The AI copyright cases that started filing in 2023 are producing actual decisions in 2025-2026. The New York Times v. OpenAI case, multiple music industry lawsuits, artists suing image generation companies, and authors suing for training data use are all moving through courts simultaneously. The outcomes will shape what AI companies can train on and what liability exists for using AI-generated content.
Here’s the current state of the legal landscape and what it practically means for developers and organizations using AI.
The Core Legal Question
At issue is whether using copyrighted content to train AI models constitutes copyright infringement, and if so, whether it falls under “fair use.”
The traditional four factors for fair use analysis:
- Purpose and character: Is the use transformative? Commercial or educational?
- Nature of the work: Factual vs. creative works (creative works get more protection)
- Amount used: How much of the original was copied?
- Market effect: Does it harm the market for the original work?
AI training data arguments on each factor:
For training being fair use:
- Training is transformative — the model learns patterns, doesn’t reproduce works
- Models typically don’t output exact copies of training data (though they sometimes do)
- The training process is different from simple copying/distribution
Against training being fair use:
- AI companies are commercial entities using creative works for commercial advantage
- Training data scrapers consumed and processed entire works
- AI outputs compete with the original works (image generators vs. illustrators, etc.)
- The market for licensing training data is being preempted by scraping
Major Cases and Where They Stand
New York Times v. OpenAI/Microsoft
Filed December 2023. The NYT alleges that OpenAI trained on millions of its articles without permission, and that ChatGPT can reproduce them verbatim in some cases. This last point is significant — the Times provided examples of ChatGPT outputting near-verbatim articles, which undermines the “transformative” argument.
As of early 2026: Discovery ongoing. The case has become focused partly on whether OpenAI has “memorized” training data and the technical question of what percentage of outputs can be tied to specific training examples.
Significance for developers: If the NYT wins on the memorization argument, it creates pressure for AI companies to implement stronger training data filters and potentially limits what models can be trained on.
Music Industry Lawsuits
The RIAA and multiple major labels filed against music AI companies (Suno, Udio) for training on copyrighted recordings. An important distinction from the text cases: music recordings have two copyrights — the composition (lyrics/melody) and the sound recording. The sound recording copyright is particularly strong.
In August 2024, a court ruled against Suno in a summary judgment motion, finding that their training on recordings likely constitutes infringement. This was the first major adverse ruling for an AI company on training data.
Significance: Music AI training without licenses appears to be legally risky. Expect either licensing deals (Suno did reach a settlement agreement with several labels) or a shift to training on licensed/public domain music.
Getty Images v. Stability AI
Getty sued Stability AI (maker of Stable Diffusion) for scraping 12 million images from Getty’s site. Key points: Getty’s watermarks appeared in some generated images, clearly showing their images were in the training set. This is harder to defend as transformative use when watermarks appear.
As of early 2026: Case is ongoing in UK and US simultaneously.
Authors Guild Cases
Multiple authors have sued AI companies for training on their books. The cases are being consolidated and proceeded through 2025. The Authors Guild argument: training on books is not transformative because the output competes directly with the books themselves.
What the Legal Landscape Means Practically
For Using AI Tools in Your Work
Using AI coding assistants (GitHub Copilot, Cursor): The legal risk here is mostly on the AI companies, not the users. If GitHub/OpenAI’s training was infringing, they bear the liability, not you.
However: be cautious about AI code that looks like it’s reproducing recognizable patterns from specific open-source projects. GitHub Copilot has filtering to try to avoid reproducing verbatim code from training data; not all tools do.
Copyright in AI-generated outputs: Currently in the US, pure AI-generated content (with no human authorship) is not copyrightable. The Copyright Office has clarified this repeatedly. If you use AI assistance but exercise creative judgment in selecting/editing outputs, you may have a copyright interest in the result.
This matters for: software, written content, images, music. If you want to copyright AI-assisted work, document your creative choices.
Using AI-generated images commercially: Until the training data cases settle, there’s uncertainty about whether AI-generated images carry the infringement risk from training data. For commercial use, prefer AI tools that use licensed or public domain training data (Adobe Firefly trained on licensed stock; Shutterstock’s AI uses licensed content).
For Organizations Deploying AI
Internal tool liability: If you build internal tools using AI APIs, the legal risk sits primarily with the foundation model provider, not you. Your terms of service agreement with OpenAI/Anthropic/etc. matters here.
Customer-facing AI products: More complex. If your product generates creative content, and that content turns out to be infringing (because the model was trained on infringing data), the question of upstream vs. downstream liability is unsettled.
Train your own models: If you fine-tune or train models, you own the training data liability. Make sure your training data is licensed or demonstrably fair use.
For Creators
The legal cases are establishing that scraping creative work for commercial AI training without compensation is contested. Some practical paths forward:
License your work proactively: Shutterstock, Adobe, and other platforms are paying creators for their work being used in training. If you’re a visual artist or photographer, this is an option.
Opt-out: Many AI training datasets now have opt-out mechanisms. Spawning’s “Have I Been Trained?” lets you check if your images are in major training datasets. Robots.txt is being considered for a more standardized approach.
Watermarking and fingerprinting: Tools like Glaze and Nightshade (for visual artists) attempt to add data to images that disrupts AI training while being imperceptible to humans.
The EU AI Act Dimension
The EU AI Act requires GPAI model providers to publish summaries of training data used. This transparency requirement will make it easier to identify when specific copyrighted works were used, strengthening litigation positions.
EU AI Act Article 53:
General-purpose AI model providers must:
- Maintain technical documentation about training data
- Implement a copyright policy
- Publish summaries of content used for training
What’s Likely to Happen
Legal analysts generally expect some form of licensing framework to emerge — either through court decisions that establish liability (forcing AI companies to license), legislation, or negotiated industry agreements. The music industry has navigated similar fights before (with streaming, user uploads, etc.) and settled on licensing structures.
The most likely outcome for text and images: AI companies that train on copyrighted content will need to either license it (as some music AI companies have done) or demonstrate that their training methodology qualifies as fair use under evolving case law.
For open-source training datasets: datasets like LAION (used to train Stable Diffusion) are under pressure. A new generation of curated, licensed datasets is emerging.
Conclusion
The AI copyright landscape is in genuine flux. The outcomes of these cases will shape how foundation models are trained, what training data they can use, and what liability exists in the AI supply chain.
For developers and organizations: the immediate practical exposure is relatively low if you’re using established AI APIs and tools. Watch the NYT case and music industry cases — those decisions will establish precedent. For any AI-generated content you’re commercializing, document your creative contributions and prefer tools with clear training data provenance.
This is an area where the law is running years behind the technology, and certainty won’t come quickly. Stay informed and review your organization’s AI usage policies annually as the case law develops.
