Generative AI raises five sets of issues not often found in IT contracts. So do other machine learning (ML) systems, particularly those trained on massive datasets. The issues are complex, but no amount of analysis will help if you don’t notice them. So I’ve listed them below.

For more information, see The Tech Contracts Master Class™. It addresses the five, along with the other key terms of IT contracts. It teaches businesspeople and lawyers to draft and negotiate contracts about SaaS, AI, and other forms of software.

Many of the five below can lead to less advantageous terms than customers usually get in software contracts. But that doesn’t mean they shouldn’t push for what they can.

1. Machine learning often reuses prompts and training data from the customer – for other clients. If so, providers need a license to that data. Customers, on the other hand, should seek confirmation of their IP rights or at least confirmation that ownership doesn’t transfer to the provider. Both parties should also address control of that data – who can do what with it – since IP rights don’t really restrict use of data.
2. Typical NDA terms won’t work if machine learning systems reuse customer prompts or training data – and won’t fully protect outputs.  Providers should try to avoid confidentiality. Customers should at least try to restrict what provider staff does with their data and with outputs, even if the AI itself can reuse it. Customers should also tell their own staff to treat all generative AI outputs as confidential until someone figures out whether it includes trade secrets. No one knows in advance whether generative outputs will be sensitive.
3. Generative/ML outputs are error-prone and so can injure customers. That probably can’t be helped, so providers should consider broad disclaimers. Customers, on the other hand, should seek promises about inaccuracies (including defamation and discrimination) but recognize the provider may not be able to offer much. But customrs should resist total disclaimers and seek warranties and SLAs covering, at least, how the artificial intelligence functions, even if they don’t cover outputs.
4. Generative/ML outputs come with unusual risks of 3rd party suits against the customer re defamation, discrimination, IP infringement, and more. Customers should always seek indemnities but, again, recognize that the provider may have no control over machine learning outputs. At a minimum, go after IP indemnities re the software itself, if not the outputs. Providers, of course, should resist indemnities.
5. AI outputs could include the provider’s own data, as well as third party data, particularly for systems trained on large swaths of data. So providers should avoid assignment of anything preexisting (like the typical exceptions to an NDA) – or even better, avoid assignment altogether. It might be enough to agree that generation of outputs doesn’t create provider IP, without transferring any. Customers should push for IP in outputs but recognize that restrictions on control – again, who can do what with outputs – may matter more.

© 2023 by Tech Contracts Academy, LLC. All rights reserved.