Comprehensive Cross-Media Intellectual Property Protection for the Oshun Ecosystem#
Executive Summary#
Oshun is not a music platform. It is a universal creative intelligence ecosystem that generates content across every media type known to human civilization — music, visual art, video, 3D models, text, game worlds, characters, narratives, animations, fashion, educational content, journalism, virtual environments, and more. Thirty-one domains produce creative output. Every piece of generated content is a potential IP liability.
The Themis Music Shield (Phase 74) addresses music copyright. But music is one media type among dozens. A generated image can infringe visual copyright. A generated storyline can reproduce a copyrighted plot. A generated character can violate right of publicity. A generated 3D model can infringe design patents. A generated article can constitute plagiarism. A generated game level can copy protected level design. A generated fashion design can violate trade dress.
This proposal establishes the Themis Universal Originality Shield — a unified, cross-media IP protection framework that gates every piece of AI-generated content across the entire Oshun ecosystem before it reaches publication. It extends the principles of the Music Shield to all media types, with media-specific detection engines and a unified risk scoring, provenance, and review infrastructure.
Scale of the Problem#
| Media Type | Generating Domains | Daily Volume (Projected) | IP Risk Level |
|---|---|---|---|
| Music/Audio | Euterpe, Calliope, Isis, Nous | 10,000+ tracks | CRITICAL |
| Visual Art/Images | Isis, Calliope, Yemaya | 50,000+ images | CRITICAL |
| Video/Film | Isis, Calliope, Veritas, Yemaya | 5,000+ clips | HIGH |
| 3D Models/Assets | Isis, Maya, Bellona | 20,000+ assets | HIGH |
| Text/Writing | Veritas, Hathor, Iris, Nisaba, Metis | 100,000+ pieces | HIGH |
| Game Content | Hathor, Maya, Yemaya | 10,000+ assets | MEDIUM-HIGH |
| Characters/Personas | Calliope, Hathor, Psyche, Aphrodite | 1,000+ characters | HIGH |
| Narrative/Storylines | Hathor, Veritas, Calliope, Yemaya | 5,000+ narratives | MEDIUM-HIGH |
| Animation/Motion | Aja, Calliope, Maya | 5,000+ sequences | MEDIUM |
| Fashion/Style | Galatea, Aglaea, Calliope | 2,000+ designs | MEDIUM |
| Voice/Speech | Nous, Euterpe, Calliope, Psyche, Iris | 50,000+ utterances | CRITICAL |
| Educational Content | Metis, Nisaba, Kuanyin | 10,000+ lessons | MEDIUM |
| Architecture/Spaces | Maya, Hathor | 5,000+ environments | LOW-MEDIUM |
| Recipes/Food Content | Hestia | 1,000+ recipes | LOW |
At these volumes, even a 0.01% infringement rate produces dozens of potential legal incidents daily. The only defensible position is comprehensive, automated originality verification across every media pipeline.
Part I: Legal Landscape by Media Type#
Understanding what is protectable — and what is not — for each media type is essential to calibrating detection systems. Overly aggressive detection wastes human review capacity on non-issues. Under-aggressive detection exposes the platform to litigation. The goal is calibration grounded in actual law.
1. Music & Audio#
Covered comprehensively by the Themis Music Shield proposal. Summary of protectable elements:
| Element | Protectability | Key Precedent |
|---|---|---|
| Melody | Strong — the #1 litigation vector | Bright Tunes v. Harrisongs (1976) |
| Lyrics | Strong — literary copyright | All lyrics cases |
| Harmony/chords | Weak — common progressions unprotectable | Griffin v. Sheeran (2023) |
| Rhythm/groove | Very weak | Gray v. Perry (reversed, 2020) |
| Timbre/sound | Generally not protectable | No direct precedent |
| Voice identity | Emerging — NO FAKES Act, state laws | Midler v. Ford (1988) |
| Song structure | Not protectable | Universal convention |
| Sound recordings | Strong — separate from composition | 17 U.S.C. § 114 |
AI-specific risk: Model memorization of training data can reproduce copyrighted melodies, lyrics, and vocal characteristics verbatim.
2. Visual Art & Images#
Generated by: Isis (text-to-image), Calliope (artist visual identity), Yemaya (production art), Aglaea (beauty/style), Galatea (fashion).
| Element | Protectability | Legal Basis |
|---|---|---|
| Specific composition/arrangement | Strong | Mannion v. Coors Brewing (2006) |
| Color palette alone | Not protectable | Functional/common |
| Art style/technique | Not protectable | Style is idea, not expression |
| Specific characters | Strong — visual appearance | DC Comics v. Towle (2013) |
| Typefaces/fonts | Weak in US, stronger in EU | Eltra Corp. v. Ringer (1978) |
| Logos/trademarks | Trademark law, not copyright | Lanham Act |
| Photographic subjects | Not protectable (facts) | Feist v. Rural (1991) |
| AI-generated output | Uncopyrightable without human input (US) | Thaler v. Perlmutter (2023) |
Key risks for AI image generation:
- Style mimicry lawsuits — Artists suing AI generators for "style theft" (Andersen v. Stability AI, ongoing as of 2025). While style is not copyrightable, reproducing specific compositions or character likenesses is.
- Training data memorization — Diffusion models can reproduce training images verbatim or near-verbatim. Getty Images v. Stability AI (2023) specifically alleges reproduction of Getty watermarks in outputs.
- Character reproduction — Generating images of copyrighted characters (Mickey Mouse pre-public domain, Marvel characters, anime characters).
- Trademark inclusion — Generated images inadvertently including brand logos, product designs, or trade dress.
- Right of publicity — Generated images resembling real people without consent.
- Deepfakes — Generated images placing real people in fabricated scenarios.
What detection must cover:
- Visual similarity to copyrighted works (composition, specific elements)
- Character recognition (copyrighted fictional characters)
- Trademark/logo detection (brand marks in generated images)
- Face recognition (real person likeness detection)
- Training data memorization (near-duplicate detection against training sets)
- Style attribution (when style crosses into specific expression copying)
- Watermark reproduction (detecting reproduced watermarks from training data)
3. Video & Film#
Generated by: Isis (text-to-video), Calliope (music videos), Veritas (news video), Yemaya (film production), Uzume (live event).
| Element | Protectability | Legal Basis |
|---|---|---|
| Specific shot composition | Moderate | Cinematographic expression |
| Editing sequence/montage | Moderate | Creative arrangement |
| Script/dialogue | Strong | Literary copyright |
| Visual effects | Moderate | Creative expression |
| Camera movements | Weak — technique | Standard cinematography |
| Color grading/look | Weak | LUT not protectable |
| Scene recreation | Strong if specific enough | Substantial similarity |
| Documentary footage | Facts not protectable, expression is | Feist principle |
Key risks for AI video generation:
- Scene reproduction — AI video models trained on film/TV can reproduce specific scenes, shot compositions, or visual sequences from copyrighted works.
- Character depiction — Generated video depicting copyrighted characters in motion.
- Real person depiction — Generated video of real people (deepfakes). California AB 2655 and other state laws specifically target this.
- Dialogue reproduction — Generated dialogue that reproduces copyrighted screenplay text.
- Score/soundtrack — Generated video with AI music that infringes (covered by Music Shield, but video pipeline must integrate).
- News footage — Veritas generating news content that reproduces copyrighted news footage or reporting.
What detection must cover:
- Scene-level visual similarity (shot composition, editing patterns)
- Dialogue/script similarity (text comparison of spoken content)
- Character recognition in video frames (same as image, per-frame)
- Face recognition in video (real person identification, per-frame)
- Audio track analysis (delegate to Music Shield)
- Trademark detection in video frames
- Deepfake detection (synthetic media identification)
4. 3D Models & Assets#
Generated by: Isis (3D generation via NeRF/Gaussian splatting/Point-E), Maya (procedural generation), Bellona (asset export).
| Element | Protectability | Legal Basis |
|---|---|---|
| Original 3D sculpture/model | Strong | Sculptural copyright |
| Functional 3D design | Design patent, not copyright | Baker v. Selden (1880) |
| 3D character model | Strong | Character copyright |
| Architecture (real buildings) | Complex — Architectural Works Act | 17 U.S.C. § 120 |
| Vehicle/product design | Design patent + trade dress | Automotive design cases |
| Game asset (weapon, prop) | Moderate | Creative expression |
| Terrain/landscape | Weak — natural features | Scenes à faire |
Key risks for AI 3D generation:
- Character model reproduction — Generating 3D models that closely resemble copyrighted characters from games, films, or animation.
- Product design copying — Generating 3D models of trademarked products (Nike shoes, Apple devices, luxury goods).
- Architectural copying — Generating 3D buildings that reproduce real copyrighted architectural works.
- Asset pack memorization — Models trained on 3D asset libraries reproducing exact meshes from commercial packs.
- Trade dress violation — 3D designs that are "confusingly similar" to established product designs.
What detection must cover:
- 3D shape similarity (mesh comparison, point cloud matching)
- Texture/material similarity
- Character model recognition (known copyrighted 3D characters)
- Product/brand identification in 3D models
- Architectural work comparison
- Asset library deduplication (detect training data memorization)
5. Text & Written Content#
Generated by: Veritas (journalism, articles), Hathor (dialogue, narrative), Iris (conversational responses), Nisaba (scholarly analysis), Metis (educational content), Calliope (artist bios, social content).
| Element | Protectability | Legal Basis |
|---|---|---|
| Specific word sequences | Strong — literary copyright | Standard copyright |
| Plot/storyline | Expression protectable, ideas not | Nichols v. Universal (1930) |
| Character personalities | Protectable if sufficiently delineated | DC Comics, Paramount |
| Facts/data | Not protectable | Feist v. Rural (1991) |
| Short phrases/titles | Generally not protectable | De minimis |
| Compilations | Selection/arrangement protectable | Feist |
| News facts | Not protectable | INS v. AP (limited) |
| Academic ideas | Not protectable | Idea/expression dichotomy |
Key risks for AI text generation:
- Verbatim reproduction — LLMs memorize and reproduce training text. Demonstrated extensively in research. The NYT v. OpenAI (2023) lawsuit specifically alleges near-verbatim reproduction of NYT articles.
- Plot reproduction — Generated narratives that closely follow copyrighted plot structures. The "abstraction-filtration-comparison" test (Altai, 1992) determines the protectable expression level.
- Character reproduction — Generating characters with the same name, personality traits, backstory, and relationships as copyrighted characters.
- Academic plagiarism — Educational content that reproduces textbook passages or scholarly work without attribution.
- News plagiarism — Generated journalism that reproduces another outlet's reporting without attribution. AP v. Meltwater (2013) establishes aggregation limits.
- Style guides and brand voice — Generated marketing copy that inadvertently reproduces another brand's distinctive voice or taglines.
What detection must cover:
- Text similarity (n-gram, embedding-based, exact match)
- Paraphrase detection (same meaning, different words)
- Plot/narrative structure comparison
- Character profile similarity
- Fact vs. expression separation
- Source attribution verification
- Trademark/brand name detection in text
- Academic integrity checking (citation verification)
6. Characters & Personas#
Generated by: Calliope (AI artist personas), Hathor (game NPCs), Psyche (AI avatars), Aphrodite (intimate personas).
| Element | Protectability | Legal Basis |
|---|---|---|
| Visual appearance | Strong | Character copyright |
| Name | Trademark if distinctive | Lanham Act |
| Personality/traits | Protectable if "sufficiently delineated" | Nichols test |
| Voice | Right of publicity | Midler v. Ford (1988) |
| Backstory | Expression protectable, archetype not | Idea/expression |
| Catchphrases | Trademark if distinctive | Lanham Act |
| Likeness to real person | Right of publicity | State laws |
Key risks:
- Accidental character reproduction — AI generating characters that closely match existing copyrighted characters in name, appearance, personality, or backstory combination.
- Celebrity likeness — Generated personas that resemble real celebrities in appearance or voice.
- Archetype vs. copy — The "brooding antihero" is an archetype (not protectable), but "a billionaire who builds a suit of power armor and has a witty personality" is specific enough to be Tony Stark.
- Cross-media character infringement — A character designed for a game that inadvertently reproduces an anime, comic, or film character.
What detection must cover:
- Character visual similarity (against database of known characters)
- Character profile comparison (name + traits + backstory compound matching)
- Celebrity likeness detection (face/voice matching against real people)
- Name collision detection (against trademark databases)
- Archetype vs. specific character differentiation
- Cross-media character matching
7. Narrative & Storylines#
Generated by: Hathor (game narratives, quests), Veritas (long-form content), Calliope (artist career narratives), Yemaya (film scripts).
| Element | Protectability | Legal Basis |
|---|---|---|
| Specific plot points (expression) | Protectable | Sheldon v. Metro-Goldwyn (1936) |
| General themes/ideas | Not protectable | Nichols v. Universal (1930) |
| Plot structure (hero's journey) | Not protectable — scènes à faire | Standard narrative |
| Specific dialogue | Strong — literary copyright | Direct expression |
| Character arcs | Moderate — depends on specificity | Abstraction-filtration |
| World rules/magic systems | Moderate if sufficiently original | WB v. RDR Books (2008) |
| Quest/mission design | Weak — gameplay mechanics | Tetris v. Xio (limited) |
Key risks:
- Plot lifting — Generated storylines that follow the specific plot of a copyrighted work too closely. The line between "inspired by" and "copied from" is the abstraction-filtration-comparison test.
- Fan fiction boundary — Generated content that continues, extends, or reimagines copyrighted narratives without license (derivative works).
- World-building reproduction — Generated worlds with the same specific rules, history, and geography as copyrighted fictional universes (e.g., Middle-earth, Westeros, the Wizarding World).
- Quest/mission plagiarism — Game quests that reproduce the specific narrative structure of quests from other games.
What detection must cover:
- Plot structure comparison (sequence of events at multiple abstraction levels)
- Character arc comparison (character development trajectories)
- World-building element matching (magic systems, fictional histories, maps)
- Dialogue similarity (direct text comparison)
- Trope vs. specific expression differentiation
- Known universe detection (is this set in an existing fictional universe?)
- Quest/mission structure comparison (for game content)
8. Animation & Motion#
Generated by: Aja (motion capture pipeline), Calliope (artist animation), Maya (game animation), Psyche (avatar animation).
| Element | Protectability | Legal Basis |
|---|---|---|
| Choreography | Protectable if sufficiently original | 17 U.S.C. § 102(a)(4) |
| Motion capture data (raw) | Disputed — performance vs. data | No clear precedent |
| Character animation style | Weak — technique/style | Idea/expression |
| Specific dance moves | Individual moves not protectable | Hanagami v. Epic Games (2022) |
| Dance sequences | Protectable as choreographic work | Registration required |
| Mocap of real performer | Right of publicity | State laws |
Key risks:
- Choreography reproduction — Generated animations that reproduce copyrighted choreographic sequences. Fortnite dance lawsuits (largely unsuccessful for individual moves, but sequences may be protectable).
- Motion style replication — Replicating a specific performer's movement style closely enough to invoke right of publicity.
- Mocap data licensing — Using motion capture data from performers without proper licensing for the generated use case.
What detection must cover:
- Motion sequence similarity (joint angle trajectories, pose comparison)
- Choreography sequence matching (dance move sequences)
- Motion style fingerprinting (performer identification from movement)
- Performer consent verification (for mocap-derived content)
9. Fashion & Style Design#
Generated by: Galatea (fashion design), Aglaea (style/beauty), Calliope (artist fashion).
| Element | Protectability | Legal Basis |
|---|---|---|
| Fabric patterns/prints | Copyright on 2D artwork | Star Athletica v. Varsity (2017) |
| Garment shape/cut | Generally not copyrightable | Functional design |
| Design patent (ornamental) | Strong if patented | Design Patent Act |
| Trade dress (brand look) | Trademark if distinctive | Lanham Act |
| Color combinations | Generally not protectable | Functional |
| Jewelry design | Strong — sculptural | Copyright on 3D art |
Key risks:
- Pattern/print reproduction — Generated fabric designs that reproduce copyrighted prints or patterns.
