A Chatbot? Are you Sirious?

Since blogging that I Need an AI BS-Meter a number of people have sent me pointers to a subset of AI I loosely think of as Result Explainers -- everything from pending government regulations (EU's Global Data Protection Regulations -- GDPR) to the latest in academic research (Local Interpretable Model-agnostic Explanations -- LIME). As the authors of the EU's GDPR state, widespread adoption of AI cannot occur until vendors are able to communicate results in a "concise, intelligible and easily accessible form, using clear and plain language." This got me thinking, "What should Result Explainers look like?" Should they generate trust scores, a series of Google-Maps like directions that get you from data to results, a series of diagrams? And as my colleague Patrick at Lab41 has pointed out, "Why should we trust a Result Explainer if we don't trust AI to begin with? As you might expect there isn't one right answer. That said, recent advances in recommenders, digital assistants, user interface design and initiatives like DARPA's recently announced Explainable Artificial Intelligence (XAI) grand challenge suggest we may be on the brink of a few breakthroughs. Again, as the authors of the EU's General Data Protection Regulations note, while the resulting classifiers, models, predictors, etc. can be very powerful they also frequently confound explanation -- e.g., the output of SVMs and Gaussian processes can be difficult to render, ensemble methods hide information as a result of aggregation and averaging, neural nets create high data dimensionality, and so on. End users care a lot more about results than they do about models. Unfortunately assessing result quality takes us right back to the models, as nonparametric models are only as good as the data used to train them (along with the type of model structure and associated parameters that were selected). But these models frequently hide information. Part of the magic of AI is that it finds stuff based on features that previously may not have been well understood. Unfortunately, the features models train on are frequently unclear. Assigning labels to pre-trained models can help mitigate some of this ambiguity -- e.g., "This model was trained with over 100,000 high-res color images of cats." These labels may be misleading though, as the model may contain feature biases that are not well understood -- e.g., "the training data is dominated by images of "well-fed, indoor cats from Japan."