Microsoft’s Brad Smith & Utah Gov. Cox: AI, Teaching, & Peace 9/9/26

0:00You spent millions deploying AI across your organization, but six months in, it still misses basic context. Sound familiar? That's not an AI problem, that's a data problem. Siloed systems leave your AI blind, while dark data confuses whatever it does see. EverPure unifies your data so your model can see it, understand it, and put it to work. No waiting around, no digging for answers. Get real results from your AI at everpuredata.com. When did work become so much work? The meeting about the meeting, the hundreds of files to find one insight, setting aside the things you want to do for the things that pop up.

0:39Your workday's gone. But what if the insight surfaced itself? Or you could ship the deck without the distractions. Gemini Enterprise helps you get that done. It's AI that knows your business with agents that take stuff off your plate. Make work less work with Gemini Enterprise from Google Cloud. Bring in show music, please. Hi, I'm CNBC producer Katie Kramer. Today on Squawk Pod. Is artificial intelligence out of control? Okay, so what? What do you do? Anthropic has a public defection over fears that AI can kill humans and soon. Which decade? This decade. This decade. It's 2026. already. Is it safe in the classroom?

1:28Teachers and tech are teaming up. Brad Smith, president of Microsoft, on why he's building an AI standard for school. Look, I don't think anybody wants a tech company to tell teachers how to teach. I mean, let's be serious. But they do want us, I believe, and we can and should and are, stepping up with the kinds of guardrails that ensure, for example, that student data isn't used to train an AI model. And teachers union leader Randy Weingarten says those guardrails are key. If you can, by the touch of your prompt, get an answer. You are not learning how to think. Then can't we all just get along?

2:07Utah Governor Spencer Cox with a warning for our communities. We've lost our healthy tribes. People are lonely. They're desperate for something. And so we're finding it in very unhealthy places. Politics has become religion for so many people. Plus, Joe Kernan with a Telluride Film Festival report. Three hours, you never really look away. There's times where you'd like to look away. The movie the business world can't wait to see. There's a lot of cringe. It's Wednesday, September 9th, 2026. Squawk Pod begins right now. Stand Becky by in three, two, one. Good morning, everybody. Welcome to Squawk Box right here on CNBC.

2:49We are live from the Nasdaq market site in Times Square. I'm Becky Quick, along with Joe Kernan and Andrew Ross Sorkin. Everybody's here. We're ready to go. An anthropic researcher saying that he quit the AI company because it, along with OpenAI, where he used to work, are making a headlong rush towards superintelligence and is, in his words, quote, gambling with our lives. On X, that researcher, his name is Jacob Cox, and writing the following. The people building AI earnestly believe that could kill us all by the end of the decade. This is not a marketing stunt. If anything, many executives and senior researchers will couch their phrasing in the press to sound sensible.

3:30But I hear the same people express fear privately. Caxton said that anthropic employees understand the potential dire implications of self-improving AI, but they don't trust that open AI will shepherd the technology responsibly so they feel they have to do it themselves. Now, another Anthropic researcher chimed in saying we really do earnestly believe AI could kill all humans. He said Anthropic doesn't have a plan to solve alignment for superintelligence and isn't clearly on track to do it. This is connected in many ways to the conversation we were having just yesterday on the set about the head of the head of safety and other things at OpenAI calling for.

4:15I don't know what he was calling for, but really explaining how they still don't understand how these systems even work. And ultimately, therefore, what could ultimately happen. And therefore, the question is the one that you asked yesterday, which is, OK, so what? What do you do? You can drop out like this guy is doing. Right. But we don't even believe that there's regulation that could necessarily change the dynamic because these folks are saying they don't even understand how it actually works. Which decade? This decade. This decade. It's 2026 already. Yeah. Four years. Four years. I might take that with a grain of salt.

4:54I mean, I've been I've used the word singularity about a million times and it's used in you don't know really what happened during the Big Bang. That was kind of a singular. We don't know what was happening before. You don't know what's happened. It's somehow used when you talk about black holes because you don't know. And it was also used a lot in when machine learning gets to a certain point, you don't know what's on the other side of it. So I certainly believe we don't know whether it's as dire as this. I'm not in a position to. I mean, you don't, you take it with a grain of salt, don't you, Sorkin?

