AI Copyright Lawsuits Are Becoming a Debate About Who Owns the AI Economy

The legal fight over how artificial intelligence companies use copyrighted material is expanding and it has now drawn the direct involvement of the U.S. Department of Justice.
The New York Times has sued OpenAI and Microsoft. The Seattle Times and Newsday have filed a separate lawsuit. Other publishers, authors and rights holders are pursuing related claims over the use of their work in training and operating generative AI systems. The DOJ has now formally stepped into the consolidated litigation to support OpenAI’s position that training AI models on copyrighted material can qualify as fair use.

These cases raise important questions about copyright, fair use and competition. But they also point to a broader issue that is directly relevant to the future of Action Model:
As AI systems become more capable by learning from human-created information and behaviour, who should participate in the value those systems generate?

That question extends beyond whether a particular article, book or image was copied. It concerns the structure of the AI economy: who supplies the intelligence, who controls the infrastructure and who owns the resulting systems.

The current legal dispute

The New York Times sued OpenAI and Microsoft in December 2023, alleging that millions of its articles were used without permission to develop AI systems and that those systems could compete with the journalism that funded the original reporting. OpenAI and Microsoft have disputed the claims.

The central legal issue is whether the use of copyrighted material to train AI models can qualify as fair use.

OpenAI has argued that training involves transforming large volumes of information into systems capable of performing new tasks. Publishers, meanwhile, argue that commercial AI companies are using their work to build products that can reproduce, summarize or replace access to the original material without an appropriate licence or compensation model.

The concern is not limited to the existence of a model that has learned from a large body of information. Publishers are also concerned about what happens when that model generates an answer that closely reproduces the source, or when it makes a copy of a copy available to users at scale.

In that scenario, the AI system may not simply be learning from an article in the abstract. It may be reproducing the article’s expression, structure or reporting in a form that can substitute for visiting the original publication. A user may receive a condensed version, a near-verbatim passage or a synthesized response assembled from multiple sources without ever reaching the publishers that paid to create and verify the underlying work.
That distinction is central to the publishers’ concerns. The issue is not only whether content was included in training data. It is also whether the resulting system can make copies, summaries or derivative versions of that content available in ways that compete with the original source.

The issue is also not limited to The New York Times. On September 4, 2026, The Seattle Times and Newsday filed a copyright lawsuit against OpenAI and Microsoft in the Southern District of New York. Their complaint similarly argues that their journalism was used in connection with commercial AI systems and that those systems could weaken the business models that support original reporting.

Their concerns reflect a wider problem facing the information economy. Publishers invest in reporting, editing and verification. AI systems can then deliver information directly to users, potentially reducing traffic, subscriptions and advertising revenue for the organizations that produced the underlying work.
The risk becomes more serious when an AI system can provide a copy of a copy: a response that draws from the original reporting, reproduces its substance or expression and makes the result available without sending the user to the source. Even when the output is not identical to the original, it may still capture enough of the source’s value to reduce the incentive to access, subscribe to or support the publisher that created it.
If that pressure causes high-quality publishers to reduce their output or close, future AI systems may also have less reliable human-created information to learn from.

This is why the lawsuits matter beyond the parties involved. They are testing not only the boundaries of copyright law, but also the relationship between AI companies and the people and institutions whose work makes AI more useful.

The Department of Justice steps into the lawsuit

On September 1, 2026, the U.S. Department of Justice filed a formal Statement of Interest in the consolidated OpenAI copyright litigation. This is one of the most influential interventions in the case so far because it places the position of the U.S. government directly before the court as the parties seek summary judgment.

The DOJ supported OpenAI’s broad argument on AI training, stating that:

“The training of AI models on copyrighted material, in and of itself, does not violate copyright laws.”

The department argued that training large language models can be highly transformative and that an overly restrictive interpretation of copyright law could slow scientific progress, economic growth, national security capabilities and U.S. competitiveness in artificial intelligence. It also argued that the creative and public benefits of AI training can outweigh broader claims of competitive harm.

This is a significant development for OpenAI. The DOJ is not simply observing the debate; it has formally stepped into the lawsuit under a federal law that allows it to represent the interests of the United States in pending litigation. Its position may carry considerable influence as the court considers how fair-use principles should apply to AI training.

However, the DOJ’s intervention does not close the case or determine its outcome. The filing is not a court judgment, and the judge is not required to adopt the government’s reasoning. OpenAI, Microsoft and the publisher plaintiffs are still pursuing competing summary-judgment arguments, and the central copyright questions remain unresolved.

The DOJ’s position also reinforces an important distinction between training and output. It argues that training a model on copyrighted material does not automatically constitute infringement. That does not necessarily protect an AI product when its output reproduces protected expression or acts as a substitute for the original work. The government has separately recognised that creators should remain protected from AI-generated outputs that infringe their content.

The legal debate therefore cannot be reduced to a single question about whether AI companies can train on copyrighted material. Courts must also consider how the material was obtained, how it was used and whether the resulting systems make protected work available to users in a substantially similar or substitutive form.

The DOJ’s intervention may help shape the legal boundary. It still does not answer the wider economic question at the centre of this article: if human knowledge and behaviour make AI more capable, who should participate in the value those systems create?

