Picture a startup creating a new AI tool that enables its customers to summarize internal business documents. The startup accesses a third-party AI model and builds its own software around it. 

The provider's first customer is ready to sign, but before they do, they have some important questions: 

Will the documents the customer uploads remain confidential? Will either the startup or the third-party model provider use the data they upload to train models available to other companies? 

The provider's AI terms of service vaguely described authorizing data usage for “service improvement”. The sales team told the customer that its data and information would stay private on the provider's end, but what about the company who created the original model that the provider is licensing? 

A gap between what the sales team promises and what the AI vendor agreement actually allows can jeopardize a deal. 

Whether you're building an AI model from scratch or licensing an existing one from a third party, AI licensing agreements should accurately describe data practices and reflect the promises made during the sale. 

Understanding the key AI licensing terms involved in an agreement and how they differ from standard software licensing can help ensure that providers deliver on customer expectations. 

There are many nuances to these terms that vary on the use case and business reality of each provider and customer. While this article is designed to offer an introduction to some of the new terms that apply specifically to the AI market, working with a lawyer well versed in these issues is strongly recommended.

How AI Changes Standard Software Licensing Agreements

Companies have used standard software licensing and software as a service (SaaS) agreements for years to give other businesses and customers the legal permission to use their software, while they retain the legal ownership of it. 

Traditional software licensing agreements have key terms around who can use the software, their purposes, and for how long. Agreements also address payment, performance commitments, indemnification, security, liability, and termination.

AI licensing agreements feature many of the same provisions. However, AI functionality requires the parties to address new questions such as: 

 

• Can customer data be used to train a model? 

• What rights does the customer receive around the usage of AI outputs? 

• Where does responsibility fall when an output is inaccurate or triggers a third-party claim?

 

Traditional software licensing terms provide a solid foundation, but providers need to adapt them to reflect how AI technology processes information, generates output, and uses third-party AI models. 

Before signing an agreement with a customer, providers of AI-powered products or software should review their licensing terms and adapt standard software provisions to accommodate how AI products operate differently than traditional software technology.

The most important terms and provisions that providers need to understand before they enter an AI licensing agreement typically fall into a few different categories. 

License Scope and Acceptable Use

An AI software licensing agreement must address the scope and intended use of the license grants. Whether the agreement grants a license to install software or access to a hosted AI service, it should clearly define how the product may actually be used. 

For example, an agreement may limit the customer to internal use for a particular purpose, or it can permit them to use it for general internal purposes without restriction. It can limit usage to a certain number of users or it can charge based on consumption of data. 

The variety of limits are limited only by the creativity of the vendor. Regardless, the limits on usage need to be made clear in the agreement and understood by both the vendor and the customer. 

Providers using third-party AI software or models must also be aware of their own scope and acceptable usage from their technology suppliers. You cannot promise your customers usage rights that you don't have yourself. 

Technology, Customer Data, and AI Output Ownership

An AI licensing agreement should clearly outline who has what rights to the software and AI technology, the raw data that the customer feeds into the AI, anonymized and aggregated data, and AI outputs. The agreement should address each category separately, as the market does.

Provider AI Technology and Customization

To have an ongoing business, the provider retains ownership of the software, interfaces, documentation, proprietary models, and other technology it developed. If the service uses any third-party models or components, these rights belong to their respective owners and any upstream agreements.

The customer receives permission to use your service according to the terms of the agreement during the term.  

If you as the provider build a custom integration for the customer, fine-tune a model, or create another type of custom feature, your agreement should establish who owns it, and whether you can reuse those customizations for other customers.  

Customers may ask to own these items but vendors should insist these rights should be limited to their use in connection with the service provided. While the customer may be paying for this work, it is typically not core to their business and really only has value in connection with the use of the provider's service. 

In contrast these items are core to the provider's ability to continue to provide the service to other customers.  

Customer Data and Training Rights

One area where an AI license agreement is materially different from a standard software or SaaS license is how customer data is handled. 

Between the parties, the customer licensing the AI powered technology will typically retain ownership of the data that it inputs into the system. The agreement then grants the provider a limited license to process those inputs as needed to operate, maintain, and support the service.

