Midjourney在版权诉讼中反将一军，要求大厂公开自己是否也偷偷用版权数据训练AI——如果大家都是这么干的，凭什么只告它？
Midjourney向法官提交动议，要求迪士尼、环球和华纳兄弟披露其内部AI训练实践、训练数据集和模型权重。当前法院仅将发现范围限制在面向消费者的AI，Midjourney希望扩大至内部商业计划和研发报告。如果这些工作室也在使用未授权版权内容训练内部AI用于故事板和构思，那么Midjourney的合理使用抗辩将获得行业标准支持。工作室指责这是一场转移注意力的钓鱼取证。
Midjourney wants Disney, Universal and Warner Bros…
Midjourney wants Disney, Universal and Warner Bros to disclose their internal AI training practices in discovery. The argument is straightforward: if the studios are training models on unlicensed copyrighted content for storyboarding and ideation, then it's industry standard practice and Midjourney's fair use defence strengthens.
The magistrate limited discovery to consumer-facing AI only. Midjourney just filed a motion asking Judge Kronstadt to overturn that, requesting access to AI business plans, training datasets, model weights, board presentations and research reports. The studios say it's a fishing expedition to distract from infringement of their characters.
I'm leaning toward Midjourney's side here.
If studios are doing exactly what they're suing Midjourney for, that's hypocrisy. Internal tools trained on third-party copyrighted data would prove industry custom and directly support fair use. The asymmetry is troubling - studios can train on whatever they want behind closed doors, but sue commercial tools that do the same thing publicly. Discovery should reveal if that's happening.
The training phase shouldn't be where we draw the line. Humans train by consuming massive amounts of freely available copyrighted content. We read news articles online, watch YouTube videos, listen to streaming music, browse images and websites. That's all copyrighted material accessed for free. We absorb it, get inspired by it, learn from it. We call that learning. AI training from publicly available free content is functionally the same thing.
The line should be at distribution and monetisation. When you're making money by letting users generate specific copyrighted characters at scale, that's when it becomes a problem. The harm isn't in the learning, it's in the commercial output capability. Studios see revenue loss when users can generate Mickey Mouse on demand.
The solution probably looks like what's emerging in music AI. Labs are starting to license content and share revenue with rightsholders. You pay for the ability to remix or generate, and the original creators get a cut of the distribution income. That model could work across all generative AI if we actually built the framework for it.
But right now we have no coherent legal framework for any of this. Studios want protection for their IP but also want freedom to experiment internally. Midjourney wants to operate commercially but claims fair use. The rules haven't caught up to the technology, and until they do, cases like this are just fights over who gets to set the precedent.
If the studios are training on unlicensed data internally, Midjourney absolutely deserves to know that. https://t.co/IORIF5SKM7