copyright-censor

Preemptive copyright/trademark string check for music (and image/video) slider prompts.

The library is a conservative filter: high recall, false positives are acceptable, false negatives are not. It is not legal advice, not a takedown engine, and not a pirate tool. It never ships copyrighted lyrics and never scrapes lyric sites.

A concept-slider UI can call check() on every keystroke and highlight the exact words that look risky.

Live demo

GitHub Pages: difflabai.github.io/copyright-censor

Local (from repo root so ESM + the gzip catalog resolve):

npm run demo

Open http://localhost:4173/ or http://localhost:4173/demo/. Sample chips cover allow / review / block (chart titles, franchises, reproduce-this heuristics).

Policy

Prompt	Verdict
Generic mood / era / production language (dreamy 80s synthwave)	allow
Cataloged artist, no specific work (in the style of …)	review
Specific cataloged work, franchise, trademark, or “reproduce this” phrasing	block

Overall verdict is the worst span: block > review > allow.

How coverage works

check() does not scan a 200-row JSON list. At build time, public identifier datasets are compiled into a compact token-level Aho-Corasick automaton (data/catalog.bin.gz):

• MusicBrainz artist names (CC0 extract)
• MusicBrainz recording titles (title field only)
• TV / game / anime / manga names
• Wikidata labels for media franchises, film series, video-game series, fictional universes

Rebuild with npm run build-catalog. The default check() path is still synchronous and local — the shipped gzip is loaded once at import, then every keystroke walks the automaton. No network at check() time. An optional rebuild/refresh exists only as a maintainer script.

This is high-recall coverage of cataloged identifiers, not a legal oracle and not 100% of every copyrighted sentence ever written. Unknown works, unpublished titles, and generic lyric lines will not match unless a heuristic fires.

Identifiers only. The compile step never reads synopsis, plot, or lyric columns.

Install

npm install copyright-censor

ESM only, browser + Node 20.10+, no runtime dependencies.

API

import { check, checkPair, createCensor, catalogMeta } from 'copyright-censor';

catalogMeta.patterns; // compiled identifier count

check('dreamy 80s synthwave');
// { verdict: 'allow', spans: [], reasons: [] }

checkPair({
  positive: 'warm analog pads',
  negative: 'no Neon Glass Harbor',
});

spans are character offsets into the original string, so a textarea or contenteditable can mark words live.

check(text, options?)

Returns { verdict, spans, reasons }.

checkPair({ positive, negative }, options?)

Checks both slider fields. Worst verdict wins. Each span also has field: 'positive' | 'negative'.

createCensor(options?)

Load the automaton once, then call censor.check() / censor.checkPair() on every keystroke.

const censor = createCensor({
  blocklist: myJson,          // extra identifiers merged onto the catalog
  extraTerms: [{ term: 'My Label Artist', kind: 'artist' }],
  media: 'music',             // or 'image' | 'video' | 'all'
  allowlist: ['harbor'],      // never flag these phrases
});

Set replaceBlocklist: true / replaceCatalog: true to ignore the shipped extras / compiled catalog (useful in tests).

Options

Option	Meaning
blocklist	User JSON extras. Merged onto the catalog.
replaceBlocklist	Ignore shipped extras and the compiled catalog.
replaceCatalog	Ignore the compiled catalog only.
extraTerms	Extra { term, kind, verdict? } rows.
media	Drop overlay entries that do not apply.
allowlist	Extra mood/era phrases that must never flag.

Extras JSON

data/blocklist.json is an optional overlay, not the engine. Edit it or pass your own:

{
  "artists": ["Zorblin Faye", { "term": "Queen", "commonWord": true }],
  "works": ["Neon Glass Harbor"],
  "franchises": ["Crystal Finch Saga"],
  "trademarks": [{ "term": "Lumistitch", "media": ["image", "video"] }]
}

• artists default to review
• works / franchises / trademarks default to block
• commonWord: true only flags a capitalized token or a nearby rights cue
• Single-token common English work titles are dropped at compile time (love, home, stay) so the slider stays usable
• Matching is word-boundary, case-insensitive, accent-folded

Do not add copyrighted lyrics. Titles, names, and franchise labels only.

Heuristics (no stored lyrics)

These fire without a catalog hit:

• Reproduction: cover of, lyrics from, word for word, recreate the song, chorus:, screenshot from, …
• Likeness: in the style of, inspired by, voice of, official logo
• A long quoted passage (8+ words) is treated as a pasted lyric-like excerpt — tests use invented or public-domain strings only

CLI

npx copyright-censor "dreamy 80s synthwave"
npx copyright-censor --json "cover of Whispering Shadows"
npx copyright-censor --positive "warm pads" --negative "no crowd noise"
npx copyright-censor --blocklist ./my-list.json --media music "…"

Exit codes: 0 allow, 1 review, 2 block, 64 usage.

What this is not

• Not legal advice and not a guarantee you are clear to ship a generation
• Not a lyrics database, fingerprint, or Content ID replacement
• Not 100% of every copyrighted sentence — only cataloged identifiers plus intent heuristics
• Not a tool for finding, storing, or reproducing copyrighted works

If a lawsuit would wreck the product, treat review as “do not send to the model until a human looks at it,” and block as “do not send.”

License

MIT