Reproducibility and audit checks anchored to the manifest and reviewer docs

Non-deployment framing

What the benchmark is not for

  • Automatic punitive moderation decisions
  • Claims of real-world moderation readiness
  • Claims of contextual utterance coverage or broad speaker generalization
Limitations should be read as design boundaries. CantHarm is lexical-form-centered, uses one controlled clip per retained form, has non-speaker-disjoint splits, and exposes a rubric-driven scalar severity target. These are deliberate scope boundaries, not hidden omissions.