Module 7: Governing AI Deployment · BoK IV.B
Deploying a proprietary model
Developing AND deploying your own model creates a dual role with heightened liability from both providing and using the technology. It brings five challenges (transparency limits, data sources, ownership, risky purposes, increased requirements) but four opportunities, chiefly sourcing your own training data for better transparency.
Developing AND deploying your own model means a dual role → heightened responsibilities and liability from both providing the technology and using it. But there are upsides.
- Transparency → proprietary datasets limit how open the organisation can be with documentation for procurers, oversight bodies or the public.
- Data sources → if training data was scraped from the web and includes copyrighted material, who owns it and who may use it?
- Ownership → when a user creates a new work with the model, who owns the new work? Set clear guidelines in acquisition frameworks, contracts, terms of service and agreements.
- Risky purposes → if the algorithm could serve a risky purpose, limit others' potential to use it that way via user agreements, contractual terms and regulation.
- Requirements → secretive development can mean different or increased requirements, e.g., a data breach may trigger different duties for a proprietary vs non-proprietary model.
- Source the training data yourself → better understanding of the data and its origin → better transparency.
- Better governance reporting → easier to meet regulatory requirements given model ownership and build.
- Less exposure to security and other problems of open-source and third-party models.
- Purpose fit → built with the exact need in mind instead of retraining or tailoring someone else's model.
Key terms - quick answers
What is “Proprietary model”?
A model an organisation both develops and deploys, creating a dual provider/user role and liability.