Copyright does not answer every economic question

Copyright law can determine whether a protected work was copied, whether a use was transformative, whether it competes with the original and whether a legal exception applies.
It can also help distinguish between a system that learns general facts or patterns and one that makes protected expression available to users in a way that substitutes for the source.
But copyright cannot, by itself, determine how the broader value created by AI should be distributed.
A model may be trained on books, articles, code, images and public discussions. It may also improve through human feedback, evaluation, corrections and usage patterns. As AI develops into systems that can complete tasks rather than simply generate text, another form of contribution becomes increasingly important: human action.
People demonstrate how software is used, how workflows are completed, how decisions are made and how problems are solved. These actions can provide the examples needed to build AI systems that operate effectively in real environments.
That creates a distinction between two questions:

1. Was a particular work used lawfully, and did the resulting system reproduce or substitute for the source?

2. How should the people whose knowledge, feedback and actions improve AI participate in the value created by the resulting system?

The first is primarily a legal question. The second is an economic and institutional one.
The current AI industry has largely separated contribution from ownership. People create the information and behaviour that make systems more capable, while a relatively small number of companies control the models, infrastructure, distribution channels and commercial returns.
The reproduction problem makes that separation especially visible. A publisher may create an original investigation. An AI system may learn from it, generate a condensed version and deliver that version directly to a user. The original creator supplied the work, while the platform controls the interface through which the value is accessed.
That structure may be efficient for building products quickly, but it leaves open a fundamental question: whether the people who help produce machine intelligence should remain only sources of data, users of the final products or customers of the companies that control them.

Why this matters for Action Model

Action Model is focused on a part of the AI stack that is becoming increasingly important: the data of human action.
Large language models learn from what people write. Large Action Models are intended to learn from what people do: how they interact with tools, complete tasks, navigate systems and turn instructions into outcomes.
This shift creates both a technical opportunity and an ownership challenge.
If AI agents are expected to perform useful work in the real world, they will need more than static information. They will need examples of successful actions, context, feedback and decision-making. The people who provide those contributions may therefore influence not only the content of an AI system, but also its ability to operate.
Action Model’s approach is to build a community-owned Large Action Model around this contribution. The objective is to create a system in which participation is not treated solely as an input to be collected, while ownership and value remain concentrated elsewhere.
Instead, the network is designed around a closer relationship between contribution and participation. People contribute actions, feedback and expertise. Those contributions help improve the model. As the network develops, contributors can participate in the ecosystem built around the intelligence they helped create.
This is not a claim that every contribution can be valued in a simple or identical way. Nor does it eliminate the legal and practical challenges involved in data rights, privacy, attribution, verification and compensation.
It does, however, establish a different principle:
The people who help create useful intelligence should have a path to participate in the systems and value that emerge from it.

From content reproduction to action ownership

The current copyright lawsuits focus primarily on published content. They also highlight a broader concern: when AI systems make a copy of a copy available, the economic value of original human work can be separated from the people and institutions that produced it.
That is likely to be only the first phase of the ownership debate.
As AI agents become more capable, companies will seek increasingly valuable datasets showing how humans perform tasks. These datasets may include software interactions, operational workflows, research processes, customer-service decisions, creative production and other forms of skilled activity.
The economic importance of this information could be substantial. A system that can reliably complete a complex workflow may be more valuable than one that can merely describe the workflow. The data required to train that capability may also be more difficult to produce, verify and replace.
In this context, the equivalent of a copy of a copy may not be a duplicated article. It may be an AI agent that has absorbed a person’s workflow and can reproduce the result of that expertise without recognizing or compensating the people whose actions made the capability possible.
This raises questions that existing copyright categories may not fully address:

• Who owns a recorded workflow?

• Who should benefit when repeated human actions improve an AI agent?

• How should contributors be recognized when their input is combined with thousands of others?

• What rights should people have over the use of their actions?

• Can contribution be measured without compromising privacy?

• How can value be distributed without creating incentives for low-quality or manipulated data?

• When an AI system reproduces the outcome of human expertise, how should the original contributors participate in the value created?

These questions will require technical, legal and economic solutions. They also create an opportunity to design AI networks differently from the centralized platforms that dominated the previous era of the internet.

The choice ahead

One possible model is familiar. People generate the data, feedback and actions. A small number of companies aggregate those contributions, build proprietary systems and retain control over the resulting value. The public participates mainly as users, data sources or customers.
In that model, the distance between original contribution and commercial output can grow over time. A publisher creates an investigation; an AI system produces a substitute summary. A worker demonstrates a process; an AI agent performs a similar process. A community generates knowledge; a platform packages that knowledge into a product.
Another model would give contributors a more meaningful role in the networks they help build. Contribution could be measured and verified. Participation could be connected to governance or economic rights. The infrastructure could be designed so that the growth of AI does not automatically require the concentration of ownership.
The lawsuits involving OpenAI and publishers will not settle all of these questions. The DOJ’s intervention could significantly influence how the court approaches fair use in AI training, but it cannot determine the judgment. A court may decide whether particular uses of copyrighted material are lawful, whether an output is sufficiently transformative and whether making a copy of a copy available competes with the original source. But legality alone will not determine whether the overall AI economy is fair, sustainable or broadly participatory.

That debate will continue as AI systems learn from more than published content. It will expand to include feedback, preferences, decisions and actions. It will involve not only publishers and technology companies, but also workers, creators, researchers, software users and communities whose activity helps make AI more capable.
The central issue is therefore larger than whether AI companies can train on a particular article.

It is whether the people who supply the knowledge and actions that make AI valuable should have any ownership or economic participation in what follows.

It is also whether AI systems should be allowed to turn original human work into copies, summaries and substitutes while leaving the original contributors outside the value chain.
Action Model is being built around the view that they should have a path to participate.

Train it. Earn it. Own it.
→ actionmodel.com

Sources: Associated Press · Statement of Interest of the United States