Illustration: Lazaro Gamio/Axios

The first step to ethical artificial intelligence is teaching the computer to explain its decision making, something known in the field as explainable AI.

Why it matters: Right now many deep learning algorithms don't make it clear how they arrived at their predictions or conclusions. That lack of visibility into the data, steps and calculations that went into an outcome makes it hard to root out bias or other algorithmic errors that could impact results like who gets a loan or how much a factory should produce.

What's happening: Explainable AI, also sometimes called transparent AI, has become a top priority for nearly all the big companies in the AI field, including Microsoft, Google, Intel, IBM and Oracle. The topic is also expected to come up in Thursday's White House meeting on AI.

  • IBM's guidelines put it pretty simply: "Companies must be able to explain what went into their algorithm’s recommendations. If they can’t, then their systems shouldn’t be on the market."

That sounds straightforward, even obvious. But it actually isn't a feature built into many of the deep learning systems that are currently available.

No one size fits all: AI was a huge topic at Google's I/O developer conference this week, with some focus on explainability as well.

  • One very basic example is the new personalized scores in Google Maps for locations. In that case, Google's AI can also say why it thought that restaurant is a good fit by showing a few of the factors that led to the recommendation.
  • By contrast, doctors wanting to know why an AI system made a clinical diagnosis are going to want a far more detailed explanation.
  • "The kind of information you need in different scenarios really is fundamentally different," Google principal scientist Greg Corrado told Axios. "It's not possible to have one uniform standard for what constitutes explainability."

How it works: There are different ways for an AI system to explain itself.

  • One is to be able to show which variables led to different decisions and how heavily each was weighted. In the Google restaurant example, the program might explain its recommendation by noting a user tends to greatly prefer restaurants that are kid-friendly and often goes to pizza places.
  • In other instances, it could be enough to offer people the ability to adjust different pieces of data to see if that changes the conclusion.

What is explainable enough? Microsoft Research director Eric Horvitz likens the problem to getting a car fixed. You don't have to know exactly how the carburetor on a car works, as long as your mechanic does. That opens up the question of just how explainable AI needs to be for specific users and specific purposes.

"I think we need to do more research on what is a satisfying answer to a human being," he said.

It's not just Big Tech: DARPA, the Defense Department's advanced research arm, has a program on Explainable AI. The stakes are obviously huge when AI is helping guide decisions of who to attack, how and when.

"The DoD has to have, I would argue, a much higher bar ," DARPA's Brian Pierce told Axios.

Government's role: IBM hopes to raise the topic at today's White House AI summit.

“If the government’s going to do anything in terms of encouraging or even potentially regulating AI, the main focus has got to be on this issue of explainability," said Chris Padilla, who leads IBM's government affairs efforts.

What else: Explainability is a necessary ingredient for ethical AI, but it's really just a start. Other keys are eliminating bias, both in the data used to "train" the programs and in the algorithms themselves.

  • There is a separate question concerning what tasks should be reserved only for humans.
  • Many believe, for example, that allowing AI-powered autonomous weapons systems is a bad idea.

Go deeper: Here are several more looks at the need for (and means of creating) explainable AI.

Go deeper

Updated 23 mins ago - Politics & Policy

Coronavirus dashboard

Illustration: Sarah Grillo/Axios

  1. Global: Total confirmed cases as of 11:30 p.m. ET: 31,478,387 — Total deaths: 968,726 Total recoveries: 21,622,862Map.
  2. U.S.: Total confirmed cases as of 11:30 p.m. ET: 6,895,685 — Total deaths: 200,768 — Total recoveries: 2,646,959 — Total tests: 96,612,436Map.
  3. Health: The U.S. reaches 200,000 coronavirus deaths — The CDC's crumbling reputation — America turns against coronavirus vaccine.
  4. Politics: Elected officials are failing us on much-needed stimulus.
  5. Business: Two-thirds of business leaders think pandemic will lead to permanent changes — Fed chair warns economy will feel the weight of expired stimulus.
  6. Sports: NFL fines maskless coaches.

Trump pushes to expand ban against anti-racism training to federal contractors

Trump speaking at Moon Township, Penns., on Sept. 22. Photo: Mandel Ngan/AFP via Getty Images

President Trump announced late Tuesday that the White House attempt to halt federal agencies' anti-racism training would be expanded to block federal contractors from "promoting radical ideologies that divide Americans by race or sex."

Why it matters: The executive order appears to give the government the ability to cancel contracts if anti-racist or diversity trainings focused on sexual identity or gender are organized. The memo applies to executive departments and agencies, the U.S. military, federal contractors and federal grant recipients.

Dan Primack, author of Pro Rata
4 hours ago - Economy & Business

GoodRx prices IPO at $33 per share, valued at $12.7 billion

Illustration: Sarah Grillo/Axios

GoodRx, a price comparison app for prescription drugs at local pharmacies, on Tuesday night raised $1.14 billion in its IPO, Axios has learned.

By the numbers: GoodRx priced its shares at $33 a piece, above its $24-$28 per share offering range, which will give it an initial market cap of around $12.7 billion.

Get Axios AM in your inbox

Catch up on coronavirus stories and special reports, curated by Mike Allen everyday

Please enter a valid email.

Subscription failed
Thank you for subscribing!