In the dynamic realm of generative AI, the question of who bears the responsibility for intellectual property (IP) issues arising from AI-generated content is a contentious one. While some generative AI vendors vow to support their customers in the face of IP lawsuits, others adopt policies that leave customers to fend for themselves.
Generative AI, encompassing text, image, and music generation, has the potential to inadvertently infringe upon copyrights. However, the stance taken by companies in the AI market varies significantly.
Certain vendors have committed to defending their customers, both financially and otherwise, when confronted with copyright disputes stemming from their generative AI tools. Meanwhile, others have adopted policies that insulate themselves from liability, shifting the burden of legal costs onto their customers.
Although the terms of service agreements for most generative AI tools are publicly available, they are often couched in legal jargon. Seeking clarity, we reached out to these vendors to understand their policies regarding customer protection in the event of copyright violations facilitated by AI-generated content.
The responses, and in some cases, the lack thereof, shed light on the industry’s approach to this pressing issue.
Generative AI’s Learning Process
Generative AI models learn by analyzing vast amounts of data to create content like essays, code, art, and music. This data often includes publicly available material. Some of the data is in the public domain, while other portions are protected by restrictive licenses, demanding attribution or compensation.
The legality of using data without permission is a contentious topic, currently being debated in courts. What may land generative AI users in legal trouble is the act of regurgitation, wherein AI models produce verbatim copies of training examples.
Notably, Microsoft, GitHub, and OpenAI are currently facing a class-action lawsuit accusing them of violating copyright law due to their AI, Copilot, regurgitating licensed code snippets without giving credit. Authors in California and New York have also sued OpenAI for alleged IP theft. Similarly, image-generating tool vendors like Stability AI and Midjourney are entangled in lawsuits brought by artists and stock image sites.
A Growing Concern
The potential for copyright violations with generative AI tools has made intellectual property a significant concern, with nearly a third of Fortune 500 companies expressing apprehension about the technology, according to a survey by Acrolinx.
Despite these concerns, investors continue to pour billions into generative AI startups. However, the sustainability of this trend remains uncertain.
The Question of Indemnity
Given the uncertainty surrounding generative AI, one might expect vendors to unequivocally support their customers, if only to allay concerns about IP-related legal challenges. However, this is not always the case.
Some terms of service agreements, particularly the indemnity clauses that specify when customers can expect reimbursement for damages related to third-party claims, reveal that not every vendor is willing to risk a legal decision that could jeopardize their approach to AI model training or their business model.
For instance, Anthropic, which recently secured substantial investments, reserves the right to “hold harmless” itself and its partners from damages related to the use of its generative AI, including those tied to IP issues. When asked directly if Anthropic would support a customer facing a copyright lawsuit, the company declined to comment.
AI21 Labs, another well-funded generative AI startup, also declined to provide a clear answer. However, their policy suggests that they might assume control of a lawsuit against a customer but will not cover the associated expenses, which would be borne by the customer.
OpenAI, one of the most prominent generative AI vendors, limits its liability in its terms of use. In many cases, it won’t be a party to or defend against copyright lawsuits targeting its users. This policy is not uncommon among text-based generative AI vendors.
Even vendors in image and video generation, where copyright violations are more apparent, offer limited contractual support. For example, Stability AI leaves customers responsible for defending themselves against copyright claims and offers no payout carve-out for potential liability. Similar policies apply to Midjourney and Runway.ai.
A Mixed Landscape
Despite the prevailing ambiguity, some vendors, perhaps in response to enterprise customer concerns, are willing to commit to protecting customers facing copyright infringement lawsuits, to some extent.
Amazon, with its Bedrock platform, promises to indemnify customers against IP claims related to its in-house text-analyzing models and code-generating service, CodeWhisperer. However, this indemnification is subject to conditions such as a professional subscription and allowing AWS to control the defense.
IBM also provides IP indemnity for its generative AI models through its Watsonx service, with no cap on indemnification liability.
Google offers some defense for customers facing third-party allegations of IP infringement but may suspend the use of the allegedly infringing model if no suitable remedies are found. Google-backed Cohere also pledges to “defend, indemnify, and hold harmless” customers against IP claims.
Microsoft recently made headlines by announcing that it would cover legal damages for customers using its AI products if they face copyright infringement lawsuits, provided they use specified products with built-in content filters.
However, the scope of Microsoft’s indemnity policy is nuanced. It applies to paid versions of certain AI services but excludes “preview” versions. IP lawyers have raised concerns about the clarity and exclusions in Microsoft’s indemnity provisions.
Adobe offers “full indemnity” protection for users of its generative AI art platform, Firefly, but this protection is limited to enterprise customers and comes with liability caps.
Shutterstock and Getty Images also provide indemnity to their enterprise clients, who use models trained on licensed images.
The Road Ahead
As generative AI vendors seek to attract enterprise customers under investor pressure, indemnification protections may become more prevalent. However, these policies are unlikely to be uniform and may have exceptions that limit their effectiveness.
Indemnities are not a panacea for avoiding the risk of third-party infringement claims. It is essential for customers considering generative AI tools to carefully evaluate these indemnities when deciding which provider to choose for their projects.