5:29Not completely. You think we could all be dead by 2030? This guy's walking away from a lot. He's right before Anthropics. No, I know. You know, IPO. I think there's some amazing things that could come, but I actually think it's possible that people will design. 2050, maybe. Well, no, no. I think that what they're talking about, they're not talking about robots taking over the world in a Terminator style. Would they be killing us or would we die from what? No, what I think they're describing is the possibility of a couple of things happening. One is just the ability for a human or the system itself to genuinely build or at least help people build terrible, you know, biotech, biotech weapons.

6:17Pathogens. Pathogens. So that's one way. But the other way is if you believe it can break into all of the systems and you could shut down. For example, if you could shut down the electricity grid, really shut it down in the middle of the winter where people are going to die because all of a sudden they don't have access to electricity and they don't have heat and you can shut down utilities. I mean, you unleash these agents. You don't know what they're going to do. And I guess the football that Trump carries around with him has codes that I don't know why quantum computing couldn't figure out what the codes are to do that.

6:49You know, the bright side of all this, Andrew. What's that? At least we're not all going to die from climate change in 30 years. Put that on your back burner, baby. What you've been losing so much sleep about. The big tidal wave headed for upper Manhattan. Probably you're going to be dead already. So, I mean, think how great that is, that we don't have to. What's interesting, this guy's perspective. Sweat it out, literally. Sweat it out. You'll outlive that part. That's the good news for you. So this guy's perspective, having worked at both Anthropic and at OpenAI and saying, look, it's really these two companies that we're competing against.

7:29Anthropic can't slow down because OpenAI won't. It's not the Chinese that he mentions. We do know, having been in this business this long, that if it bleeds, it leads. How many people, very smart, very connected, very in tune with what's happening, would disagree with this completely? Are we going to rush them on and have discussions about them? By the way, that's what I mean. So, I mean, I would take all this with a salt lick based size. The reason I'm not taking it with a salt lick. Why not? Well, because Elon said it. Dario said it. Now you like it. Now you believe Elon. I have always believed Elon.

8:12Elon said it. Dario said it. Sam has said it. So I'm not sure. All of them said we're all going to be dead by 2030. It's a greater than 0 % chance. The whole point of everything we've been saying for the past six years since this project began was how dangerous it ultimately could be. Then why would any of them be plunging headlong into it? That's a good question. Because there's the excitement on one end. This is that people have to, they're handicapping which version of it's going to happen. And because they're all saying somebody else will do it if I don't. Which brings the question, do we need some regulatory form?

8:51And if we did, how do we even do that outside the United States? Do you get the top minds together from all these nations to agree to it? Because you fall back. If OpenAI and Anthropoc agree to it, then we say, okay, the Chinese are going to do it. It reminds me of Lloyd's great line. What Demis is doing. About all the inequality and the rising of populism. He goes, I'm not that worried. I have a doorman. I have a doorman building. And, you know, maybe I got to I live out in Shorthill. I think, you know, you think you're OK? I think I hope I was going to say Demis from Google, who's really the OG of this whole business.

9:28All of these people at one point or another worked for Demis. He's the one took himself out. He's the one who's been working behind the scenes to try to build a global coalition so that you could actually create some real regulations around it. So can you go off the grid like Ted Kaczynski? I mean, would that work? I don't know, Joe. You can try. That's until they have the robot. You can try. You don't know if you can go deep into, I don't know, the wilderness in Idaho. But then if you have robots, this is where you get to the Terminator side of things. I'm going to be much more optimistic. You know, the NFL starting today.

