Josh Clark on design principles for addressing flaws in machine learning. (via waxy.org)\

"The answer machines have an overconfidence problem. It’s not only a data-science problem that the algorithm returns bad conclusions. It’s a problem of presentation: the interface suggests that there’s one true answer, offering it up with a confidence that is unjustified.

So this is a design problem, too. The presentation fails to set appropriate expectations or context, and instead presents a bad answer with matter-of-fact assurance. As we learn to present machine-originated content, we face a very hard question: how might we add some productive humility to these interfaces to temper their overconfidence?

I have ideas."