- Trade dress infringement — Generated designs that are "confusingly similar" to established fashion brands (e.g., Birkin bag shape, red soles).
- Designer knockoffs — Generated designs clearly derived from specific designer collections.
What detection must cover:
- Fabric pattern/print similarity (image comparison)
- Trade dress matching (against known protected trade dress)
- Design patent comparison (ornamental design similarity)
- Brand identifier detection (logos, monograms, distinctive marks)
10. Virtual Worlds & Game Environments#
Generated by: Maya (metaverse), Hathor (game worlds), Yemaya (production environments).
| Element | Protectability | Legal Basis |
|---|---|---|
| Level layout/design | Moderate — creative expression | Limited precedent |
| Game mechanics | Not protectable (methods) | Lotus v. Borland (1995) |
| Environmental art | Strong if original | Visual copyright |
| Architectural design | Moderate — Arch. Works Act | Real buildings: § 120 limits |
| UI design/HUD | Weak — functional | Lotus |
| Procedural generation rules | Not protectable — method | Baker v. Selden |
| Soundscapes/ambience | Weak | Compilation at best |
Key risks:
- Level design copying — Procedurally generated levels that closely reproduce levels from copyrighted games.
- Landmark reproduction — Virtual worlds that reproduce real copyrighted architectural landmarks.
- Game environment cloning — Generating environments that are recognizably from specific games (e.g., a Minecraft-like blocky world, a Dark Souls-like interconnected level).
What detection must cover:
- Level layout comparison (topological similarity of navigable spaces)
- Environmental art similarity (texture, lighting, atmosphere comparison)
- Architectural work detection (real building reproduction)
- Game-specific aesthetic detection (recognizable game environments)
11. Educational Content#
Generated by: Metis (courses, lessons), Nisaba (scholarly analysis), Kuanyin (spiritual/mindfulness education).
| Element | Protectability | Legal Basis |
|---|---|---|
| Specific text/prose | Strong | Standard literary copyright |
| Facts/data/methods | Not protectable | Feist; Baker v. Selden |
| Test questions (specific) | Moderate — creative expression | ETS protects questions |
| Course structure/syllabus | Weak — method of organization | Functional |
| Educational diagrams | Moderate if original | Visual copyright |
| Academic papers | Strong | Literary copyright |
Key risks:
- Textbook reproduction — Generated lessons that reproduce passages from copyrighted textbooks.
- Test question copying — Generated assessments that reproduce specific copyrighted test questions.
- Academic plagiarism — Generated scholarly content that reproduces published research without attribution.
What detection must cover:
- Text similarity against educational databases
- Citation verification (claims properly attributed)
- Test question similarity (against known test banks)
- Diagram/figure similarity
12. Journalism & News Content#
Generated by: Veritas (articles, newsletters, social posts, video news).
| Element | Protectability | Legal Basis |
|---|---|---|
| Factual reporting | Facts not protectable | Feist v. Rural |
| Specific prose/expression | Strong | NYT v. OpenAI (2023, pending) |
| Headlines | Generally not protectable | Short phrases |
| Photography | Strong | Visual copyright |
| Investigative findings | Mixed — hot news doctrine | INS v. AP (1918) |
| Data compilations | Selection/arrangement protectable | Feist (limited) |
Key risks:
- Article reproduction — Generated articles that reproduce passages from copyrighted journalism. This is the core allegation in NYT v. OpenAI.
- Hot news misappropriation — Using another outlet's exclusive reporting before it's widely known.
- Photo reproduction — Generated articles including descriptions or recreations of copyrighted photographs.
- Source protection — Generated content that inadvertently reveals confidential sources.
What detection must cover:
- Text similarity against news databases (AP, Reuters, major outlets)
- Hot news detection (is this fact recently broken by another outlet?)
- Source/attribution verification
- Photo description comparison
13. Voice & Speech Synthesis#
Generated by: Nous/audio (TTS), Euterpe/voice (singing synthesis), Calliope/voice (AI artist voices), Psyche (avatar speech), Iris (assistant voice).
This is a rapidly evolving legal area with active legislation:
| Jurisdiction | Law | Status | Key Provision |
|---|---|---|---|
| US Federal | NO FAKES Act | Proposed | Federal right against unauthorized AI replicas |
| US Federal | ELVIS Act extension | Proposed | Extends right of publicity to AI voices |
| Tennessee | ELVIS Act | Enacted 2024 | First state to protect voice from AI replication |
| California | AB 2602, AB 1836 | Enacted 2024 | AI replica contracts and posthumous rights |
| EU | AI Act | Enacted 2024 | Transparency requirements for synthetic media |
| UK | No specific law yet | Pending | Common law passing off may apply |
Key risks:
- Voice cloning of real people — The most legally dangerous area. Creating a voice that sounds like a specific real person without consent.
- Vocal style replication — Even without exact cloning, reproducing the distinctive vocal characteristics (timbre, phrasing, technique) of a known artist may violate right of publicity.
- Posthumous voice use — Using AI to recreate deceased performers' voices without estate consent.
- Consent scope creep — Using a consented voice for purposes beyond the original consent scope.
What detection must cover:
- Already covered by Themis Music Shield vocal analysis engine for singing.
- Must extend to: speaking voice comparison, conversational voice comparison, consent scope verification for all voice synthesis contexts.
Part II: The Universal Originality Shield Architecture#
Design Philosophy#
The Themis Music Shield established the right architecture for music: multiple specialized detection engines → risk scoring → decision → human review → remediation. The Universal Shield extends this pattern to all media types while sharing infrastructure.
Key architectural principles:
-
Media-specific engines, shared infrastructure — Each media type has detection engines tuned to its unique characteristics (melody matching for music, visual similarity for images, text comparison for writing). But risk scoring, provenance, human review, and analytics are shared across all media.
-
Unified risk scoring — All media types feed into the same risk scoring framework with media-specific weights and thresholds calibrated by legal precedent for each media type.
-
Single audit trail — One provenance system tracks every generated work across all media types. Legal discovery queries one system, not seventeen.
-
Cross-media compound detection — A music video is music + visual + text (lyrics). A game is narrative + visual + audio + 3D. Compound detection across media types catches what single-media detection misses.
-
Progressive scanning — Fast pre-filters eliminate obviously clear content quickly. Expensive deep scans run only on content that survives pre-filters. Human review sees only the most ambiguous cases.