10:09The White House escalating its trade fight with Canada. The U.S. is going to ban imports of some alcoholic beverages, motorcycles and dairy products from Canada starting September 29th. Some aluminum furniture and cheeses. Oh, my God. No. We'll be hit with a 50 percent tariff, but not banned. The moves are in response to Canada's tariffs this week on the U.S., which were in retaliation for new U.S. tariffs from last month on X last night. Canada's minister responsible for trade issues with America said he's in contact with U.S. trade rep Jamison Greer. He's been on the show. And Canada will work in good faith towards a better trade relationship when the U.S.

10:52is ready to engage. Are you going to talk at all about anything interesting you've seen? I want everyone to see that. Explain what we're going to talk about. It's called You Can See Everything. And it was the highlight of the film festival. in Telluride. I mean, leading up to it, it was, we were all wondering what is going to be so monumental because it was so secretive. This is the announced. Every year, there's a film that's going to be shown that you don't know what it is until you get there. This wasn't announced until everyone was seated and your phone was in a bag. You were able to keep your phone, but it was in a bag that you couldn't open almost like when you go and you can't take that thing off when you buy something.

11:42You had to have a magnetic. And I'll just, a couple of quick comments. Nathan Fielder is maybe my favorite filmmaker at this point. It was three hours, and you definitely, you go in with one idea about Elizabeth Holmes and her husband, and I think you exit with, I did, anyway with a different idea, but it's not, mine is not a mainstream media take, Andrew, so not a conventional wisdom take. Don't say more. Right. Don't say more. I don't want you to say more because I want to go see this film. You could go see it, but I have a much different view of her than I did before. Let's discuss this in a month from now.

12:28I don't know how much I want to push this. I don't want to push this because it's going to ruin the movie. But I'm curious. Three hours, you never really look away. There's times where you'd like to look away. Because it's so... There's a lot of cringe, but then it takes three hours to have the whole experience of where you end up with how you feel about the whole thing. And there's a lot of, you know, my son's writing. You couldn't even write about it. You weren't even allowed to write about what happens. I know. I don't want you to tell more. You'll get in trouble if you say anything. End it.

12:57I'm not sure whether you get in trouble, but it's amazing. It's great for Telluride. It's the best film festival, I think. I mean, there's Cam and Dennis and everything else. But in the U.S., it's just phenomenal. And this is great for Julie, who runs it, who I know. And it was really intense. I mean, it was really an experience. I can't wait for it. Yeah, it was great. OpenAI saying its artificial intelligence has solved one of the world's toughest math problems. It is called the Navier Strokes Existence and Smoothness Problem, and it concerns the rules that describe how liquids and gases move.

13:36It's one of the seven unresolved math problems that were dubbed Millennium Prize Problems. This is back in 2000. Until now, only one had been solved. OpenAI published its findings along with the 165-page proof. The company said it took as many as 10 ,000 AI agents working together on the problem for about 88 hours to find the solution. and took millions of dollars in computing resources. So that's the exciting part about AI. And the scarier part is if you can do that, what else can you do? Tease will be next. Coming up on Squawk Pod, how AI could and should be used in classrooms around America.

14:18Microsoft has teamed up with the American Federation of Teachers to set up guardrails to protect students. AFT President Randy Weingarten joins us. On letters, recommendation letters, if you're a guidance counselor and you have 100 to write in a weekend, if you use it as the first pass and then you personalize it, that's okay. But that is different than as a kid not knowing anything. And Microsoft President Brad Smith. We don't want to slow down innovation where we need it, but we need to ensure that safety and trust go forward at the same pace as innovation. And right now, innovation is going faster.

14:56We'll be right back. When did work become so much work? The meeting about the meeting, the hundreds of files to find one insight, setting aside the things you want to do for the things that pop up. Your workday's gone. But what if the insight surfaced itself? Or you could ship the deck without the distractions. Gemini Enterprise helps you get that done. It's AI that knows your business with agents that take stuff off your plate. Make work less work with Gemini Enterprise from Google Cloud. This episode is brought to you by Schwab Market Update, an original podcast from Charles Schwab. Join host Keith Lansford for this information-packed daily market preview delivered in 10 minutes or less, including projected stock updates, monetary policy decisions, and key results and statistics that may impact your trading.

