Overview
Generative artificial intelligence (AI) systems are finding use in a variety of applications, including software development. A major downside to these systems is that the current popular implementations are all cloud services, meaning that a third-party company can see all the queries made into the system and all associated data used with AI-enabled tools. As a result, these companies are using AI to harvest your private data both to sell to data brokers and to train the AI models themselves. Several of the large AI companies have made no secret of using AI to serve targeted ads based on user data. The European Union isn't helping matters by requiring AI companies to watermark AI-generated material, since the companies can use that requirement as an excuse to add a traceable identifier to generated content that can lead back to the person who used AI to create it.
While one way to avoid this surveillance is to forego the use of AI entirely, a more practical solution may lie in open-weight large language models (LLMs) that can be run on a local computer. By running the AI algorithm on hardware that you control, it is theoretically possible to use AI technologies privately. However, this possibility is still only theoretical at the moment, since it isn't entirely clear that the readily available local AI tools are not introducing their own privacy problems and/or making calls to third-party services.
The objective of this research project is to evaluate open-source tools for running local AI models to determine the extent to which such tools are actually private. Some of these applications are capable of accessing the Internet, which means they're also capable of exfiltrating data. Similarly, there is no guarantee that these applications won't generate and embed some sort of unique fingerprint into the output.
An Incomplete List of Tools
Some of the local AI applications to evaluate may include, but are not limited to:
Related Research and Development
If off-the-shelf, open-source applications are found to have suitable privacy properties, then these applications could become useful in a Private Desktop Environment. It is possible that an off-the-shelf application can provide privacy as long as it is prevented from accessing the Internet. In this case, a bubblewrap container could be used to run the AI model in a network-isolated namespace.
Alternatively, it might be a more productive project to develop a local desktop AI application that uses the open weights to run the models directly. This approach would require more work to implement but would provide finer-grained controls to ensure that a locally run model could not exfiltrate data or launch an unanticipated attack on some other system.