System Architecture#
┌─────────────────────────────────────────────────────────────────────────┐
│ CONTENT GENERATION PIPELINES │
│ Calliope Euterpe Isis Hathor Veritas Maya Yemaya Aja [all] │
└────────────────────────────┬────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────────────┐
│ THEMIS UNIVERSAL ORIGINALITY SHIELD │
│ │
│ ┌──────────────────────────────────────────────────────────────────┐ │
│ │ MEDIA ROUTER │ │
│ │ Detect media type → route to appropriate engine set │ │
│ └──────────────────────────────┬───────────────────────────────────┘ │
│ │ │
│ ┌──────────┬──────────┬────────┼────────┬──────────┬──────────┐ │
│ │ MUSIC │ VISUAL │ VIDEO │ TEXT │ 3D │ OTHER │ │
│ │ SHIELD │ SHIELD │ SHIELD │ SHIELD │ SHIELD │ SHIELDS │ │
│ │(Phase 74)│ │ │ │ │ │ │
│ ├──────────┤──────────┤────────┤─────────┤──────────┤──────────┤ │
│ │ Melodic │ Compos. │ Scene │ Text │ Mesh │ Motion │ │
│ │ Harmonic │ Charact. │ Dialog │ Semant. │ Texture │ Choreo. │ │
│ │ Rhythmic │ Face │ Face │ Plot │ Charact. │ Fashion │ │
│ │ Spectral │ Trademark│ Audio │ Source │ Brand │ Recipe │ │
│ │ Lyric │ Style │ Deepfk │ Charact.│ Archit. │ Educatn │ │
│ │ Vocal │ Watermark│ Tmmark │ Brand │ Patent │ Level │ │
│ │ Struct. │ Neural │ Neural │ Citation│ Neural │ World │ │
│ │ Fingerpr.│ Embed. │ Embed. │ Embed. │ Embed. │ Embed. │ │
│ └────┬─────┘────┬─────┘───┬────┘───┬─────┘────┬─────┘────┬─────┘ │
│ │ │ │ │ │ │ │
│ └──────────┴─────────┴────┬───┴──────────┴──────────┘ │
│ │ │
│ ┌──────────────────────────────▼───────────────────────────────────┐ │
│ │ CROSS-MEDIA COMPOUND DETECTOR │ │
│ │ Music video = audio shield + visual shield + lyric shield │ │
│ │ Game = narrative shield + visual shield + audio shield + 3D │ │
│ │ AI artist = vocal shield + visual shield + music shield + text │ │
│ └──────────────────────────────┬───────────────────────────────────┘ │
│ │ │
│ ┌──────────────────────────────▼───────────────────────────────────┐ │
│ │ UNIFIED SYSTEMS │ │
│ │ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │ │
│ │ │ Risk │ │ Provenance │ │ Human │ │ │
│ │ │ Scoring │ │ & Audit │ │ Review │ │ │
│ │ └──────┬──────┘ └─────────────┘ └──────┬──────┘ │ │
│ │ │ │ │ │
│ │ ┌──────▼──────┐ ┌─────────────┐ ┌──────▼──────┐ │ │
│ │ │ Decision │ │ Reference │ │ Remediation │ │ │
│ │ │ Engine │ │ Databases │ │ Pipeline │ │ │
│ │ └──────┬──────┘ └─────────────┘ └─────────────┘ │ │
│ └─────────┼────────────────────────────────────────────────────────┘ │
│ │ │
└────────────┼────────────────────────────────────────────────────────────┘
│
┌────────┼────────┐
▼ ▼ ▼
PASS FLAG REJECT
Part III: Media-Specific Detection Engines#
1. Music Shield (Already Specified — Phase 74)#
See docs/domains/themis-music-shield-proposal.md and Phase 74 in TODOS.md. 232
tasks covering 8 detection engines, reference database, risk scoring,
provenance, human review, remediation, model governance, and analytics.
2. Visual Originality Shield (@themis/visual-shield)#
Protects all AI-generated visual content — images, artwork, illustrations, design assets, visual branding.
2.1 Image Composition Similarity#
- Global descriptor matching — Extract holistic image descriptors (GIST descriptor, spatial pyramid matching) that capture the overall composition — subject placement, horizon line, perspective, negative space distribution. Compare against reference database of copyrighted visual works.
- Layout analysis — Parse image into compositional elements (foreground subject, background, framing) and compare structural layout. Two images with the same subject positioned identically, same perspective, same framing may be compositionally copied even with different content.
- Color histogram comparison — Extract color distributions and compare. Color alone is not protectable, but combined with composition it strengthens a claim.
- Spatial frequency analysis — Wavelet-based comparison of spatial frequency content. Captures texture patterns and visual detail distribution.
2.2 Neural Image Embedding Comparison#
- CLIP embedding similarity — Extract CLIP (Contrastive Language-Image
Pre-training) embeddings from generated images and reference database. CLIP
captures semantic visual similarity — two images of "the same thing" will have
high cosine similarity even with different artistic styles. Threshold:
0.85 cosine = FLAG, > 0.92 = REJECT.
- DINOv2 embedding similarity — Self-supervised visual embeddings that capture structural visual similarity without relying on text-image alignment. Better at detecting visual copies with different color palettes or artistic filters.
- SSCD (Self-Supervised Copy Detection) — Meta's specialized model for detecting visual copies, near-duplicates, and manipulated versions of copyrighted images. Purpose-built for this exact task.
- Perceptual hashing — pHash, dHash, aHash on generated images. Fast pre-filter for near-duplicate detection. Hamming distance < 10 = likely copy.
2.3 Character Recognition#
- Fictional character detection — Train or fine-tune a classifier on a database of copyrighted fictional characters (Disney, Marvel, DC, Nintendo, anime, etc.). Detect when a generated image contains a recognizable copyrighted character. Must handle stylistic variations — a "realistic Pikachu" is still Pikachu.
- Character feature matching — Extract character-defining features (color scheme, silhouette, distinctive accessories, body proportions) and compare against character database. A character with yellow skin, blue hair, and a specific body shape may be "Homer Simpson" even without exact reproduction.
- Character name-in-prompt detection — When the generation prompt contains a copyrighted character name, flag before generation even begins.
2.4 Face & Person Recognition#
- Face detection and identification — Detect faces in generated images using face detection models (RetinaFace, MTCNN). Extract face embeddings (ArcFace, AdaFace) and compare against a database of protected individuals — celebrities, public figures, opted-out persons.
- Likeness threshold — Cosine similarity > 0.70 on face embeddings triggers FLAG. > 0.85 triggers REJECT unless consent is verified.
- Consent verification — Integration with consent database. If a person has consented to AI depiction (e.g., for a licensed product), bypass the flag.
- Deepfake detection — Detect AI-generated faces that may be placing real people in fabricated scenarios. Combine face recognition with context analysis.
2.5 Trademark & Logo Detection#
- Logo detection — Object detection model fine-tuned on brand logos (Nike swoosh, Apple logo, McDonald's arches, etc.). Detect when generated images inadvertently include recognizable brand marks.
- Text detection (OCR) — Detect text in generated images. AI image generators sometimes produce text that resembles brand names or copyrighted text.
- Product identification — Detect recognizable commercial products in generated images (specific car models, consumer electronics, luxury goods) that could constitute trademark use.
2.6 Style & Training Data Analysis#
- Training data memorization detection — Compare generated images against known training dataset images (LAION, CC12M, etc.) using perceptual hashing and SSCD. Detect near-verbatim reproduction of training images.
- Watermark detection — Detect reproduced watermarks from training data (Getty Images, Shutterstock, Alamy watermarks). This is direct evidence of training data memorization.
- Style quantification — While style is not copyrightable, quantify the degree to which a generated image reproduces the specific stylistic elements of a particular artist (beyond mere inspiration) for advisory purposes. Use style transfer metrics and artist attribution models.
2.7 Reference Database#
- Copyrighted artwork database — Index visual descriptors and embeddings for commercially significant artwork — museum collections, major illustrators, stock photo libraries, album covers, movie posters, book covers, game art.
- Character database — Visual profiles of 10,000+ copyrighted fictional characters with multiple reference images per character.
- Face database — Face embeddings for public figures, celebrities, and opted-out individuals.
- Trademark database — Visual representations of 100,000+ registered trademarks and brand logos.
- Training data sample database — Representative sample of common AI training datasets for memorization detection.
3. Video Originality Shield (@themis/video-shield)#
Protects all AI-generated video content — music videos, short films, news clips, promotional videos, game cinematics.
3.1 Frame-Level Analysis#
- Per-frame visual analysis — Sample frames at 1fps (or higher for short clips) and run each through the Visual Shield's image analysis pipeline. Detect character reproduction, face recognition, trademark presence in any frame.
- Keyframe extraction — Identify visually distinct keyframes using scene change detection. Analyze keyframes more thoroughly than intermediate frames.
- Temporal aggregation — Aggregate per-frame results over time. A single frame containing a copyrighted character is a flag. Multiple consecutive frames are a stronger flag.
3.2 Scene-Level Analysis#
- Scene similarity — Extract scene-level features (shot composition, camera angle, lighting, color grading) and compare against a database of copyrighted scenes from films and television.
