Microsoft Sees Low Risk for Customers in AI Copyright Lawsuits

Microsoft’s promise to defend customers from copyright infringement lawsuits stemming from its Copilot artificial intelligence tools likely won’t require a mass deployment of its lawyers.

The company believes its systems are unlikely to produce content that so closely copies its source material that it violates copyright, a legal executive at the company told Bloomberg Law after a Sept. 7 announcement.

Legal professionals largely agreed that the risk of customers facing a copyright infringement suit related to AI’s outputs was low—though generative AI platforms themselves still face a substantial legal reckoning over the inputs, or material that went into training the models.

Microsoft is extending indemnification to protect commercial customers of its Copilot tools—which provide generative AI assistance as part of Office, GitHub, and other products—when they’re sued for copyright infringement based on outputs. Microsoft will defend the customer and pay any adverse judgments or settlements, the company said.

The move aims to quell fears among potential customers that using generative AI-derived content will expose them to copyright suits. But that chance is small, according to the Microsoft lawyer and several legal experts.

“This is a smart move by Microsoft because the risk of Copilot output adopted by one of their customers being found to be infringing is incredibly low,” said Matthew Sag, professor of law, artificial intelligence, machine learning and data science at Emory University Law School.

The indemnification doesn’t take pressure off the most critical legal questions about generative AI and intellectual property, as authors, artists, and other creators sue Microsoft and Open AI for training the AI models on copyrighted materials. But those questions are confronting the platform creators themselves, not customers.

“The copyright concerns around generative AI are make or break questions for the industry. Microsoft has to validate the legitimacy of using their tools or the entire sector goes away,” said Eric Goldman, a professor at Santa Clara University School of Law who focuses on intellectual property and technology law.

‘Good Move for PR’

Lawsuits so far haven’t targeted end-users of generative AI, and users shouldn’t be liable for how AI companies train their machines, said Theresa Weisenberger, a partner at BakerHostetler LLP who works on IP and technology law. But it’s unclear how the law would treat AI customers using infringing outputs without knowing they resembled an existing work, she said. She said the announcement should at least help reassure potential users of the products.

“My clients are understandably concerned about using generative AI based on this type of litigation,” Weisenberger said. Generative AI’s copyright questions are a “high-profile risk” and the law isn’t settled, she said.

She also warned users to heed Microsoft’s parameters for coverage, ensuring they don’t manipulate prompts or take other steps that would foreseeably produce infringements.

Microsoft built features into the models designed to reduce purely duplicative outputs that could raise copyright concerns—for example, natural language filters that “try to limit the number of verbatim outputs that we can include in any particular response,” the Microsoft lawyer said.

Microsoft’s indemnification commitment “puts their money where their algorithmic development is,” said Chris Callison-Burch, an associate professor of computer and information science at the University of Pennsylvania’s School of Engineering and Applied Science who studies AI. It’s a good move for PR and reassurance, he added, which “also aligns with their technical goals of trying to mitigate this unintentional replication of copyrighted data.”

Limitations on who can claim coverage, as well as the fact that Microsoft would likely already be a co-defendant in lawsuits stemming from their AI, would limit the number of applicable cases, Goldman said.

The company would rather control the litigation than have their customers control it because Microsoft cares more about the result than any individual customer does, Goldman said.

To be sure, cases of duplication are not unheard of. For example, in a lawsuit against OpenAI and GitHub, programmers said chunks of their code are reproduced in GitHub Copilot’s suggestions.

The Real Issue

The models were trained by scraping vast quantities of data across the internet, including materials that are under copyright. Ongoing litigation, including a suit brought by comedian Sarah Silverman, will decide whether that constitutes copying under copyright law, and if so, whether it’s fair use. Fair use is the legal principle that permits the unlicensed use of copyrighted works in certain circumstances, typically taking into account the nature and purpose of the use, among other things.

Outputs would also have to be analyzed under fair use, including the extent of their effect on the market for the original work and derivatives—often considered the most important factor.

Shubha Ghosh, a professor at Syracuse University College of Law and director of the Syracuse Intellectual Property Law Institute, said AI companies may find refuge in the Second’s Circuit’s 2015 opinion involving Google Books.

There the court deemed Google copying tens of millions of books to be fair use because Google used the resulting data for linguistic analysis and only made excerpts available online. The research purposes and limited excerpts were transformative, while the use didn’t significantly affect the originals’ market.

AI companies could argue that machine-learning copying is generally even more transformative since a scanned work provides merely a tiny fraction of the input into the resulting output.

While Thursday’s announcement may reassure customers, bigger legal questions still remain, including ones that extend beyond copyright.

“It’s quite an intimidating legal environment that generative AI faces, and I’m generally concerned it can’t survive all of the legal bases,” Goldman said.