Midjourney’s latest terms-of-service change was small on the surface, but it lands inside one of the biggest unresolved questions in generative AI: what happens when an AI company builds powerful tools from copyrighted material?
The image-generation startup has not taken the more cautious path some rivals have chosen. Instead, its approach suggests confidence that courts will ultimately accept the argument that AI training can be protected by fair use.
A small policy change with large implications
Last week, Midjourney made a brief change to its terms of service related to IP disputes. The update replaced more casual language with clauses that were more formal and legal in tone.
That wording shift matters because Midjourney is already operating in a heated copyright environment. The company builds AI image generators and is also moving toward video generation, placing it squarely inside the debate over how training data is collected and used.
Generative AI systems such as Midjourney’s are trained on huge collections of examples. Those examples can include images and text gathered from public websites and repositories across the web.
AI vendors commonly argue that fair use protects this practice. In their view, training a model can qualify as a transformative use of copyrighted material because the system is being used to create a secondary product rather than simply republish the original works.
Many creators reject that framing. Their concern is not only that their work may be included in training data without permission, but also that models can produce outputs that resemble or reproduce material they were trained on.
Why creators are pushing back
The copyright dispute around AI training is not only theoretical. A growing body of studies has shown that models can, and sometimes do, regurgitate training data.
That point undercuts the cleanest version of the AI industry’s argument. If a model can reproduce recognizable pieces of its training material, creators have stronger reasons to question whether the process is sufficiently transformative.
Some AI companies have tried to reduce that tension. According to the source article, certain vendors have made licensing deals with content creators, created opt-out systems for training datasets, or promised to cover legal fees if customers are drawn into copyright lawsuits because of their use of generative AI tools.
Midjourney is not described as taking that kind of proactive route. Instead, the company has been notably direct in its use of copyrighted works.
At one point, Midjourney maintained a list of thousands of artists whose works were, or would be, used to train its models. The list included illustrators and designers connected to major brands such as Hasbro and Nintendo.
A study also presented convincing evidence that Midjourney used TV shows and movie franchises in its training data. The examples named in the source include “Toy Story,” “Star Wars,” “Dune” and “Avengers.”
The fair use gamble
Midjourney’s position may prove correct. If courts decide that fair use does apply to model training, the company could keep operating much as it has, including scraping and training on copyrighted data both old and new.
That outcome would be important far beyond Midjourney. It would strengthen the position of AI vendors that rely on large-scale web data and would weaken creator arguments that permission or licensing should be required before copyrighted works enter training datasets.
But the opposite outcome would be severe. If fair use does not protect Midjourney’s training practices, the source article says the decision could decimate the company overnight.
That is what makes the company’s apparent confidence so consequential. Midjourney is not merely adjusting legal language. It is signaling that it believes the legal system will allow the core data practices behind its product to continue.
The risk is amplified by the company’s reported business position. Midjourney has reportedly reached around $200 million in revenue without outside investment. That gives the startup momentum, but copyright litigation can be expensive, and the stakes are existential if the company’s training methods are found to be legally vulnerable.
What this means for the AI market
The Midjourney dispute captures a broader split in generative AI strategy. One group of companies is trying to lower legal exposure through licensing, opt-outs or customer protections. Another path is to defend broad training practices and wait for courts to clarify the limits.
For creators, the issue is control. Their work can help shape a model that may later generate competing images, styles or concepts. When there is no licensing agreement or opt-out mechanism, they have fewer practical ways to influence how their work is used.
For AI vendors, the issue is scale. Modern generative models depend on vast collections of training examples. Limiting those datasets to fully licensed material could change the economics and capabilities of AI development.
For users, the question is more indirect but still important. If a customer uses AI-generated work from a tool that later becomes tied to copyright litigation, the legal and commercial uncertainty may affect how safe that tool feels for professional use.
Midjourney’s latest terms change does not resolve any of these issues. It simply makes the company’s posture clearer. The startup appears ready to stand behind the idea that AI training on copyrighted material can survive legal scrutiny.
That may turn out to be a winning bet. It may also become one of the most expensive assumptions in the AI industry.