- Shot sequence matching — Compare sequences of shots (editing patterns) for structural similarity. A sequence of specific shot types in the same order may indicate scene reproduction.
- Montage analysis — Detect when an editing sequence reproduces the specific rhythm and structure of a copyrighted montage.
3.3 Dialogue & Audio Analysis#
- Speech-to-text extraction — Transcribe all spoken dialogue in generated video using ASR.
- Dialogue comparison — Run transcribed dialogue through the Text Shield's text similarity engine to detect reproduction of copyrighted screenplay dialogue.
- Audio track analysis — Extract and analyze the audio track through the Music Shield for any background music, sound effects, or vocal content.
3.4 Deepfake Detection#
- Synthetic media detection — Classify generated video as synthetic using deepfake detection models. All AI-generated video must be properly labeled per EU AI Act requirements.
- Face manipulation detection — Detect face-swapping, face reenactment, and other facial manipulations that place real people in fabricated video.
- Body manipulation detection — Detect body-swapping or full-body deepfakes.
- Provenance labeling — Ensure all generated video carries C2PA (Coalition for Content Provenance and Authenticity) metadata identifying it as AI-generated.
4. Text Originality Shield (@themis/text-shield)#
Protects all AI-generated text — articles, dialogue, narratives, lyrics, educational content, marketing copy, social posts.
4.1 Exact & Near-Exact Match Detection#
- Verbatim text matching — N-gram hash-based exact match detection against a reference corpus. Window sizes: 5-gram through 20-gram. Even a 50-word verbatim match against a copyrighted source is actionable.
- Near-duplicate detection — MinHash/SimHash for detecting paraphrased or lightly edited copies. Jaccard similarity on shingle sets. Threshold: > 0.60 Jaccard on 3-gram shingles = FLAG.
- Cross-language detection — Detect when generated text in one language closely translates a copyrighted text from another language. Use multilingual embeddings (mUSE, LaBSE) for language-invariant comparison.
4.2 Semantic Similarity Analysis#
- Passage-level embedding comparison — Encode passages (paragraph-length) as dense vectors using sentence transformers. Compare against reference database via ANN search. Two passages conveying the same specific idea in different words score high.
- Document-level embedding comparison — Encode entire documents and compare. Catches cases where overall structure and argument are copied even though individual sentences differ.
- Contextual similarity — Compare not just text but the context in which ideas are presented. Two texts presenting the same sequence of specific arguments in the same order is stronger evidence than scattered similar passages.
4.3 Narrative Structure Analysis#
- Plot structure extraction — Decompose narrative text into plot events using LLM-based analysis. Extract: inciting incident, rising action, climax, falling action, resolution. Compare event sequences.
- Character relationship graphs — Extract character names and their relationships (ally, enemy, romantic, familial). Compare character relationship graphs between generated and reference works. Isomorphic character graphs with same relationship types suggest plot copying.
- World-building element extraction — Extract named locations, rules/laws of the fictional world, historical events, faction names. Compare against known fictional universes.
- Quest/mission structure — For game narrative, extract quest objectives, branching points, reward structures. Compare against known game quests.
4.4 Source Attribution & Citation#
- Claim detection — Identify factual claims in generated text that should be attributed to sources.
- Source verification — For claims that appear to originate from specific sources, verify that proper attribution is included.
- Academic integrity — For educational content, verify that referenced studies/statistics exist and are correctly cited.
4.5 Brand & Trademark Detection#
- Brand name detection — NER (Named Entity Recognition) for brand names, product names, and trademarked terms in generated text.
- Tagline/slogan detection — Compare generated text against a database of copyrighted taglines and advertising slogans.
- Distinctive phrase detection — Detect use of phrases strongly associated with specific brands or works ("May the Force be with you", "Just do it", "Winter is coming").
4.6 Reference Database#
- Published text corpus — Index passages from commercially published books, articles, screenplays. Sources: Project Gutenberg (public domain), licensed excerpts from publishers, publicly available article abstracts.
- News article database — Rolling 90-day index of articles from major news outlets for hot news detection.
- Academic paper database — Abstracts and key passages from academic papers for educational content verification.
- Trademark/slogan database — 50,000+ trademarked phrases and slogans.
- Fictional universe database — Structured databases of major fictional universes (character names, locations, lore, magic systems) for world- building collision detection.
5. 3D Asset Originality Shield (@themis/3d-shield)#
Protects AI-generated 3D models, environments, and assets.
5.1 3D Shape Similarity#
- Point cloud comparison — Convert 3D models to point clouds (10K-100K points) and compare using Chamfer distance or Earth Mover's Distance. Rotation-invariant comparison using PCA alignment or spherical harmonics.
- Mesh topology comparison — Compare mesh structure using graph-based methods — genus, vertex degree distribution, face connectivity patterns. Topologically identical meshes with different vertex positions may be deformations of the same base mesh.
- Voxel occupancy comparison — Voxelize models at multiple resolutions (32³, 64³, 128³) and compare occupancy grids using IoU (Intersection over Union). Fast pre-filter for gross shape similarity.
- Neural 3D embeddings — Use pre-trained 3D shape encoders (PointNet++, DGCNN, Point-MAE) to extract fixed-length embeddings from 3D models. Compare via cosine similarity in vector database.
5.2 Texture & Material Comparison#
- UV-mapped texture comparison — Extract texture maps from 3D models and run through Visual Shield's image comparison for copyright analysis.
- Material parameter comparison — Compare PBR material parameters (albedo, roughness, metallic, normal maps) for distinctive material reproduction.
- Procedural pattern detection — Detect known copyrighted procedural textures or patterns applied to 3D models.
5.3 Character & Product Recognition#
- 3D character recognition — Compare generated 3D characters against a database of known copyrighted 3D character models. Multi-view rendering comparison (render from 8+ viewpoints, run each through Visual Shield).
- Product/brand detection — Detect when generated 3D models represent trademarked products (specific car models, consumer devices, branded items). Match against product 3D databases.
- Architectural work detection — Compare generated 3D architecture against database of copyrighted real-world buildings (those copyrighted under Architectural Works Copyright Protection Act).
5.4 Asset Library Deduplication#
- Training data memorization — Compare generated 3D models against known 3D asset libraries (Sketchfab, TurboSquid, Unity Asset Store) used in model training. Detect near-exact reproduction of commercial 3D assets.
- Attribution verification — When a generated 3D model closely matches a CC-licensed source, verify proper attribution.
6. Character Originality Shield (@themis/character-shield)#
Protects against unauthorized reproduction of copyrighted fictional characters and real person likenesses across all media types.
6.1 Visual Character Matching#
- Multi-view character comparison — Render/analyze character from multiple angles and compare against character database using visual embeddings. Handles 2D illustrations, 3D models, and video frames.
- Distinctive feature extraction — Extract character-defining features: color palette (skin, hair, outfit), body proportions, facial features, distinctive accessories (masks, weapons, logos). Compare feature vectors.
- Silhouette comparison — Compare character silhouettes. Many iconic characters are identifiable by silhouette alone (Batman, Mickey Mouse, Pikachu).
6.2 Character Profile Matching#
- Name collision detection — Check generated character names against databases of: trademarked character names, famous fictional character names, celebrity names. Flag exact matches and phonetic near-matches.
- Personality trait comparison — Extract personality dimensions (Big Five, alignment, archetype) and compare against known characters. A character with the exact same rare combination of traits may be a derivative.
- Backstory comparison — Extract key backstory elements (origin event, motivation, relationships, abilities) and compare using text similarity. The combination of specific backstory elements constitutes protectable expression even when individual elements are common tropes.
- Ability/power comparison — For characters with special abilities, compare
ability descriptions. Distinctive power sets (e.g., spider-related abilities
- web-slinging + spider-sense) are recognizable even with different character identity.
6.3 Cross-Media Character Detection#
- Unified character database — Maintain a cross-media character database linking visual representations, text descriptions, voice profiles, and animation styles for each copyrighted character.
