person using Generative AI (models that generate text, images, music, etc. when given a prompt) can infringe on the copyrights of others through no fault of their own. But if, or rather when, that happens, who will be liable for legal costs and damages?
it depends.
In the rapidly changing landscape of generative AI, companies monetizing technology, from startups to big tech companies like Google, Amazon, and Microsoft, are approaching intellectual property risk from very different angles. .
Some vendors have promised to use their generative AI tools to financially and otherwise defend customers who find themselves on the wrong side of copyright lawsuits. Some companies have published policies that shield them from liability by making their customers pay legal costs.
The terms of service for most generative AI tools are public, but written in legalese. For clarification, I asked vendors about their policies to protect customers who may infringe copyrights on AI-generated text, images, videos, and music.
The responses and non-responses were enlightening.
backflow of data
Generative AI models can “learn” from examples to write essays and code, create artwork, compose music, and even write lyrics to accompany that music. They are trained with millions to billions of e-books, artwork, emails, songs, audio clips, voice recordings, etc., most of which come from public websites.
Some of these examples are in the public domain, at least for vendors trawling the web for training data. Some do not, and others are under restrictive licenses that require citations or certain forms of compensation.
The legality of vendors using data for training without permission is another issue being challenged in court.But what could generative AI be about? user The problem arises when regurgitating, or when the generative model spits out a mirror copy of the training example.
At the top, Stable Diffusion, an image generation AI, displays an image generated from random captions in the model’s training set. At the bottom you will see the image you are asked to match with the original. Image credits: Somepari et al.
Microsoft, GitHub, and OpenAI are currently sued in class action motion It claims that Copilot, a code-generating AI, violates copyright law by regurgitating licensed code snippets without providing credit. Elsewhere, thousands of writers have signed an open letter condemning generative AI technologies that “imitate and regurgitate” their “language, stories, styles and ideas.”
Cases keep coming in.
Authors from California and New York sued OpenAI is suspected of stealing intellectual property from its works. Image generation tool vendors such as Stability AI and Midjourney have been the subject of lawsuits brought by artists and stock image sites like Getty Images. And Universal Music Group’s goal is to Ban The AI-generated music mimics the styles of musicians represented on streaming platforms and sends takedown notices to remove songs.
Probably it’s no surpriseSecond, in a recent survey of Fortune 500 companies by Acrolinx, nearly one-third said intellectual property was their biggest concern regarding the use of generative AI.
The threat that generative AI tools will violate copyright hasn’t stopped investors from pouring billions into startups developing such tools. But some may wonder whether this situation will continue for much longer.
issue of compensation
In the midst of uncertainty, generative AI vendors may think they will support their customers with the strongest terms, if for no other reason than to allay customers’ concerns about intellectual property-related legal issues.
But that’s wrong.
The wording of some clauses in service contracts, specifically indemnification clauses, or clauses specifying when a customer can expect to recover damages from a third-party claim, makes it mandatory for all vendors to It is clear that there is no desire for the court to issue a judgment that is You need to rethink your approach to training generative models, or worse, your business model.
For example, Anthropic recently signed a deal with Amazon to raise up to $4 billion. reportedly Seeking an additional $2 billion in investment from Google and others, reserves The right to “hold harmless” ourselves and our partners from harm arising from the use of generative AI, including those related to intellectual property.
Point blank, I asked Anthropic, a company that provides strictly text generation models, whether they would legally or financially support customers who were involved in copyright lawsuits over the output of their models. The company declined to comment.
AI21 Labs, another well-funded generative AI startup that is developing a suite of text editing tools, also declined to respond.so i saw it policy.
A21 Labs says the company may “assume exclusive defense and control” of claims against its customers if the customers elect not to defend or settle the claims themselves. But that doesn’t compensate for that privilege. This will be the responsibility of the customer.
OpenAI — arguably the most successful generative AI vendor today, with over $10 billion in venture capital and approaching $1 billion in revenue — told me that terms of servicethe Company’s liability is limited to the Amount. [a customer] paid for [an OpenAI] The service that created [a] or $100 for claims made within the 12 months prior to liability arising. ” This is the best-case scenario for the customer. OpenAI’s policy makes clear that the company does not become a party to or defend copyright lawsuits directed against users in many, if not most, cases.
Vendors that build the AI that generates images and videos tend to be a little more transparent about potential copyright violations, but they’re also less cooperative on the contractual front compared to their text-first rivals.
