Exceptions (e.g., Fair Use): In the US, “Fair Use” might allow limited copying for purposes like criticism, commentary, teaching, or research. Its application to AI training is highly debated and uncertain.
“Style” or “Sound”: Copyright generally protects the specific expression (melody, lyrics, recorded sounds), not the underlying idea, genre, style, or a general “sound” or voice.
Is it fair for AI companies to train models on vast amounts of copyrighted music without asking for permission or paying creators? Why or why not?
Should an AI be allowed to generate music that perfectly mimics a specific artist’s unique “sound” or vocal style, even if it doesn’t copy their exact songs?
Part I: How Generative AI Works (Simplified)
Generative AI (GenAI): AI that creates new content (text, images, music).
Examples: ChatGPT (text), DALL-E (image), various music AI tools.
Models:
LLMs (Large Language Models): Trained on text (e.g., GPT).
Diffusion Models: Often used for images/video.
Training: Requires massive computing power and huge datasets. Often dominated by large tech companies.
Part I: The Training Process - Copying & Tokenizing
Copying: Data (music files, text, images) is typically copied locally for efficient training.
Tokenizing: The copied data is broken down into smaller pieces called “tokens”.
For text: tokens might be words, parts of words, or characters.
For music: tokens might represent notes, chords, timings, or audio features.
Relationships: The AI learns patterns and relationships between these tokens.
Part I: Training Misconception
Myth: Training destroys the original work, leaving only abstract learning.
Reality: The process preserves representations (embeddings/vectors) of the original data, including relationships between elements.
This allows the AI to reproduce parts of the training data or generate statistically similar content.
The tokenized dataset itself can be seen as another reproduction.
Part I: How AI Generates Output
Prediction Machine: Based on a prompt (user request), the AI uses its training (the learned relationships between tokens) to predict the most likely next token (word, note, pixel).
It does this repeatedly to generate a sequence (sentence, melody, image).
Part I: Legal Implications of Tech
Copying (Reproduction):
Initial copying for training.
The tokenized dataset itself.
Outputs that are substantially similar to training data.
Rights Management Information (RMI): Metadata (author, title, etc.) is often stripped during training, potentially violating laws like the DMCA in the US.
AI Company Indemnification: Offers from companies (like Google, Microsoft, OpenAI) to protect users from copyright lawsuits often have significant limitations and exclusions.
Part I: Key Takeaways Summary
Training LLMs involves copying data.
Copying copyrighted data infringes rights unless an exception (like fair use) applies (highly debated).
Tokenization also creates a reproduction, preserving aspects of the original works.
AI outputs can infringe if substantially similar or derivative of training data.
Training may involve illegal removal of RMI (metadata).
AI company indemnification offers have significant limits.
Any legal exception for TDM must likely pass the international “three-step test” (balancing creator rights and public interest).
Discussion Question 2
Does understanding how LLMs work (copying, tokenizing, preserving relationships, predicting) change your view on whether using copyrighted music for training is fair or constitutes infringement?
How is AI “learning” different from human learning in a way that matters for copyright?
Part II: Copyright Has Always Adapted
Copyright law isn’t static; it evolves with technology:
Player Pianos -> Mechanical reproduction rights
Radio/TV -> Broadcasting rights
Cinema -> Film recognized as copyrighted work
Internet -> Digital transmission / “Making available” rights (WCT/WPPT 1996)
New rights (exclusive or remuneration) were often created to ensure creators were compensated for new major uses.
Part II: Why Adapt Now for AI?
AI is a profound technological change.
It uses creators’ past work to generate competing content.
This threatens creators’ ability to earn a living, potentially shrinking future creativity.
Delegating human interpretation/creation to machines has deep cultural implications.
Goal: Not to stop AI, but to ensure human creators can coexist and thrive alongside it, using the established system of copyright.
Part II: The Proposal - A Right to Remuneration
Create a NEW right specifically for creators.
What it covers: Payment (remuneration) for the use of copyrighted material within an AI model when that model is used to generate competing content.
Distinction: This focuses on the output phase’s reliance on the training data, not just the initial input copying (which is covered by existing reproduction rights).
Part II: Advantages of This Proposed Right
AI training can continue largely unhindered.
AI companies pay for a key, valuable input (the creative data).
Exempts non-commercial research uses (e.g., universities).
Compensates creators when AI uses their life’s work to compete with them.
Part II: Legal Details - Subject Matter & Ownership
Applies to: Copyright-protected musical works (likely compositions first).
Initial Owner: The human creator(s).
Transferable: Creators could assign or license this right (e.g., to publishers, CMOs).
Normative Basis: Empowers creators, giving them agency when their work fuels competing AI output.
Part II: Legal Details - National Treatment
If “Sui Generis” (Unique Right): Countries could choose reciprocity (only pay creators from countries with a similar right). This incentivizes other countries to adopt it.
If Part of Copyright: Subject to national treatment (must treat foreign creators the same as domestic ones, under treaties like Berne/TRIPS).
Part II: Legal Details - Compulsory License? Levy?
Remuneration Right: Can be structured similar to existing compulsory/statutory licenses (use allowed if payment made), which is generally permissible under international law.
Better than a Levy: A simple levy (like on blank tapes/hard drives) based on input data size doesn’t reflect actual use or market impact of the AI output. The proposed right connects payment more directly to the AI’s productive use.
Part II: Practicalities - The Distribution Challenge
Common Argument Against: How can we possibly track which works were used by the AI to generate a specific output and distribute the money fairly? It’s too complicated!
Part II: Practicalities - Potential Solutions
Transparency is Key: AI platforms can be designed to identify source material or influences (though they may resist). Requires legal obligation.
EU AI Act mandates some training data disclosure.
CMOs: Collective Management Organizations already distribute royalties based on complex usage data/proxies. They have experience.
Usage Data: Ideally, track how often tokenized works are “pulled” by the AI during generation.
Proxies: If direct tracking is impossible, use proxies (like commercial success of source works, genre representation in dataset).
Focus: Ensure money reaches individual creators.
Discussion Question 3
How could we realistically track AI’s use of specific songs or musical elements to pay creators fairly?
What are the pros and cons of relying on Collective Management Organizations (CMOs) versus requiring AI companies themselves to provide detailed usage data for distribution?