- Multi-modal matching — When a character appears across media types (visual design + voice + dialogue + animation), compound matching across all modalities strengthens detection.
7. Motion & Animation Shield (@themis/motion-shield)#
Protects against choreography copyright infringement and unauthorized performer replication.
7.1 Motion Sequence Comparison#
- Joint trajectory comparison — Compare skeletal joint angle trajectories between generated and reference motion sequences using DTW. Normalize for body proportion differences.
- Pose sequence matching — Extract key poses and compare pose sequences using pose embedding similarity (PoseFormer, MotionBERT).
- Dance move vocabulary — Classify individual dance moves into a vocabulary (turns, kicks, waves, locks, pops, etc.) and compare move sequences. Individual common moves are not protectable; specific sequences are.
7.2 Choreography Detection#
- Choreographic work database — Index registered choreographic works and well-known dance sequences from commercial performances.
- Sequence matching — Compare generated dance sequences against database. A sequence of 8+ specific moves in the same order with similar timing constitutes a match.
- Performance style analysis — Detect when generated motion replicates the distinctive movement style of a specific performer (for right of publicity concerns).
7.3 Performer Consent#
- Mocap source verification — When generated motion is derived from motion capture, verify the mocap performer has consented to the generated use.
- Performer identification from motion — Use gait/movement biometrics to identify whether generated motion closely replicates a specific performer's signature movement patterns.
8. Fashion & Design Shield (@themis/fashion-shield)#
Protects against design patent infringement, trade dress violation, and pattern/print copyright infringement.
8.1 Pattern & Print Analysis#
- Textile pattern comparison — Compare generated fabric patterns/prints against databases of copyrighted textile designs using image similarity (CLIP, pHash, SSCD).
- Repeat pattern detection — Detect the repeating unit in a pattern and compare the unit against known copyrighted patterns.
- Color-way comparison — Compare color combinations used in generated designs against protected color-ways.
8.2 Design Silhouette & Trade Dress#
- Garment silhouette comparison — Compare overall garment shapes against known design patent silhouettes and protected trade dress.
- Distinctive design element detection — Detect design elements strongly associated with specific brands (red soles, interlocking G pattern, check pattern, monogram canvas).
- Design patent comparison — Compare generated designs against registered design patents using visual similarity.
9. Environment & Level Shield (@themis/environment-shield)#
Protects game environments, virtual worlds, and architectural reproductions.
9.1 Level Layout Comparison#
- Topological layout matching — Abstract level layouts to navigation graphs (rooms as nodes, connections as edges) and compare graph topology. Isomorphic level graphs with same connectivity suggest layout copying.
- Spatial flow comparison — Compare player movement patterns through levels — critical path length, branching factor, pacing (alternation of combat/ puzzle/exploration).
- Landmark placement matching — Extract positions of key landmarks within levels and compare spatial distributions.
9.2 Environmental Art Comparison#
- Atmosphere comparison — Compare lighting, fog, color grading, sky/weather to detect reproduction of distinctive environmental aesthetics from copyrighted games.
- Biome/theme matching — Detect when generated environments closely reproduce the specific aesthetics of copyrighted game environments (e.g., "this looks exactly like Firelink Shrine from Dark Souls").
9.3 Architectural Protection#
- Real building detection — Compare generated virtual architecture against a database of copyrighted real-world buildings.
- Fictional architecture detection — Compare against distinctive fictional architecture from copyrighted works (Hogwarts, Rivendell, etc.).
Part IV: Cross-Media Compound Detection#
The Compound Problem#
A music video is not just music. It is music + visual art + video + text (lyrics) + character (AI artist persona) + potentially choreography. Each media type's shield runs independently, but the compound result — multiple media types simultaneously matching the same copyrighted source — must be detected and escalated.
Multi-Modal Content Types in Oshun#
| Content Type | Component Media Types | Generating Domains |
|---|---|---|
| AI artist music video | Music + video + visual + lyrics + voice + dance | Calliope, Euterpe |
| Game level | 3D + narrative + audio + environment + characters | Maya, Hathor, Euterpe |
| Film scene | Video + dialogue + audio + character + visual | Yemaya, Isis, Hathor |
| News article with media | Text + images + video | Veritas |
| Educational lesson | Text + images + diagrams + audio | Metis |
| AI avatar interaction | Voice + text + animation + character | Psyche, Iris |
| Virtual world region | 3D + environment + audio + narrative + characters | Maya, Hathor |
| Fashion presentation | Visual + 3D + animation + music | Galatea, Calliope |
| Interactive experience | All of the above combined | Aphrodite, Maya |
Compound Detection Algorithm#
- Decompose — Split multi-modal content into component media types.
- Scan independently — Run each component through its media-specific shield.
- Correlate matches — For each matched reference work, check if matches occur across multiple media components targeting the same source. A music video where the audio matches Song X AND the visuals match Song X's music video is a compound match.
- Escalate non-linearly — Compound matches across N media types receive escalation multiplier: 2 types × 1.5, 3 types × 2.5, 4+ types × 4.0.
- Cross-reference — Check if the matched source across different media types belongs to the same IP holder. Music matching Universal's catalog + visuals matching Universal's music video = compound against the same entity.
Part V: Unified Infrastructure#
1. Universal Reference Database (@themis/universal-reference-db)#
A single, unified database infrastructure serving all media shields:
| Collection | Size Target | Sources |
|---|---|---|
| Music melodies | 20M+ works | MusicBrainz, AcoustID, MIDI databases |
| Audio fingerprints | 100M+ recordings | AcoustID, Content ID |
| Visual embeddings | 50M+ images | Museum collections, stock libraries, album art |
| Character profiles | 50,000+ characters | Manual curation from major franchises |
| Face embeddings | 1M+ individuals | Public figure databases, opt-out registry |
| Trademark/logo database | 100,000+ marks | USPTO, EUIPO, WIPO databases |
| Published text corpus | 10M+ passages | Books, articles, screenplays, lyrics |
| 3D model embeddings | 5M+ models | Sketchfab, TurboSquid, training datasets |
| Fictional universe DB | 500+ universes | Manual curation of major fictional IPs |
| Choreography database | 10,000+ sequences | Registered works, viral performances |
| Fashion design patents | 500,000+ designs | USPTO, EUIPO design registrations |
| News articles (rolling 90d) | 10M+ articles | Wire services, major outlets |
Shared vector database infrastructure — pgvector for structured data, Qdrant for high-dimensional embeddings. Shared ingestion, deduplication, enrichment, and maintenance pipelines.
2. Universal Risk Scoring (@themis/universal-risk-scoring)#
Extends the Music Shield's risk scoring to all media types:
- Media-type-specific weights — Each media type has its own element weight table calibrated by the legal protectability of each element in that medium.
- Cross-media compound scoring — When content spans multiple media types, compound detection across media types applies escalation multipliers.
- Unified risk levels — Same 5-level classification across all media: PASS → ADVISORY → FLAG → HIGH_RISK → REJECT.
- Per-jurisdiction assessment — Same jurisdiction-aware analysis (US, EU, UK, etc.) for all media types with media-type-specific legal standards.
3. Universal Provenance System (@themis/universal-provenance)#
Single audit trail for all generated content:
- Content-addressed storage — Every piece of generated content gets a unique hash. All provenance events reference this hash.
- Full generation chain — Prompt → model → parameters → intermediate outputs → final output, recorded for every content piece.
- Cross-media linkage — When a single creative project produces multiple media types (e.g., a music video produces audio + video + lyrics), the provenance system links all outputs to the same project.
- Legal discovery interface — Single query point for lawyers: "show me everything about this piece of content — how it was made, what it was scanned against, what was found, who approved it, and why."
4. Universal Human Review (@themis/universal-review)#
Shared human review infrastructure for all media types:
- Unified review queue — All flagged content enters one priority queue, regardless of media type.
- Specialized reviewer pools — Musicologists for audio, art historians for visual, IP attorneys for all high-risk, game designers for game content, journalists for news, etc.
