AI safety restrictions

[Daniel-Olson]

Background: At the Critical Path Institute Data Collaboration Center's face-to-face meeting last week, we discussed considerations with using generative AI or LLMs with our datasets. In the conversation, we noted that as part of data classification, we could mark datasets as safe for use, and this could be added to the Data Use Ontology (DUO) as a controlled vocabulary term.

DUO is used for data classification at Critical Path. Hence, we would like this new term to be added to DUO, for use with our datasets.

Proposed solutions: Here are two potential terms that may cover the AI safety concerns:

1. A broad term may be something like “No Generative Artificial Intelligence Restrictions” as a child of “Data Use Permissions”.
2. Alternatively, we could flag individual datasets with “Artificial Intelligence Specific Restrictions” as a child of “Data Use modifier” if we have any safety concerns at all about a given dataset.

[ddooley]

Note that if further distinctions are needed about the kind of AI involved, this ontology might have pertinent terms. https://github.com/berkeleybop/artificial-intelligence-ontology