In addition to models that generate images and text, Stability AI also develops models that generate music; mention I agree to the API terms. The company leaves it up to its customers to protect themselves from copyright infringement claims, and unlike some other generative AI vendors, there are no payout carve-outs if found liable.
Midjourney and Runway.ai didn’t respond to my emails, but I found the terms.mid journey policy Releases Company from liability for third party IP damages. Runway.ai Equally so.
fine print
Some vendors are now becoming more sensitive to the concerns of enterprise customers who are considering adopting generative AI or seeking to position themselves as a more “safe” alternative. We will not hesitate to protect our customers in the event of a potential lawsuit. For copyright infringement. Up to a certain point.
Amazon, which recently launched a platform for running and fine-tuning generative AI models called Bedrock, will indemnify customers against claims that the models infringe on third-party intellectual property rights. (in other words, to defend). However, Amazon’s indemnification policy only applies to Titan, the company’s in-house family of text analysis models, and Amazon’s code generation service, CodeWhisperer.
CodeWhisperer’s coverage is broader and applies to all IP claims involving trademarks. However, you need at least a CodeWhisperer Professional subscription with the copyright protection filtering feature enabled. Free users of CodeWhisper are not given the same protection. The customer must also agree to let her AWS control the defense and resolve it “as AWS deems appropriate.”
IBM also provides IP coverage for Slate and Granite, generative AI models available through Watsonx generative AI services.
“In accordance with IBM’s approach to indemnification obligations, IBM does not cap indemnification liability for models developed by IBM,” an IBM spokesperson told TechCrunch in an email. “This still applies today. [and] Future Watsonx models developed by IBM. ”
Google didn’t respond to my email.However, from the company side clauseit appears that Google offers Several Defending customers against third-party claims of intellectual property infringement arising from text and image generation models. However, Google says it may suspend customers’ use of the allegedly infringing model if a “commercially reasonable” remedy is not found.
Google-backed Cohere also includes: regulations Its terms suggest that Cohere will “defend, indemnify and hold harmless” customers facing claims from third parties alleging that Cohere’s models infringe on their intellectual property. This is not surprising given Cohere’s enterprise focus.
Microsoft recently announced that if a customer using its AI products is sued for copyright infringement, it will act on behalf of that customer as long as the customer uses the “guardrails and content filters” built into its products. made a flashy announcement that it would pay legal damages.
Which products does it involve? That’s the tricky part.
Microsoft said its compensation policy covers paid versions of its AI-powered Copilot portfolio of services (including Microsoft 365 Copilot for Word, Excel, and PowerPoint) and the enterprise version of its chatbot on Bing. The attack reportedly targets a certain Bing Chat Enterprise. It also applies to his GitHub Copilot, Microsoft’s code generation service co-developed with OpenAI.
But in that deep blue policyMicrosoft has made it clear that customers using “previews” of generative AI capabilities powered by Azure OpenAI services are responsible for responding to third-party claims of copyright infringement.
Kate Downing, a Santa Cruz-based intellectual property lawyer, takes particular issue with the compensation provisions for co-pilots. are discussing Given the ambiguity of this provision and its exclusions, the initial costs of enforcement may be unacceptably high for businesses.
In contrast, Adobe claims to offer “full indemnity” protection to users of Firefly, its generative AI art platform, and that its models are trained on stock images to which Adobe already owns the rights. It is claimed that it has been done. However, the user must be a business customer and is subject to the same Adobe liability limits that apply to her other technology-based IP claims.
Adobe’s former rival Shutterstock also offers coverage to all business customers, a policy the company introduced late this summer. The same goes for Getty Images. (Getty Images and Shutterstock, like Adobe, use licensed images to train their models.)
road ahead
Indemnification protection is likely to become more common as generative AI vendors, particularly startups, face investor pressure to acquire enterprise customers. After all, those customers want assurance that they won’t be sued over alleged copyright infringement.
But if the current situation is any indication, the policies will not be similar. And some exceptions exist that make them more appealing in theory than in practice. In other words, it’s more of a marketing ploy than legitimate protection.
as Recent According to an article by British law firm Ferrer & Company, compensation does not provide a “get-out-of-jail-free card” nor is it a panacea.
“Our key message is: Do not assume that providing provider indemnification is the complete answer to the risk of third-party infringement claims,” the company wrote in a blog post. “Instead, weigh the offer of such compensation when deciding whether to use that provider’s generative AI tools for your project.”
Gen AI customers would do well to keep this in mind.