What copyright law actually says about AI-generated content.

Not if it is purely AI-generated. US and EU law both require human authorship for copyright protection.

Thaler v. Perlmutter (March 2025): the DC Circuit Court ruled machines cannot be authors under US copyright law. Zarya of the Dawn (2023): the Copyright Office protected human-written text and creative arrangement but denied protection to raw Midjourney images.

What gets copyright protection

• Human-authored text and strategic direction
• Creative selection and arrangement of AI elements
• Substantial human modifications to AI outputs
• Works where humans control the creative expression

What does not get protection

• Raw AI outputs from prompts
• Minor adjustments to AI content
• Iterative prompt refinement alone
• Content where AI determined the creative choices

For copyright protection, the human creative contribution must be substantial and documented.

Ownership depends on contracts and the level of human contribution. Most AI platforms grant users contractual rights to outputs, but copyright protection depends on whether the human contribution meets the legal threshold for authorship.

Employment context

• Content created by employees: the employer owns it under the work-for-hire doctrine
• Content created by contractors: ownership is determined by the contract
• Content created by agencies: requires an explicit transfer agreement

Always verify

• Platform commercial use terms
• Whether the platform trains on your data
• Indemnification coverage, if any
• Geographic restrictions on usage

The user bears the liability, not the AI platform. Platform indemnification, where it is offered, provides only limited protection.

Platform indemnification examples

• OpenAI Enterprise: Copyright Shield covers claims up to 12 months of fees paid, on Enterprise and API only, not the free tier
• Google Gemini paid plans: covers both training data and generated outputs; the free tier is excluded
• Adobe Firefly Enterprise: indemnification capped at $10,000 per output or claim
• Midjourney and Stability AI: no indemnification offered

What indemnification typically does not cover

• Trademark claims, which are explicitly excluded
• Content you knew or should have known was infringing
• Outputs where you disabled safety filters or citations
• Use after receiving an infringement notice from a rights holder

The $1.5 billion Anthropic settlement (September 2025) involved authors proving Anthropic downloaded pirated books for training data.

It depends on jurisdiction and content type.

United States, currently

• No general legal requirement for marketing content
• The FTC requires disclosure if AI creates deceptive endorsements, testimonials, or product claims
• State laws are emerging; California and New York are considering disclosure mandates

European Union, from August 2026

• The EU AI Act mandates labeling for AI-generated images, audio, and video
• Required format: "clearly labelled as AI generated" by machine-readable means
• Text content is currently exempt
• Penalties for non-compliance reach €35M or 7% of global annual revenue

Platform-specific requirements

• TikTok, Instagram, and YouTube are adding AI content labels
• Some platforms require creator disclosure when AI generation tools are used

Transparency requirements are tightening globally. Plan for disclosure capabilities even where they are not yet required.

Free tools provide no legal protection. Paid enterprise platforms offer indemnification, privacy, and clear commercial rights.

Free tools versus enterprise platforms

• Copyright indemnification: none on free tools, available on enterprise
• Data privacy: free tools may train on your data, enterprise does not
• Commercial use rights: often restricted on free tools, unrestricted on enterprise
• Usage tracking: prompts may be stored on free tools, private on enterprise

The Midjourney free tier restricts commercial use entirely under a Creative Commons Noncommercial 4.0 licence. Using free-tier outputs commercially violates the licence terms. Enterprise platform pricing typically ranges from $30 to $60 per user per month.

Yes, if the content is purely AI-generated without substantial human creative contribution. Pure AI outputs lack copyright protection under current US and EU law, and content in the public domain can be copied freely.

What creates legal protection

• Substantial human creative direction
• Selection and arrangement decisions
• Editing and modification choices
• Strategic composition

The Zarya of the Dawn case established this: individual AI images received no copyright protection, but the human-authored arrangement and text did. Proving substantial human contribution requires documentation; without evidence of the creative process, defending against copying becomes difficult.

Trademark infringement is serious exposure, and platform indemnification explicitly excludes trademark claims. Copyright protects creative expression; trademark protects brand identifiers and prevents consumer confusion.

Common AI trademark risks

• Visual similarity to protected logos, packaging, or designs
• Trademarked colour combinations such as UPS brown, Tiffany blue, or T-Mobile magenta
• Trade dress: the overall look and feel of products or packaging

Indemnification does not cover trademarks because the infringement occurs through your commercial use creating consumer confusion, not through AI generation itself. That is your liability. GEMA v. OpenAI (Munich, November 2025) found ChatGPT infringed copyright by reproducing song lyrics, and OpenAI must pay licensing fees.

Risk management

• Avoid prompts referencing competitor brands by name
• Screen outputs for brand resemblance before commercial use
• Reverse image search all visual content
• Commission legal review for major campaigns

Style and aesthetic are not copyrightable. Specific creative expressions are.

Legal: acceptable prompts

• "Luxury automotive photography aesthetic"
• "Minimalist tech product design language"
• "Premium spirits brand sophistication"
• "High-energy sports marketing vibe"

Illegal: infringement risk

• "Apple iPhone campaign clone"
• "BMW 2024 ad series recreation"
• "Patagonia campaign exact match"
• Prompts naming specific artists, photographers, or brands

Thomson Reuters v. Ross Intelligence (February 2025) rejected the fair use defence for AI training on a competitor’s content, finding it commercial, non-transformative use that created a competing product and caused market harm. You can capture a quality level and a feeling; you cannot copy specific protected elements. "Inspired by" is generally acceptable. "Identical to" crosses into infringement.

Documentation proves human authorship and protects against infringement claims. Keep records across five areas.

Creative development

• Creative briefs and strategic direction
• Mood boards and reference materials
• Initial concept development

Production records

• Prompt history and iterations
• Raw AI outputs before editing
• Human modifications and editing decisions
• Version history with timestamps

Verification and approval

• Reverse image search results and plagiarism check reports
• Trademark screening results and brand similarity reviews
• Quality review notes and stakeholder approvals
• Final sign-offs with dates

Platform documentation

• Commercial licence confirmations
• Terms of service at the time of creation
• Indemnification coverage details

The $1.5 billion Anthropic settlement turned on evidence: Anthropic could not prove legitimate training-data sourcing. Retain records for a minimum of seven years, the statute of limitations for copyright claims in most jurisdictions, in a version-controlled system with an audit trail.

Yes, when you understand the legal requirements and limitations. The safeguards scale with volume.

Small-scale use, under 100 assets a month

• Free tools with commercial licences are acceptable
• Document human contributions
• Maintain basic verification workflows

Medium-scale use, 100 to 1,000 assets a month

• Enterprise platforms with indemnification recommended
• Systematic documentation processes
• Regular legal review of outputs and trademark screening

Large-scale production, 1,000+ assets a month

• Enterprise platforms are mandatory
• Comprehensive documentation systems
• Dedicated legal review workflows and essential indemnification coverage

By client type: agencies are liable to clients for delivered content, and platform indemnification does not transfer automatically. Brands must protect their own trademark while avoiding others, since the FTC holds them responsible for AI-generated product claims. Manufacturers and retailers need consistent processes, because one infringement claim across thousands of SKUs can shut down an entire catalogue.

Copyright principles align globally, since human authorship is required everywhere, but platform availability and enforcement vary significantly.

Platform availability

• China: OpenAI and Google services are blocked or restricted; domestic alternatives such as Alibaba Tongyi Qianwen, Baidu ERNIE, and ByteDance platforms require commercial licences
• Hong Kong, Singapore, Japan, and South Korea: full access to Western enterprise platforms, strong IP frameworks, and reliable enforcement

Copyright law variations

• China: the Copyright Law (2020 revision) requires human authorship; a first-to-file system means documentation timing matters, and registration aids enforcement
• Japan: protection similar to US and EU frameworks; registration is optional but provides litigation advantages
• Singapore: a common law system aligned with UK and US principles, with robust IP enforcement

If you create content in the US or Asia for EU markets, the August 2026 EU AI Act labeling requirements still apply. Images, audio, and video distributed in the EU must be labeled as AI-generated, with penalties up to €35M or 7% of global revenue. Multi-jurisdiction campaigns require compliance with the strictest applicable law, which is often the EU.