Can AIs Learn to Explain Themselves?


The NewCo Daily: Today’s Top Stories

Jessica Mullen | Flickr

The new machine-learning techniques transforming how digital systems operate don’t work like old software programs. Once upon a time, programmers wrote code, and the code worked — or it didn’t, and the programmers debugged. But how do you debug a neural network? In Technology Review, Will Knight explores “the dark secret at the heart of AI”: We can’t really know why AIs do what they do. They are not programmed, but trained.

When a machine-learning program provides an answer, a decision, or a choice that its human operators believe is wrong, they can tell it so, and it will incorporate that data into its next choices. But most AIs can’t turn around and tell us how they reached a particular outcome. This “explainability” problem poses practical, legal, and moral questions we’re only beginning to scope out.

In Europe, for instance, governments are moving towards requiring companies to explain to users how their automated systems made decisions. Some critics are starting to argue that we have (as Knight puts it) a “fundamental legal right… to interrogate an AI system about how it reached its conclusions.” But this is a right we can’t enforce — we don’t have the tools.

We’ve never before built machines that operate in ways their creators don’t understand,” Knight writes. Now that we are, it looks like we’re going to have to teach them to explain themselves to us.

Amazon Isn’t All That’s Ailing Retail

The overall economy may be booming, but don’t tell that to brick-and-mortar retailers: They’ve had a disastrous two years, full of bankruptcies, store closures, and stock collapses (Derek Thompson in The Atlantic). All that has, in turn, hurt retail employment.

It’s easy to blame Amazon and the rise of e-commerce, and that is certainly one critical factor. Amazon’s sales have quintupled since 2010, mobile shopping works much better than it did a few years ago, and online research makes customers behave in a more directed manner, with fewer incidental purchases.

Thompson also points to two other factors. First: In the late 20th century the U.S. simply built way too many shopping malls. The Great Recession a decade ago exposed the rot and now these facilities are shuttering — and dragging down many of the retailers that invested in them. Second: Americans are spending less on clothes and more on eating out and traveling. Experiences are edging out dry goods.

What’s next? According to Thompson, look out for the way self-driving cars and shopping will combine into weird new phenomena — like the autonomous showroom vehicle or drugstore convenience van. In the future, the store may just come to you, on its own wheels.

For Science Marchers, Staying Nonpartisan Is Tough

The April 22 March for Science is avowedly “nonpartisan.” But the Trump administration is aggressively anti-science — it wants to cut funds, ignore research conclusions, and scuttle work to limit climate change — so that’s going to be a hard stance to maintain (Andrew Leonard in Nautilus).

The partisan-ization of science was an inevitable side-effect of the research community’s consensus on climate change. Democrats read the data and took up the cause of environmental regulation. Republicans are allergic to regulation of business and aligned with the fossil-fuel companies being blamed for warming. As Leonard writes, “Scientists who pointed out the deleterious environmental effects of specific economic policies were bound to get caught in the partisan crossfire.”

In this world, the size of a rally like the March for Science will matter. Even more important is what kinds of participation and activism marchers are inspired to take up after the event. That’s how the Tea Party rallies of 2009–10 won control of Congress. It could happen again.

The Gig Economy Goes Full-Time

The biggest platform companies in the on-demand or gig economy today still treat their workers as independent contractors, also known as “1099 employees,” after the label of the tax forms they receive. But there’s an alternative movement of startup platforms that give their workers the steady pay — and the respect — that comes with being treated as a real employee (Miranda Katz in Backchannel).

Sure, the companies that do this — like Managed by Q (office services), Honor (home care), Shyp (couriers), Luxe (valet parking), Eden (tech support), and Sprig (food delivery) — are spending more. They’re also getting a more reliable, engaged, and loyal workforce. And they’re avoiding potentially costly “misclassification” lawsuits.

Of course, there are companies — and workers — for whom the 1099-style relationship makes sense, and there ought to be room in the economy for both approaches. The firms that choose a traditional employer role will have a hard time achieving the kind of hypergrowth that put gig-based platforms like Uber on the map. But they’re creating real jobs, too, and that’s worth a lot.

The NewCo Daily will be taking tomorrow — Friday, April 14 — off. See you on Monday!

Leave a Reply