However, the license does not give a provider ownership of customer data, nor does it allow them to use that data in any possible way. The customer may be required to confirm that it has the rights and permissions to use the data they submit.

Training rights are a separate issue. If a provider is given the right to process customer data, they don't automatically have the right to use that data for a second purpose, such as using that data to train/improve their models for other customers. Providers will often try to get these rights by asking for the ability to use customer data to ‘improve' the product. Customers often only allow providers to use their data in an anonymized and aggregated form for this purpose.

The agreement should clearly state whether the customer's content or data can be used to train or improve a model specific to the customer, a model shared with other customers, or neither. If a third-party model processes the customer's content, the provider must be sure that its promises to customers match the third party's terms and actual data practices. Customers may object to use of their raw data to train AI models but often agree to the use of anonymized and aggregated data.  

AI Output Ownership

An agreement should also address AI output ownership. An agreement can state that between the provider and the customer, the customer owns the output of an AI product, such as code, images, summaries, and other material generated by the product. 

However, an AI vendor agreement cannot promise that output from an AI system can qualify for copyright protection. That is because the U.S. Copyright Office maintains that output generated entirely by AI is unable to be copyrighted. Human authored elements or modifications may qualify for protection. 

AI output ownership provisions also do not guarantee that an AI output will be unique or free from third-party claims of infringement. The customer may “own the output”, but that doesn't necessarily mean that they have the ability to use it commercially without legal risk. Those are separate matters. 

Confidentiality and Data Protection

Confidentiality and data protection provisions define how the provider can use and disclose customer information, including how information is handled by employees, contractors, hosting companies or third-party AI providers. 

Data protection terms should also address security measures, incident responses, retention and deletion, and what happens to data after the agreement ends.

Warranties, Indemnification, and Limitations of Liability

Due to the probabilistic nature of AI technology, AI outputs can be inaccurate, incomplete, or infringe on protected materials. Thus, AI terms of service need to include language that addresses these risks for both the provider and the customer. 

AI agreements should define the provider's warranties and disclaimers, including what it promises about the product's performance. Providers can't realistically promise that every output will be accurate or legally compliant. Providers will regularly require customers to validate output before using it – particularly in sensitive areas of operation like health care applications.

Indemnification clauses describe who must defend and cover third-party claims. The provider may agree to cover defined intellectual property claims involving its technology and the customer's authorized use of the AI product. The customer may be responsible for claims arising from its inputs, unlawful conduct, output, or for prohibited uses. The provider's indemnity often excludes claims arising from modifications made by the customer or misuse. 

The limitation of liability provision caps total damages one party may recover from the other. The overall cap typically bears a reasonable relationship to fees paid during a certain period. In some cases, obligations like indemnification, data security claims, and confidentiality may be treated separately from limits of liability and have higher limits or “super caps”. 

Providers that license a third-party model for their AI product should ensure that the protections they offer their customers align with the protections received from their own providers. This can avoid agreeing to cover a claim that your own provider doesn't cover. 

Fees, Service Changes, and Termination

An AI software licensing agreement should also cover the financial and commercial relationship between the provider and customer. 

Fees provisions define how usage of the AI product or software is measured and billed. It can specify the allowances or credits the customer receives, the rates charged for overages, and explain how and when pricing can change.

Service changes address model substitutions, discontinued features, notice requirements, and outline the customer's options if a change materially reduces functionality. 

AI terms of service should identify when the provider may suspend access, when either party may terminate the engagement, what notice is required, and whether the breaching party has an opportunity to resolve the problem. 

Finally, the agreement should also address payment of any outstanding fees, refunds, early termination charges, data export and deletion, and the customer's right to continue using previously generated outputs. 

The agreement should explain what happens while the service is active, when the product changes, and when the relationship ends. 

Need Help Drafting an AI Licensing Agreement? Contact SVTech

AI licensing agreements should reflect how the product actually works. That means defining what customers can do with the AI product, how their data will be handled, and what rights they receive in AI outputs. 

However, these issues are still emerging. While the AI market is beginning to settle, it remains in flux and highly nuanced.

SVTech helps AI companies and software providers draft and negotiate AI agreements that protect their technology, allocate risks, and build strong relationships.

Contact SVTech today to talk about licensing your AI product or software.