Copyright and Generative AI: A Practical Guide for Legal Teams
You might not think about copyright every day, but you own it. If you've written a memo, taken a photo, or recorded a song, you've created a protected work. Generative AI puts new pressure on those rights, and the stakes are high on both the development and output sides.
Tech companies are hungry for high-quality, human-made content to train models. Creators want clarity on whether AI-assisted outputs can be protected. Courts, agencies, and companies are moving fast, and the rules are being tested in real time.
Copyright, in brief
Copyright protects original works fixed in a tangible form-books, art, software, music, films, choreography, architectural designs, and more. The owner controls reproduction, distribution, public display, performance, and derivative works.
As the US Copyright Office notes: create an original work and fix it, and you are the author and owner. That baseline is unchanged by AI-but how it applies around AI inputs and outputs is where the friction starts.
Where AI and copyright collide
- Outputs: Can a user claim protection for AI-enabled text, images, audio, or video?
- Training inputs: Can developers train on copyrighted material without permission, or does that trigger infringement unless licensed?
Can AI-generated images or text be copyrighted?
Short answer: it depends on human authorship. The US Copyright Office has made clear that works generated entirely by AI, without sufficient human creative control, are not eligible for registration. Human authorship is the touchstone.
AI-assisted edits-adding or removing objects, de-aging, noise reduction, or other refinements-can be registrable, but the AI use must be disclosed. Applicants should limit their claim to the human-authored portions and exclude purely machine-generated material. In rare cases, selection, arrangement, or meaningful transformation of AI outputs can meet the threshold, but you'll need evidence of creative control and contribution.
For official guidance, see the US Copyright Office's AI resources here.
Can copyrighted content be used to train AI?
With a license, yes. Many publishers have struck deals with AI developers to authorize use of their catalogs. Without permission, creators argue it's infringement-reproduction, distribution, public display, or derivative use-absent a valid defense.
Dozens of lawsuits are pressing that point, including high-profile actions against major AI companies. Courts will sort whether copying for training crosses the infringement line or fits within an exception.
Fair use: where the fight is concentrated
Fair use allows certain uses of copyrighted works without permission, assessed under four factors. It's fact-intensive and context-specific-there are no automatic wins.
- Purpose and character: Commercial vs. noncommercial, and whether the use is transformative.
- Nature of the work: Factual vs. highly creative.
- Amount and substantiality: How much is used and whether it captures the "heart" of the work.
- Market effect: Whether the use substitutes for or harms the market for the original.
With AI, key questions include whether the analysis applies to inputs (mass ingestion during training), outputs (any recognizable expression), or both. Tech companies argue training is transformative and non-substitutive; creators counter that wholesale ingestion at scale and output substitution undercut markets. Early rulings are mixed and fact-specific, and more are coming.
Statutory text: 17 U.S.C. § 107 (Fair Use).
Practical steps for in-house counsel and firms
- Licensing and procurement: Require training-data transparency, source disclosures, and change notifications. Negotiate warranties, indemnities, DMCA 1202 representations, and caps that match exposure. Address dataset retention and model retraining obligations on takedown.
- Creator-side protection: Keep creation logs showing human contribution, prompts, edits, and version history. Register works early. When filing, exclude machine-generated portions and claim only the human-authored elements.
- Policy and compliance: Publish clear internal rules for AI use, especially for confidential data, trade secrets, and export-controlled material. Set review gates for public releases that used AI in the process.
- Monitoring and enforcement: Track model outputs that mirror your works. Consider notices, negotiated licenses, or test cases where copying appears substantial or market-substitutive.
- Risk controls for media assets: Use content authenticity signals, watermarks, and unique markers to support provenance claims. Consider collective licensing or private deals when economics make sense.
- Contracts with AI vendors: Specify output ownership, training restrictions on your data, audit rights, and remedies for misattribution or leakage. Address model updates that may alter risk profiles.
Key issues to watch
- Fair use outcomes: How courts treat training ingestion, intermediate copying, and output similarity. Watch how "transformative" is applied at scale.
- Derivative works and substantial similarity: Standards for AI outputs that echo protected expression, and what evidence moves the needle.
- US Copyright Office activity: Guidance updates, rulemaking, and registration practices for AI-assisted works (USCO AI hub).
- Licensing markets: More publisher-platform deals, opt-in repositories, and standardized terms for training access.
- Contract norms: Warranties, indemnities, and audit frameworks becoming standard in SaaS and data agreements.
Bottom line
Human authorship still carries the day. If a person's creative choices drive the result, protection is on the table; if a model did the expressive work, it likely isn't. On training, licensing is clean, unlicensed scraping is risky, and fair use remains the contested middle.
For now, tighten contracts, document human creativity, and keep options open-license when it pencils out, contest use when it doesn't, and stay close to the cases and guidance that will set the next round of norms.
Want to upskill your team on practical AI use by role? Explore curated programs at Complete AI Training - Courses by Job.
Your membership also unlocks:
