Libraries as AI Model Repositories

Libraries and librarians have a critical role to support and enable explainable artificial intelligence (XAI). Calls for transparency and oversight of the opaque and potentially biased machine learning and deep learning systems have increased as the extent and influence of AI has also increased.

However, transparency often conflicts with the quite reasonable desire for the protection of intellectual property. Hence many regulations (e.g. the EU’s GDPR) have carve outs for IP protection dramatically restricting some XAI options like public access to code or algorithmic models. Third party oversight of AI (such as auditing) is a widely promoted XAI strategy but concerns about secure access are problematic.

Lilian Edwards and Michael Veale (Enslaving the Algorithm: From a ‘Right to an Explanation’ to a ‘Right to Better Decisions’?) offer an interesting option: libraries as secure places for model repositories. The idea resembles how access to sensitive Statistics Canada data is provided via Research Data Centres in academic libraries. The result would allow qualified and certified third parties access to the AI models and training data under conditions that would protect IP.

Three observations from this:

  1. Since third party auditing is a highly favoured strategy for XAI, I’m supportive of anything that furthers this.
  2. This suggestion establishes a new role for libraries as AI repositories. These collections will require increased sophistication in terms of collection management. However, their near-term value as auditing resources and their longer-term value as research collections makes this new role important.
  3. This proposal explicitly identifies libraries as secure and trusted. It selects libraries over other possible sites for such an archive. Edwards and Veale believe libraries are (or can be) ultra secure and have the mandate and capacity to be so. We have a track record with Stats Canada, although this one may require more diligence.

Of course, this option would almost certainly come with additional funding since libraries would be part of a formalized auditing regime mandated through regulation or legislation. For the forward-looking library (or library consortium – I’m looking at you OCUL), this could be an important opportunity and the beginning of a research collection of unprecedented value.


This entry was posted in AI. Bookmark the permalink.

2 Responses to Libraries as AI Model Repositories

  1. John says:

    Interesting idea. I can’t imagine vendors trusting libraries to defend their intellectual property rights. I don’t see why libraries would think of providing hosting resources when everything is going to the (more secure) cloud. Vendors products can be subject to auditing, and accessed from a vendor secured cloud location – – no role for libraries that I can see.

    • Mike Ridley says:

      I think the concern is the impartiality of third party auditing and an appropriate, trusted player. Regulations like the EU’s GDPR are going to require model and performance auditing; removing or limiting vendor bias or interference will likely be a signficant issue. Maybe not libraries but I don’t think it can be vendors. [Also, good to hear from you. All the best]

Leave a Reply to Mike Ridley Cancel reply

Your email address will not be published. Required fields are marked *