Caveat creator: GenAI giants’ pledges won’t pre-empt copyright suits

One year after Wall Street banned the use of generative artificial intelligence (GenAI) bots, the prevailing defensive stance towards the technology has shifted to one of active exploration.

Even banks – known for their conservatism and risk aversion – are now racing to experiment with hundreds of use cases. Some are already opening up access to GenAI’s possibilities to large numbers of employees.

But in the midst of this boom, one question has become a persistently worrisome part of the discussion: who owns the generated output?

When I look at the way their language is drafted ... either these people at tech companies are bad drafters or they are intentionally building ambiguity. And I’m going to guess the latter

Kate Downing, IP lawyer

Most tech vendors assert in their service terms that users have full ownership of their GenAI output, but from a legal standpoint, it’s not as straightforward as such assertions would suggest.

In a March conference, Nick Tsilas, Microsoft’s assistant general counsel, acknowledged that users owned the AI-generated output “in quotes”, as they must still cautiously identify the underlying intellectual property (IP) involved and determine whether they had the right to use it – or risk getting tangled in lawsuits.

To ease this concern, large tech vendors – Microsoft, OpenAI, Google and Amazon Web Services (AWS) among them – have made promises to indemnify users against IP claims. This means they will cover legal damages if their customers are sued for copyright infringement for the output generated by their models.

While this commitment sounds promising, eight lawyers specializing in technology and IP transactions spoken to for this article unanimously caution users against treating such indemnity as a panacea – and warn of significant limitations.

“From an actual risk management perspective, I’m not sure how much of this indemnity is really useful, and how much is just a marketing play,” says Van Lindberg, an IP attorney and partner at law firm Taylor English.

Lawyers say there are three major limitations. Firstly, the indemnification does not cover any modified output that diverges from the way the tools are normally used. Secondly, while it covers the costs of adverse judgements and settlement results from the lawsuits, it does not compensate for indirect costs, such as delays in work progress or expenses to redesign products that may result from lawsuits – nor can it indemnify users against reputational damage. And, finally, for some specific use cases, such as coding, numerous preconditions must be met to qualify for indemnity, of which users may not even be aware.

That said, instead of waiting for vendors to clarify their commitments – or for IP law to evolve and become completely clear – there are proactive steps companies can take to mitigate the risks. These include negotiating contracts with vendors, crafting policies to assess risks, filtering out lower-risk use cases, and educating users on copyright infringement.

“It’s a mistake for companies to just sit and wait around,” says James Gatto, an IP attorney and partner at law firm Sheppard Mullin, noting it could take five to 10 years for IP law to fully catch up with AI.

“There are lots of lawsuits pending, and many of these cases are fact-dependent. So, even if there is a decision in one case, you are not going to get judicial decisions that provide clear guidance.”

Conditional contracts

In tech vendor posts promising indemnity, reassurances regarding IP legal risk are set out in encouraging tones. But lurking within are hyperlinks that take the reader to what various vendors call ‘exact terms’, ‘documentation’ or ‘guardrails’ – where only further exploration reveals the preconditions that users must meet to qualify for indemnity.

The language in these terms varies among vendors, and many are open to interpretation. One common aspect, however, is they all disclaim responsibility for any output that has been modified or customized.

Lawyers note this condition can cause many use cases to step beyond the scope of indemnity, including one of the most popular among financial services: coding. For code to be finalized and integrated into the workstream, it normally undergoes a cycle of editing and review, so it’s almost impossible for the final output to be directly AI-generated.

When it comes to generating code that is going to be used in your long-term core products and services, I don’t think indemnity is particularly useful

Tom Magnani, Arnold & Porter

A majority of generative AI tools for code generation, including GitHub Copilot, are trained on source code available in public code repositories – for example, open-source licensed code available on GitHub. But some code is not as ‘open’ as it sounds. Depending on its licensing, users are required to cite the source properly to avoid copyright infringement, or there are limitations on its redistribution, which can be a challenging task for developers. On top of this, modifying the code leaves them even more vulnerable to lawsuits without backup from vendors.

“Indemnity is helpful when you use the output for short-term purposes, such as a short-term ad campaign,” says Tom Magnani, partner and head of technology transactions practice at law firm Arnold & Porter.

“But when it comes to generating code that is going to be used in your long-term core products and services, I don’t think it is particularly useful. You inevitably need to make modifications, either by combining the generated code with other code, or integrating it with other technology.”

This could explain why most banks currently restrict their use of GenAI coding tools to only translating internal code and providing suggestions on small code segments. Three lawyers nonetheless say their clients from other industries are actively evaluating and testing the technology’s capacity to generate longer sections of code. Given its immense potential in enhancing productivity, it is only a matter of time for the financial sector to embrace GenAI, so understanding and mitigating its underlying IP risk should be a top priority.

Muddy waters

As well as unmodified output, tech vendors can require users to meet additional conditions to qualify for indemnity, depending on specific use cases. For example, Microsoft requires users who use the Azure OpenAI service models for code generation to turn on either the ‘annotate’ or ‘filter’ mode.

Gatto at Sheppard Mullin says that, compared with the requirement of unmodified output, these conditions are less challenging to meet – as long as users are aware of them. However, many users fail to carefully go through and identify the terms, he observes.

Even if all preconditions are met for indemnity qualification, lawyers caution that there are additional indirect costs stemming from lawsuits that companies should take into consideration. For example, if a user’s output is determined by a court to be copyright infringement, while tech vendors indemnity can cover adverse judgements and settlement results, the user still needs to bear the cost of replacing copyrighted materials and dealing with any impact, says Kate Downing, an IP lawyer specializing in open-source compliance. This could include workflow delay or expenses to indemnify investors, she adds.

Juliana Neelbauer, a partner at law firm Fox Rothschild, notes that, given tech vendors are willing to fully take over the case, user companies can still be subject to claims filed against them – and, in some cases, need to “fight to get themselves removed” from these claims.

“You cannot have absolutely no involvement in the case if the generated output is being promoted on your websites, your social media sites, or used in your product suites. And if nothing [else], your company name is going to be listed throughout the initial motion as being examples and evidence of the violation,” says Neelbauer.

“I know that Microsoft says ‘you own it’, and … they will back us up if we get sued,” said Christian Kulas, an AI solutions architect at Deutsche Börse, speaking at the same March conference as Microsoft’s Tsilas. “That’s great, but my managers say: ‘OK, if that happens – if there is a lawsuit – it’s already too late.’”

It is also worth noting that not all indemnity provisions from large tech vendors are uncapped. Microsoft, Google, OpenAI and AWS were asked about their respective specifics. AWS is the only one that explicitly states that it offers uncapped indemnity, taking full financial responsibility to protect users. Google states the financial caps are “dependent on the agreement” with users. Microsoft and OpenAI have declined to comment.

IP lawyer Downing says most vendors’ indemnification terms lack clarity: “When I look at the way their language is drafted, there are only two conclusions: either these people at tech companies are bad drafters or they are intentionally building ambiguity. And I’m going to guess the latter.”

Don’t drop out

Although it’s wise to recognize these legal risks, lawyers say users should not stop experimenting with and adopting technologies that could improve their competitiveness in the market solely out of concern to avoid incurring such risks. Instead, there are proactive steps to take to properly evaluate and mitigate them.

All eight lawyers agree there is room for companies to negotiate indemnity terms with vendors, including the large ones. Negotiation varies depending on the individual cases, but in general, companies should not hesitate to invest time in clarifying ambiguous terms and in removing unreasonable preconditions, says Downing.

Consider coding use cases, for instance. Taylor English’s Lindberg says it would be extremely valuable if companies were able to negotiate indemnity that allowed minor code edits while retaining copyright coverage, given that the current terms only cover unmodified code. When asked if this was feasible, Microsoft did not directly answer the question, while AWS and OpenAI declined to answer. Google notes that if users modify the output, its indemnity “doesn’t automatically disqualify them, as Google will still cover any claims related to the unmodified output”.

There are lots of lawsuits pending, and many of these cases are fact-dependent. So, even if there is a decision in one case, you are not going to get judicial decisions that provide clear guidance

James Gatto, Sheppard Mullin

Companies can also employ techniques internally to mitigate risks, one of which involves subjecting outputs to multiple rigorous levels of review. This practice, applicable to both non-code and coding use cases, can use human expertise, automated toolkits, or ideally a combination of both, to examine if the outputs infringe copyright.

For code generation, the most practical and easiest way is to do source code scanning of AI-generated code to see if it matches any known code, says Lindberg. If it does, developers then need to determine whether to remove the code or comply with the licensing requirements.

Besides, Lindberg notes that test-driven development (TDD), a software development method where developers write tests before they write the actual code, is also useful for managing the risks. By first writing tests, developers can define clear expectations for what the code should and shouldn’t do, including adherence to IP constraints. In addition, automated tests, a core component of TDD, are run frequently, typically every time new code is integrated into the codebase or when existing code is modified. This ongoing testing ensures that every change is immediately checked for compliance and can help catch any new potential breaches of IP agreements.

Put a number on it

Fox Rothschild’s Neelbauer says good corporate lawyers should be comfortable with helping their clients quantify risks so that businesses can make decisions based on numbers, rather than on gut feelings.

Specifically, companies and their legal teams should work together to assess and compare the costs of various options. For example, the risk cost of potential lawsuits as a result of using vendor tools versus the cost of not adopting these third-party tools, and instead developing customized internal models trained using proprietary data. Notably, Neelbauer emphasizes, they should allocate “the highest dollar amount for the risks”, otherwise the risk valuation can be insufficient to cover the actual risks.

As Microsoft’s Tsilas said in March, the debate on the ownership of AI-generated output won’t be going away anytime soon. And not until the laws catch up and provide answers.

Legal experts say they anticipate enhanced regulatory clarity in the next five years – particularly as the White House signed a long-awaited executive order on the technology in 2023. This year, the legislative branch is also showing significant momentum by introducing various bills across different states and at the federal level.

Regarding US IP law specifically, Neelbauer believes it already has “great bones”, and that it’s just a matter of “how those bones apply”.

McCoy Smith, founding attorney of law firm Lex Pan Law, thinks the risks will subside as AI models become more sophisticated and evolve to train themselves with greater independence from external public datasets.

Until then, users should continue to tread carefully.