A friend asked over lunch this week if I had read anything that helped me understand the real-world effects of artificial intelligence – not theoretical future scenarios, but concrete ways it could change the world now.

I suggested a long New Yorker story about Microsoft attaching AI-driven applications onto software used by millions of people. And I suggested that he listen to NPR's Morning Edition interview with Microsoft CEO Satya Nadella.

In the NPR interview, Nadella told me that 2023 "will be looked at the year we started using AI, it just sort of became part of our lives." Even as researchers and lawmakers debate potential risks of a technology that might grow beyond human control, Microsoft is on its way to deploying it almost everywhere.

Microsoft partnered with OpenAI, creator of the widely noted chatbot ChatGPT, and has begun rolling out AI features within its nearly ubiquitous software such as Microsoft Excel, Outlook and Word. Microsoft's gigantic scale — 220,000 employees in a company valued at $2.75 trillion — gives such moves impact.

Nadella spoke via video from Microsoft's wooded campus in Redmond, Washington, where he said he uses his own company's so-called "copilot" functions to compose "the first draft" of memos and emails.

Engineers developed the "copilot" concept while thinking of the limitations of the technology and the people who use it: users are urged to understand that AI will often get things wrong.

What follows is a portion of our conversation, edited for length and clarity.

SATYA NADELLA, CEO, Microsoft

The one big design decision we made was to think about this as a copilot, not as an autopilot. Designing it in such a way that the human is in control. Thinking of anything that gets generated as a first draft, whether it's a piece of code or a piece of text or an image, that's probably a good way for us to think. Humans have to do their job too.

STEVE INSKEEP, host, NPR's Morning Edition

I wanted to play with some of your technology, so I went on to Bing [Microsoft's search engine]. I asked it to help me prepare for this interview. And I said: Tell me a few things that I didn't know about Satya Nadella. And it gave me information almost in the form of a special Wikipedia article with a number of facts, some of which were true, but also said, "Nadella ... has actually published a book of poems called Two Shorten the Road."

No. That is a hallucination.

Is this what you want people to be aware to do? That what they get back may be wrong? That they need to use their own brains?

That is right. This is not a database. It's a reasoning engine that goes and looks at all the data and tries to help you search through it and reason over it. That's why I think you have to design it with humans, teach humans to correct mistakes. And the technology itself needs to get better.

Would you explain for the layman how you're preventing this technology from being used for ill purposes? Someone says: Give me the instructions of how to make a bomb or any number of other things.

What you have to do is start putting it out [in front of users] in progressive ways, for people to find all the issues ... and then support it by all the technology, like filtering out things you don't want it to say. Put classifiers, right? When somebody says, 'Give me instructions to make a bomb,' that's a query that you can't put into Bing chat and get a response because we filter those out.

I'm impressed by how there are all these invisible instructions that go out with my query. But I also think about the flip side of trying to control the technology in that way. You become the gatekeeper. Is this right? You're going to decide which information I can get or don't receive, which might cause people to begin questioning you and your power.

Yeah. I mean, as producers of technology, one is we have to take initial responsibility for what are the safety standards around our products. But I fully expect that this is a place where our democratic process — whether it's legislation or regulators — will have a thing or two to say about what exactly is safe deployment.

Are you prepared for the possibility that in the election year that's about to begin, you will be under pressure the way that social media have been in the past to allow certain information out or criticized for not allowing certain information out?

We are already in the search business. When you search the web, you get all kinds of results back. And there's a ranking algorithm today in search, and there's already people who have concerns about the transparency. So this is not complete new ground. It does have a next level of user experience. But yes, we will be subject to the same set of pressures, and subject ourselves more importantly to the standards of making sure that what we put in front of people is sort of verifiable and accurate and helpful.

I understand that a lot of people who work on AI do so with an almost religious fervor. They believe so strongly in the technology, and they want it out in the world as quickly as possible. Isn't there a risk that that attitude, commendable as it sounds, could lead us astray?

I think people who are working on this technology, quite frankly, have the right balance. We are not just talking about technology for technology's sake. We are talking about technology and its real world impact. Inflation-adjusted economic growth all around the world is basically very low. So I think we need a new factor of production. [AI is] compatible with creating more economic opportunity. Then the second aspect is what you said, which is how do we do it safely? We can't break things because, if we break things, one, you're not going to have a business.

In addition to driving economic growth, can this technology drive economic inequality, making more and more money for the people who control it the best while putting other people out of work?

There will be changes in jobs and there will be new jobs that will be created. I'll give you an example. Today, you can participate in the increasing digitization inside your company. You can build an application by just [speaking to a computer in English rather than coding]. If you're in health care, or you are in retail, as somebody who's in the frontline with domain expertise, [you can] essentially do IT class jobs that may have better wage support.

Do you anticipate a moment where there's a large language model and maybe it works with another AI program that can help it to see, and this device thinks it's smarter than you are and would do a better job than you do in running Microsoft?

You know, already the Microsoft copilot is helping me compose emails better, is able to help me take a meeting, remember things I said and others said better. I do think that it is helping me be a better person working at Microsoft.

But do you actually worry about that moment?

I don't worry about the moment of some technology replacing me. If anything, I want technology hopefully to remove the drudgery in work that all of us have.

Copyright 2024 NPR. To see more, visit https://www.npr.org.