AI’s ability to teach itself borders on metaphysical, so harnessing the benefits of this phenomenon — by both providers and customers — requires thoughtful navigation.
Typically, a provider offers its proprietary algorithms, while the customer brings datasets and maybe a query to run. The AI platform, along with supporting tools, identifies patterns, makes predictions, and generates outputs. As usage increases, the model “heats up” — voluminous customer throughput turns out to be rich with information. Based on the customer’s inputs and repeat processing, the AI platform uses whatever it has learned to infer ever-nuanced responses and train and distill the model accordingly. The result is accretive to value all around — providers, customers, and their respective products and services all get smarter.
But who owns what?
Set aside for the moment the complexities of copyrights, third-party privacy, and other legal claims. For general coverage in these respects, each party may warrant that it has sufficient authority to work with the other (e.g., to offer the AI platform and to input data to the AI platform, respectively). Beyond that, a customer will typically look to have rights to whatever the AI platform delivers as output in response to the customer’s inputs. The provider, meanwhile, wishes to retain ownership of its models, both the baseline models and any improvements that may have resulted from interaction with the customer. From some providers’ perspectives, a reason to offer a learning platform is its ability to improve itself from inputs and be able to offer a better-honed platform for future users. While customers may have been attracted by this very capability, some of them may still push back on such ownership. Specifically, customers may seek rights to improvements that their data helped fashion. Some may even want to prevent the AI platform provider from sharing the improved AI platform with others, particularly the customer’s competitors.
Drawing bright lines may not showcase AI’s full potential. For example, the parties may agree that customer content will simply not be used to train provider models. Alternatively, the parties might assign rights based on the customer’s specific use case or what each party contributed. These solutions may sacrifice the full value of the provider’s solution and be impractical to implement.
In some such situations, a provider might consider creating separate instances of an AI model. Such an instance may be a copy of a model running in partitioned cloud environments (or on separate hardware), dedicated to an individual customer. This allows customers to train models using their own data and interaction. The provider may, in turn, aggregate learnings across its customer base — without accessing raw customer-specific data — to update the core model, which is then redistributed for the benefit of all. Through such a structure, each customer may gain rights to the improvements generated within its instance, while the provider retains the ability to offer its enhanced model universally.
It is important to anticipate these potentially competing interests. Broad-brushed contractual clauses purporting to assign rights between the parties are sometimes difficult to appreciate when presented theoretically. Parties may wish to “game out” the possible workflow and its consequences. They can then consider the feasibility of addressing who owns what to their mutual satisfaction.