Navigating Ownership with AI Content

AI-generated content raises complex copyright issues. Training data, output ownership, and infringement are highly debated with ongoing lawsuits and evolving regulations. This lesson covers what we know, what is clarified, what remains uncertain, and how to navigate responsibly.

The Current Legal Landscape:

Key Fact: Human Authorship Required

The U.S. Copyright Office has confirmed (2023–2025):

• No new law is required—existing copyright principles apply to AI.
• Copyright protects only works of human creation.
• Fully AI-generated works (with no human creative input) cannot be copyrighted.
• Case-by-case evaluation continues, especially for mixed human + AI works.

Three Separate Copyright Questions:

1. Training data: Is it legal to train AI on copyrighted works?
2. Output ownership: Who owns or can copyright AI-generated content?
3. Infringement: Can AI outputs infringe existing copyrights?

Question 1: Training Data and Fair Use:

How AI Models Are Trained:

• Trained on billions of works: books, articles, code, images, music
• Most training data is copyrighted
• Used without explicit permission from creators
• AI companies argue this is 'fair use'
• Creators argue it's copyright infringement

The Fair Use Argument (AI Companies):

Four fair use factors under U.S. law:

1. Purpose and character: Transformative use (learning patterns, not reproducing)
2. Nature of work: Factual works have less protection
3. Amount used: Entire works used, but in transformed way
4. Market effect: Doesn't substitute for original (debatable)

AI companies claim training is similar to a human reading books to learn—transformative, not reproductive.

The Copyright Infringement Argument (Creators):

• Copying entire works without permission or compensation
• Commercial use (AI companies profit from trained models)
• Market harm (AI can create competing works)
• Not transformative if AI can reproduce training data
• Different from human learning (copies stored digitally)

Current Lawsuits:

• Authors vs. OpenAI/Meta: Claim unauthorized use of books for training
• Getty Images vs. Stability AI: Claim unauthorized use of millions of photos
• Artists vs. Midjourney/Stability AI: Claim style mimicry violates rights
• Programmers vs. GitHub/Microsoft/OpenAI: Claim Copilot violates code licenses

Question 2: Ownership of AI Outputs:

Critical Update (January 2025 – U.S. Copyright Office):

• Works created entirely by AI (with no human creative input) are not eligible for copyright.
• Prompts alone do not create authorship.
• Iterative prompting (even 100+ tries) does not change this.
• Selecting the "best" AI output is not creative authorship.
• Copyright applies only where a human contributes perceptible expressive elements or significantly edits/arranges outputs.

Three pathways for protection:

1. AI Assistive Tool Use: Correcting spelling, adjusting colors, removing background → fully copyrightable (AI not making creative choices).
2. Expressive Inputs: Supplying original human artwork or text as input, with those elements clearly visible in the output (e.g. Rose Enigma case).
3. Creative Editing/Arranging: Modifying AI outputs significantly, or arranging them with text or other works (e.g. Zarya of the Dawn comic).

Case Study Examples:

• Thaler v. Perlmutter (2023): Registration denied—no human authorship.
• Rose Enigma: Human-drawn mask and roses registered, but AI-added lighting effects not protected.
• Zarya of the Dawn: Copyright covered text and arrangement, not individual Midjourney images.

Question 3: Can AI Infringe Copyright?

The Concern:

AI can generate outputs similar to copyrighted works:

• Image generators can produce images in the style of specific artists
• Code assistants can suggest code similar to copyrighted code
• Text generators can reproduce passages from training data

When Is It Infringement?

Traditional copyright infringement requires:

1. Copying: Defendant had access to and copied from protected work
2. Substantial similarity: Works are substantially similar
3. Protectable expression: Similarity is in protectable expression, not ideas

With AI:

• Access: AI had access via training data
• Similarity: Can be high—especially if outputs closely mirror existing works
• Important: Copyright law does not protect style, only expression. Style mimicry ≠ infringement unless specific expression is copied.

Practical Guidance for Users:

Understand Terms of Service: Each platform defines output ownership, data usage, and training terms.

Add Human Creativity: Don’t publish raw AI output—edit, refine, combine with your own expressive work.

Avoid Copying Specific Works: Don’t prompt AI to replicate known copyrighted texts or images.

Document Your Process: Save drafts, edits, and inputs showing your creative contribution.

Consider Disclosure: Disclose AI use when appropriate for transparency and trust.

Attribution & Citation:

Consider disclosing AI tools used (e.g., “Draft created with assistance from ChatGPT”) when AI significantly contributed. Disclose when required by client, publisher, or legal context.

Future Legal Developments:

Expect evolving legislation around AI training data, licensing mechanisms, transparency, and liability frameworks. But the principle remains firm: pure AI output has no copyright protection—your human creative input is the key.