- Cross-media review — For compound content (music videos, games), the review interface shows all media components side by side with their respective matches.
- Unified SLA — Same 4-hour SLA for all flagged content.
5. Universal Remediation (@themis/universal-remediation)#
Media-specific remediation with shared orchestration:
- Music remediation — Melody variation, reharmonization, lyric reformulation, vocal modification (from Music Shield).
- Visual remediation — Re-generation with negative constraints, composition modification, color palette shift, character redesign.
- Text remediation — Rewriting flagged passages while preserving meaning, character renaming, plot divergence.
- 3D remediation — Mesh deformation, texture regeneration, proportion adjustment.
- Video remediation — Re-render flagged scenes, adjust shot composition, modify dialogue.
- Character remediation — Modify visual appearance, alter name, adjust backstory elements.
6. Universal Analytics (@themis/universal-analytics)#
Shared monitoring and continuous improvement:
- Cross-media dashboard — Scan volume, risk distribution, latency, and FP/FN rates across all media types in one view.
- Media-type comparison — Which media types have the highest flag rates? Which need threshold adjustment?
- Model risk tracking — Which generation models across which domains produce the most flagged content?
- Legal landscape monitoring — Track new cases and legislation across all media types.
Part VI: Domain Integration Matrix#
Every generating domain must integrate with the Universal Shield. This matrix shows which shields gate which domains:
| Domain | Music | Visual | Video | Text | 3D | Character | Motion | Fashion | Environment |
|---|---|---|---|---|---|---|---|---|---|
| Calliope | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ||
| Euterpe | ✓ | ✓ | |||||||
| Isis | ✓ | ✓ | ✓ | ✓ | ✓ | ||||
| Hathor | ✓ | ✓ | ✓ | ||||||
| Veritas | ✓ | ✓ | ✓ | ||||||
| Maya | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ||
| Yemaya | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | |
| Aja | ✓ | ||||||||
| Galatea | ✓ | ✓ | |||||||
| Aglaea | ✓ | ✓ | |||||||
| Psyche | ✓ | ✓ | ✓ | ||||||
| Iris | ✓ | ||||||||
| Nous | ✓ | ||||||||
| Metis | ✓ | ✓ | |||||||
| Nisaba | ✓ | ||||||||
| Aphrodite | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | |
| Bellona | ✓ | ||||||||
| Hestia | ✓ | ✓ |
Part VII: Watermarking & Content Authentication#
Beyond detecting whether generated content infringes existing works, we must also protect Oshun's own generated content and provide transparency about AI generation.
1. AI Generation Disclosure#
- C2PA metadata — All AI-generated content carries Coalition for Content Provenance and Authenticity metadata identifying it as AI-generated. Required by EU AI Act for synthetic media.
- Visible watermarking — Optional visible "AI-Generated" watermarks on images and video. Configurable per domain and use case.
- Audio watermarking — Inaudible watermarks embedded in all generated audio identifying the generation source and timestamp.
- Text metadata — AI-generated text tagged with generation provenance in document metadata.
2. Forensic Watermarking#
Extends existing @aja/content-watermarking and
@aphrodite/stream-watermarking:
- Multi-layer approach — Spatial, temporal, and frequency-domain watermarks for robustness against different attack vectors.
- Per-output unique watermarks — Each generated piece carries a unique watermark enabling source tracing if content leaks.
- Tamper detection — Watermarks detect if content has been modified after generation.
3. Leak Detection & Enforcement#
Extends existing @aja/content-watermarking leak detection:
- Active scanning — Periodically scan public platforms for Oshun-watermarked content appearing without authorization.
- DMCA automation — When unauthorized use is detected, auto-generate DMCA takedown notices with all necessary provenance evidence.
- Chain of custody — Complete forensic trail from generation to detected unauthorized use.
Part VIII: Model Governance (Cross-Media)#
Extends the Music Shield's model governance to all generation models across all domains.
1. Training Data Audit (All Media)#
| Media Type | Training Data Risk | Key Audit Requirements |
|---|---|---|
| Music | Copyrighted songs | Licensed datasets, opt-out compliance |
| Images | Copyrighted photos/art | LAION audit, Getty/Shutterstock exclusion |
| Video | Copyrighted film/TV | Clip licensing, scene exclusion |
| Text | Copyrighted books/articles | Fair use analysis, opt-out compliance |
| 3D | Commercial asset packs | Asset library licensing |
| Motion | Performer mocap | Performer consent verification |
2. Memorization Testing (All Media)#
- Image memorization — Feed prompts designed to reproduce known training images. Measure SSCD similarity to training data.
- Text memorization — Feed prompts designed to reproduce known training text passages. Measure n-gram overlap.
- Video memorization — Feed prompts designed to reproduce known training video scenes. Measure per-frame similarity.
- 3D memorization — Feed prompts designed to reproduce known training 3D models. Measure shape similarity.
3. Red-Teaming (All Media)#
Systematic adversarial testing of every generation model before deployment:
- Can it reproduce copyrighted characters?
- Can it reproduce real people's likenesses?
- Can it reproduce copyrighted text verbatim?
- Can it reproduce copyrighted scenes from films?
- Can it reproduce copyrighted game levels?
- Can it reproduce trademarked logos or products?
Part IX: Regulatory Compliance Matrix#
| Regulation | Jurisdiction | Effective | Requirements | Oshun Impact |
|---|---|---|---|---|
| EU AI Act | EU | 2024-2026 | Transparency for synthetic media, risk assessment | All generated content must be labeled |
| NO FAKES Act | US (proposed) | TBD | Federal right against unauthorized digital replicas | Voice and likeness shields mandatory |
| ELVIS Act | Tennessee | 2024 | Voice rights protection | Voice shield mandatory for TN distribution |
| AB 2602 / AB 1836 | California | 2024 | AI replica contracts, posthumous rights | Consent verification mandatory |
| DMCA | US | 1998 | Safe harbor, takedown procedures | Takedown automation, repeat infringer policy |
| Copyright Directive | EU | 2019 | TDM exceptions with opt-out | Opt-out registry for training data |
| GDPR | EU | 2018 | Data protection for biometric data | Face embeddings are biometric — special handling |
| UK Online Safety | UK | 2023 | Deepfake distribution | Deepfake detection mandatory |
| Digital Services Act | EU | 2024 | Transparency for recommender systems | Content provenance |
Part X: Library Architecture (Full)#
libs/themis/
├── music-shield/ # Music Shield (Phase 74 — 17 libraries)
│ └── (see Phase 74 TODOS.md for full breakdown)
│
├── visual-shield/ # Visual Originality Shield
│ ├── composition-analysis/ # Image composition similarity
│ ├── neural-embeddings/ # CLIP/DINOv2/SSCD embeddings
│ ├── character-detection/ # Fictional character recognition
│ ├── face-recognition/ # Real person likeness detection
│ ├── trademark-detection/ # Brand/logo detection in images
│ ├── memorization-detect/ # Training data reproduction
│ └── watermark-detection/ # Reproduced watermarks from training
│
├── video-shield/ # Video Originality Shield
│ ├── frame-analysis/ # Per-frame visual analysis
│ ├── scene-comparison/ # Scene-level similarity
│ ├── dialogue-extraction/ # Speech-to-text + comparison
│ ├── deepfake-detection/ # Synthetic media detection
│ └── c2pa-labeling/ # Content authenticity metadata
│
├── text-shield/ # Text Originality Shield
│ ├── exact-matching/ # N-gram hash-based matching
│ ├── semantic-similarity/ # Embedding-based comparison
│ ├── narrative-analysis/ # Plot/character/world comparison
