AI Disclosure

How SongForgeAI discloses AI involvement.

Every lyric, score, and critique on this platform is AI-generated. This page documents the disclosure spec, the content boundaries, the tamper-evident provenance signal, and the takedown contact path. EU AI Act-aligned and forward-compatible with the August 2026 visible-labeling + machine-readable-metadata obligations.

Last reviewed 2026-05-21. Provenance spec v1.0.0.

1. What’s AI on this platform

SongForgeAI uses Anthropic’s Claude models (Claude Opus 5.5 writes, Claude Sonnet 4.6 scores, and Claude Haiku 4.5 handles the structured steps) to produce every lyric, every score, and every critique on the platform. The pipeline:

Lyrics — A 50-voice writing room (the Forge) drafts lyrics in multiple rounds, with adversarial critique + targeted refinement. Output: a complete song lyric scored by the published 12-metric rubric.
Scores — An 8-voice eval panel applies the open Lyric Scoring Standard. Composite score, grade, per-metric breakdown, transcendent lines, wounds.
Critique — The Crucible (free, no-login) runs an 8-voice adversarial panel against any user-pasted lyric, producing a verdict in ~30 seconds.
Fidelity audit — An 8-component composite that measures whether the lyric served the brief. The orthogonal axis to quality (see The Fidelity Standard).

Every output is AI-generated. There is no human writer ghostwriting in the background; there is no curated content library being remixed. The system produces fresh output for each request based on the user’s prompt.

2. Where we display the disclosure

Every surface that renders an AI-generated lyric carries an in-view disclosure label:

AI-generated content — lyrics produced by SongForgeAI’s 50-voice writing room.

The label appears on:

/s/[slug] — every public lyric permalink
/forge — the forge result screen, beneath the lyric body
/dashboard — the SongDetail view inside the user’s library
The /api/v1/score response carries a provenance field for downstream consumers who need to preserve the disclosure through their own pipelines

3. Machine-readable provenance

Every public lyric page emits a structured provenance payload via HTML meta tags + (for API responses) a JSON payload. The spec carries five fields:

<meta name="ai-content-provider" content="SongForgeAI">
<meta name="ai-content-generated-at" content="2026-05-20T12:00:00Z">
<meta name="ai-content-hash" content="sha-256:abc123...">
<meta name="ai-content-spec" content="eu-ai-act-provenance-1.0.0">
<meta name="ai-content-url" content="https://songforgeai.com/s/<slug>">

The hash is SHA-256 over the canonical content payload (for lyrics: <title>\n\n<lyrics>; for scores: the composite + grade + sorted metric scores). Same input, same hash, forever.

How to verify a content hash

Given any AI-generated lyric on this platform, you can independently verify its provenance hash:

Read the ai-content-hash meta tag (or the provenance.contentHash field if you fetched from the API).
Reconstruct the canonical content payload: <title>\n\n<lyrics> with literal newlines between title and lyrics.
Compute SHA-256 over the UTF-8 bytes of that payload.
Compare. If the hashes match, the content hasn’t been tampered with since publication. If they don’t, the content has been modified — by us (with a re-published hash) or by a downstream redistributor (whose hash will not match).

Reference implementation: src/lib/ai-disclosure/provenance.ts in the public repo at github.com/torgin-dev/songforge-ai.

4. What this platform will never generate

Regardless of register, regardless of prompt phrasing, regardless of user request, SongForgeAI blocks four categories of content. These are non-negotiable:

1. Sexual content involving minors

Any sexual or sexualized depiction of anyone under 18. Hard block at the model boundary; hard block at the output classifier; account suspension on attempted generation.

2. Hate speech or slurs targeting protected groups

Content that dehumanizes, threatens, or incites violence against people on the basis of race, ethnicity, religion, gender, sexual orientation, disability, or national origin.

3. Credible threats of violence, doxxing, harassment

Specific threats against named individuals; release of private contact information; orchestrated harassment. Creative violence in storytelling is OUT of scope; this rule is about real-world targeting.

4. Instructions facilitating illegal harm

Working synthesis routes for weapons or controlled substances, instructions for human trafficking, financial fraud playbooks. Songs referencing real events are fine; how-to manuals are not.

5. Takedown contact + response time

If you encounter AI-generated content on this platform that violates one of the four hard nevers above, OR believes their likeness / identity / copyrighted material was used without consent, contact:

support@songforgeai.com

Response time: We aim to acknowledge takedown reports within 3 hours during operator availability and within 24 hours otherwise. Verified violations are removed within 24 hours.

Note: this SLA is a stated commitment, not a contractual guarantee. Real takedown response times are logged at /incidents for accountability.

Related public artifacts

The Lyric Scoring Standard — the 12-metric open rubric every score is computed against.
The Fidelity Standard — the 8-component composite that scores brief adherence.
Operating principles — how decisions get made; principles paired with the inversions we reject.
Public incident log — where takedown response times + content-violation incidents are logged.
Manifesto — why this product exists + what it refuses to be.

This page is the canonical destination for the “How it works” links on every AI-content disclosure surface in the product. Updates are dated + appended; the underlying provenance utility is open-source.