At Georgia Tech, we do not currently have an official institute-wide policy about the release of open source AI models, as the code for these models is typically covered by standard open source licenses. Note that the Office of Information Technology (OIT) has provided guidance on the usage of generative AI models, especially as it relates to generating GT-branded content and for its usage with Personally Identifiable Information (PII) on Georgia Tech systems. Please review the following two links for more details on OIT’s guidelines.
The Open Source AI Definition
Note this definition is an external effort to Georgia Tech, the official endorsement on supporting this effort is at OSPO@GT Linkedin Page.
The Open Source Initiative is the primary maintainer of the definition of open source software and arbiter of what makes a software license open source. Recent advancements in open source Large Language Models (LLMs) have led to a community-driven effort to define what is meant when we say an Artificial Intelligence system is “open”. OSI has released a draft definition of an Open Source AI definition, which has resulted from monthly meetings with open source developers from around the world.
Importantly, this definition covers the model, parameters, and training data. The OSI definition defines four different categories of input training data used to train an open AI system, from open (fully reshareable) to unshareable, non-public training data. These different “levels” of open training data create some grey areas for interpreting whether a model is truly open, but they also reflect the reality that some training data may be sensitive (e.g., personal health information). The OSI definition attempts to further define the impacts of these classes of unshareable data by noting that open AI systems must still describe the data types used for training and any biases that may be implicit in the data set.
Learn more at the resources below and consult with the OSPO on any questions you may have related to the usage or release of open AI systems, especially if you are considering the release of an open AI system.