We know it when we see it

Explainable AI (XAI) is the field of making AI systems that we can understand and explain. Fairly straight forward - it does what it says on the tin.

But our best definitions for "Explainable" boil down to "we know it when we see it". This seems rather poorly specified for something within regulation such as GDPR with the right to explanation.

Context is key

One of the issues is that what you and I understand is heavily influenced by our backgrounds, our skill sets, and everything else we bring to the room.

Communication is one of the hardest challenges, be it within your team, in the board room, over the dining room table, or via a blog. How we say things really depends on when and why we are communicating to people.

So I define explainability in terms of that subjective context - the audience, the purpose and the language determines the framework for determining if something is explainable.

Right now, I'm explaining explainability to interest you, the reader, through the medium of a blog post. Success for me is if you can remember this definition and use it in your own work. I can measure that based on your feedback to this post.

Why this changes XAI research

A lot of XAI work is published in the context of academia. It is by academics, for academics, in scientific papers, in order to progress academic careers. This means that many of the explainability metrics are great for other AI researchers but fall flat when put into a board room or in front of auditors.

Instead, we need to be thinking about XAI research as an interdisciplinary field, where we test our explanations with the audiences that really matter. We motivate XAI with health care, legal work, and other sensitive areas. So the XAI work can only be evaluated by the stakeholders in those fields.

Honestly, it doesn't matter if I understand the predictions of my own models. If you don't trust them, you won't use them, and all the work will create no value. So the onus is upon me, and others in the field, to ensure that you understand what is being said.

Communication is about what is heard, not about what is said.

‍

If academic papers are more your thing, you can download my paper which covers this in far more detail