│ ├── source-attribution/ # Citation verification
│ ├── brand-detection/ # Trademark/slogan detection
│ └── cross-language/ # Multi-language detection
│
├── 3d-shield/ # 3D Asset Originality Shield
│ ├── shape-comparison/ # Point cloud/mesh/voxel matching
│ ├── texture-comparison/ # UV map analysis
│ ├── product-detection/ # Brand/product recognition in 3D
│ └── asset-deduplication/ # Training data memorization
│
├── character-shield/ # Character Originality Shield
│ ├── visual-matching/ # Character visual comparison
│ ├── profile-matching/ # Name/trait/backstory comparison
│ └── cross-media/ # Multi-modal character detection
│
├── motion-shield/ # Motion & Animation Shield
│ ├── sequence-comparison/ # Joint trajectory matching
│ ├── choreography-detect/ # Dance sequence detection
│ └── performer-consent/ # Mocap performer verification
│
├── fashion-shield/ # Fashion & Design Shield
│ ├── pattern-analysis/ # Textile pattern comparison
│ ├── trade-dress/ # Brand design detection
│ └── design-patent/ # Design patent comparison
│
├── environment-shield/ # Environment & Level Shield
│ ├── layout-comparison/ # Level topology matching
│ ├── atmosphere-analysis/ # Environmental art comparison
│ └── architecture-detect/ # Real building detection
│
├── compound-detector/ # Cross-Media Compound Detection
│ ├── media-decomposition/ # Split multi-modal content
│ ├── cross-correlation/ # Match correlation across media
│ └── escalation/ # Non-linear compound scoring
│
├── universal-reference-db/ # Unified Reference Database
│ ├── ingestion/ # Multi-media ingestion pipelines
│ ├── maintenance/ # Dedup, enrichment, jurisdiction
│ └── query/ # Unified query interface
│
├── universal-risk-scoring/ # Unified Risk Scoring
│ ├── media-weights/ # Per-media-type element weights
│ ├── compound-scoring/ # Cross-media compound escalation
│ └── jurisdiction/ # Multi-jurisdiction assessment
│
├── universal-provenance/ # Unified Provenance System
│ ├── generation-chain/ # Full generation provenance
│ ├── cross-media-linkage/ # Project-level linkage
│ └── legal-discovery/ # Discovery query interface
│
├── universal-review/ # Unified Human Review
│ ├── queue/ # Priority queue for all media
│ ├── reviewer-pools/ # Specialized reviewer assignment
│ └── cross-media-review/ # Multi-modal review interface
│
├── universal-remediation/ # Unified Remediation
│ ├── orchestrator/ # Media-type routing
│ └── media-specific/ # Per-media remediation adapters
│
├── universal-analytics/ # Unified Analytics
│ ├── cross-media-dashboard/ # Aggregated metrics
│ ├── threshold-tuning/ # Continuous calibration
│ └── legal-monitor/ # Case law and legislation tracking
│
├── watermarking/ # Content Authentication
│ ├── c2pa/ # Content authenticity metadata
│ ├── forensic/ # Multi-layer forensic watermarks
│ ├── audio-watermark/ # Inaudible audio watermarks
│ └── leak-detection/ # Active platform scanning
│
└── model-governance/ # Cross-Media Model Governance
├── training-audit/ # Per-media training data audit
├── memorization-test/ # Per-media memorization testing
├── red-team/ # Systematic adversarial testing
└── opt-out-registry/ # Cross-media opt-out management
Total new libraries: ~60 (in addition to the 17 from the Music Shield).
Part XI: Development Phasing#
Phase 74 (Already Defined): Music Shield Foundation — 232 tasks#
The most legally critical media type. Already specified in TODOS.md.
Phase 75 (Proposed): Visual + Text Shields — ~250 tasks#
The next two most voluminous and legally risky media types. Images are generated at the highest volume. Text is generated across the most domains.
Phase 76 (Proposed): Video + Character + 3D Shields — ~200 tasks#
Complex media types that build on the visual and text foundations.
Phase 77 (Proposed): Motion + Fashion + Environment Shields — ~120 tasks#
Specialized shields for lower-volume but still legally relevant media types.
Phase 78 (Proposed): Cross-Media, Watermarking, Universal Systems — ~180 tasks#
The unifying infrastructure: compound detection, universal reference database, universal provenance, universal review/remediation/analytics, watermarking and content authentication, cross-media model governance.
Total estimated tasks across all phases: ~980
Part XII: Success Metrics (Universal)#
| Metric | Target | Scope |
|---|---|---|
| False negative rate (all media) | < 0.01% | Third-party claims that shield missed |
| False positive rate (all media) | < 5% | Human overrides of FLAG/REJECT |
| Music scan latency (3-min track) | < 30s | P95 |
| Image scan latency | < 5s | P95 |
| Text scan latency (5000 words) | < 10s | P95 |
| Video scan latency (3-min clip) | < 60s | P95 |
| 3D model scan latency | < 15s | P95 |
| Reference database coverage | > 100M works | Across all media types |
| Review SLA (all flagged content) | < 4 hours (90%) | Time from flag to decision |
| Remediation success rate | > 90% | Auto-remediation without human |
| Uptime | 99.9% | All shield APIs |
| Regulatory compliance | 100% | EU AI Act, NO FAKES Act, DMCA |
Conclusion#
The Oshun ecosystem generates creative content at a scale and across a breadth of media types that is unprecedented. No existing platform operates across this many creative domains simultaneously with AI generation. This means no existing IP protection system is sufficient — we must build the most comprehensive cross-media originality verification system ever constructed.
The Themis Universal Originality Shield is not a competitive advantage. It is an existential necessity. A single high-profile infringement case across any media type can trigger regulatory scrutiny, platform bans, and reputational damage that affects all 31 generating domains. The shield must be as comprehensive as the platform itself.
Every pixel, every note, every word, every polygon, every frame, every character, every motion — must be defensible.
References#
Foundational IP Law#
- Feist Publications v. Rural Telephone, 499 U.S. 340 (1991) — facts not copyrightable
- Baker v. Selden, 101 U.S. 99 (1880) — methods not copyrightable
- Nichols v. Universal Pictures, 45 F.2d 119 (2d Cir. 1930) — abstraction test for characters/plots
- Computer Associates v. Altai, 982 F.2d 693 (2d Cir. 1992) — abstraction-filtration-comparison test
Visual Art#
- Mannion v. Coors Brewing, 377 F. Supp. 2d 444 (S.D.N.Y. 2006) — photo composition protectability
- Star Athletica v. Varsity Brands, 580 U.S. 405 (2017) — separability test for fashion designs
- Andersen v. Stability AI (N.D. Cal., filed 2023) — artist style and AI training
- Getty Images v. Stability AI (D. Del., filed 2023) — training data and watermark reproduction
Characters & Personas#
- DC Comics v. Towle, 802 F.3d 1012 (9th Cir. 2013) — character copyright test
- Midler v. Ford Motor Co., 849 F.2d 460 (9th Cir. 1988) — voice right of publicity
- Thaler v. Perlmutter, No. 1:22-cv-01564 (D.D.C. 2023) — AI-generated works not copyrightable
Text & Narrative#
- Sheldon v. Metro-Goldwyn Pictures, 81 F.2d 49 (2d Cir. 1936) — plot copying
- Warner Bros. v. RDR Books, 575 F. Supp. 2d 513 (S.D.N.Y. 2008) — fictional world protection
- NYT v. OpenAI (S.D.N.Y., filed 2023) — AI text reproduction
Game & Software#
- Tetris Holding v. Xio Interactive, 863 F. Supp. 2d 394 (D.N.J. 2012) — game visual expression
- Lotus Dev. v. Borland Int'l, 49 F.3d 807 (1st Cir. 1995) — UI methods not copyrightable
Dance & Choreography#
- Hanagami v. Epic Games, 85 F.4th 1100 (9th Cir. 2023) — choreography in games
Legislation#
- EU AI Act (Regulation 2024/1689)
- NO FAKES Act (S. 3696, 118th Congress)
- Tennessee ELVIS Act (Tenn. Code § 47-25-1101)
- California AB 2602, AB 1836 (2024)
- DMCA, 17 U.S.C. § 512
- EU Copyright Directive 2019/790
- GDPR (Regulation 2016/679) — biometric data provisions