Cloud giants offer limited protection to companies against AI copyright infringement claims

Amazon, Microsoft and Google are offering new services such as virtual assistants and chatbots as part of a multibillion-dollar bet on generative AI (systems that spit out human-like text, images, and code in seconds). I’m competing.

AI models are “trained” based on data found on the internet, such as photos and text. This raises concerns that rights holders, from media companies to image libraries, could pursue legal claims against third parties that use AI tools trained on copyrighted data. is occurring.

The three major cloud computing providers are committed to protecting their business customers from such intellectual property claims. However, an analysis of indemnification provisions published by cloud computing companies shows that legal protection applies only to the use of models developed by or under the supervision of Google, Amazon, and Microsoft. Masu.

“Compensation payments are very smart business. . . . They can make people think, ‘I can afford this,'” said Matthew Sugg, a law professor at Emory University.

But Brenda Leung, a partner at Luminos Law Firm, said: “It’s important for companies to understand that.” [the indemnities] Very narrowly focused and defined. ”

Google, Amazon and Microsoft declined to comment.

Compensation provided to customers includes the use of third-party models, such as those developed by Anthropic, an AI startup that counts Amazon and Google as investors, if those tools are available on the cloud company’s platform. But it’s not included.

In Amazon’s case, it only covers content generated by its models, such as Titan, and the company’s various AI applications.

Similarly, Microsoft only provides protection for the use of tools that run on its own models and tools developed by OpenAI, a startup with which it has a multibillion-dollar partnership.

“People needed those guarantees to buy products because they… [the legal] There are risks,” said an IP lawyer working on the issue.

Meanwhile, three cloud providers are adding safety filters to their tools aimed at filtering out potentially problematic content that is generated. While the technology community was “satisfied that there are very few instances of infringement,” it did not want to offer “unlimited” protection, the lawyer said.

While the compensation policies announced by Microsoft, Amazon and Alphabet are similar, cloud companies say customers may want to negotiate more specific compensation in contracts tailored to their needs, but that’s not yet common. said sources close to the.

OpenAI and Meta are among the companies battling the first generative AI test case, brought to you by noted author and comedian Sarah Silverman. They primarily focus on allegations that the companies developing the models illegally used copyrighted content to train the models.

Test cases in particular can “take a significant amount of time to resolve,” which can create a period of “uncertainty,” giving users concerned about the possibility of further litigation additional “security issues.” ” Coverage is offered in tiers, Angela said. Mr. Dunning is a partner at the law firm Cleary Gottlieb.

However, Google’s coverage does not apply to models that customers have “tweaked” using internal data. This allows companies to train generic models to produce better and more specific results. Microsoft coverage, on the other hand, still applies.

Although Amazon does cover these customized Titan models, the protection is void if the alleged infringement is due to a tweak.

However, it may be difficult to prevail on legal claims brought against users of generative AI tools rather than their creators.

In dismissing parts of a lawsuit brought by three artists against AI companies Stability AI, DeviantArt, and Midjourney a year ago, U.S. Judge William Orrick said one “problem” was that the tools He said it was “not plausible” that all images produced by were dependent. “Copyrighted Training Images”.

For copyright infringement to apply, it must be proven that the AI-generated image is “substantially similar” to the copyrighted image, Orrick said.