Disclaimer

Effective date: 2026-04-14

ArtifactNet provides forensic indicators, not conclusive legal proof. Verdicts are statistical estimates based on machine learning models and are subject to error. Do not use ArtifactNet as the sole basis for legal action, public accusation, contract enforcement, or copyright dispute resolution.

1. What ArtifactNet Does

ArtifactNet analyzes audio recordings and predicts the probability that the recording contains AI-generated content. It returns one of three categories:

• AI: high-confidence indication of AI-generated content
• Real: high-confidence indication of human-performed content
• Unknown / Uncertain: insufficient confidence to make a determination

2. What ArtifactNet Does NOT Do

• It does not identify the specific AI model or service that generated the audio (although it may suggest a likely generator family).
• It does not detect whether a song was composed by AI but performed by humans.
• It does not detect post-processing such as autotune, vocoder, or stem replacement on otherwise human recordings.
• It does not provide a forensic chain of custody admissible as legal evidence without independent expert review.

3. Known Limitations

3.1 Codec sensitivity

Audio recompression (mp3 / aac / opus) can shift verdicts. We continuously improve codec invariance, but heavily processed mp3 streams (e.g., low-bitrate YouTube re-encodings) may produce false positives or false negatives.

3.2 Generator coverage

ArtifactNet is trained on commercial generators known as of the model release date. Newer or private models may produce outputs that deviate from learned distributions. We update the model regularly to address this.

3.3 Hard cases

The following categories have higher error rates:

• Lo-fi instrumentals (real but with synthetic textures)
• Heavily quantized or auto-generated MIDI renderings
• Stable Audio outputs (lower TPR than other generators)
• Live recordings with extreme dynamic range

3.4 Statistical nature

Even at 99% accuracy on a benchmark, 1 in 100 verdicts will be wrong. Treat each verdict as one input among many in your decision process, not as ground truth.

4. Recommended Use

• Music platforms: as a triage signal for review queues, not for automatic content removal.
• Rights holders: as supporting evidence alongside acoustic fingerprinting (AcoustID, audible.com) and provenance metadata.
• Researchers: as one of several detection tools in a comparative framework, with appropriate citation.
• Journalists: as a preliminary check before independent verification.

5. Not Recommended Use

• Automated content takedowns based solely on ArtifactNet verdicts
• Public accusations of artists without independent corroboration
• Legal evidence in court without expert witness analysis
• Real-time content moderation at scale without human review of edge cases

6. Dispute Resolution

If you believe ArtifactNet has incorrectly classified your work, you may request a manual review by emailing [email protected] with:

• The audio file or a stable URL
• The verdict you received
• Any supporting evidence (recording session metadata, stems, raw takes)

We aim to respond within 5 business days.

7. No Warranty

The Service is provided "AS IS" without warranties of any kind, express or implied, including but not limited to merchantability, fitness for a particular purpose, or non-infringement. We do not warrant that verdicts are accurate, complete, or suitable for any particular use.