15:50Download the latest episode and subscribe at schwab.com slash market update podcast, or find Schwab Market Update wherever you get your podcasts. Hey, Chicagoland, the Wayfair store is in your neighborhood at Edens Plaza and Wilmette. Finally, you can feel the fabric, sit on the sectionals, and even open the refrigerators. Plus, our in-store designers will help you bring it all together with free one-on-one design support for any project on any budget. Yep, we said free. Oh, and did we mention the cafe? So what are you waiting for? Come see all that's in store. Visit the Wayfair store today at Edens Plaza and Wilmette.

16:27Wayfair, every style, every home. You're listening to Squawk Pod from CNBC with Joe Kernan, Becky Quick, and Andrew Ross Sorkin. Here's Joe. Our next guests are unveiling legally enforceable AI protections for schools nationwide. Joining us now, Microsoft Vice Chair and President Brad Smith. Good morning. You're on frequent guests. And American Federation of Teachers, President Randy Weingarten. And we're applauding your efforts, wondering how it came about. We want to hear about exactly what it means and what's involved. But off camera, the first thing we said was, does any of it matter if we're all dead in a couple of years from from what we were talking about earlier from Anthropic?

17:14And I mean, this is I don't remember many times in history where we've had something that universally is thought of as so dangerous that we're plunging headlong into. Well, it's not the first time. And let's remember, World War II ended, and we talked about nuclear weapons, and we found a way now to avoid war for 80 years and harness the power of the atom for nuclear power. Now we're dealing with AI. It can be used to do dangerous things. It can be used to do good things. It's got to be controlled well. This is the leader of the second largest trade or teachers union in the country. We're one of the largest tech companies.

17:57We found a way to work together over the last few months and forge what is a legally binding and enforceable contract to keep students safe, their data private, provide more transparency to parents. We're not here to say that we have found the solution to everything that will be needed. But this is part. Can you explain it, Randy? And who contacted who? We contacted them. So look, we saw what was going on with AI a few years ago. And there has been a long-term relationship between the FL and Microsoft as a way of trying to work together when you can labor and industry in terms of making, actually making America better.

18:46So through that, we said to Brad, look, we're seeing that AI is about to control us. We want teachers in classrooms to actually be in control, to be in the driver's seat. So we started that AI Institute over a year ago. But part of that was we needed to ensure parents and teachers that data was going to be safe, that their data was going to be private. And you can't do that unless you do it with the tech providers because they do the infrastructure. So from that point on, we said, let's get to an agreement on safety and privacy. And over this summer, Brad and I and our teams, we were trying to do this with Anthropic and with OpenAI as well.

19:41And I suspect that they will come along very shortly. But we, Brad saw it and saw the need. And so we did it more than for the AI Institute. We said, let's make this procurement policy for schools across America because there's no trust. Parents have no trust. Teachers have very little trust in AI. So what this agreement is, is literally a standard, 10 different aspects of a standard to ensure that kids' data will be private, that when they use kids, use AI or parents or teachers use AI in schools, kids' data will be private. Now, that's what will happen in Microsoft products throughout the country if a school district wants it.

20:28And that's what's going to happen in our AI institute. Do you also have huge concerns, though, about whether students are using AI to cheat, to get the answers, or whether they're learning how to like how do you tackle that part of it? I have all those concerns. And in May, I did a speech called Devices Down, Hands On, Eyes Up, to try to do a whole blueprint on all of this. My main concern is, I mean, this may sound blasphemous. My main concern is not cheating. My main concern is learning. My main concern is cognitive offloading and that, you know, what Jonathan Haidt and Jared Covey and others have seen.

21:10So that's why we called for no screens K through two, which L.A. and New York are now doing. We called for no A.I. facing instruction, you know, kids, A.I. kids facing instruction in elementary schools, because we need to actually create the kind of critical thinking and knowledge that kids need. They need to develop the muscle for it. But this is different. This is about ensuring that that parents have transparency and that their data is. We talk about the practicality of this. So I've got kids. You guys have kids. They have laptops. They have iPads. They have phones. They bring all three of those things on occasion to school.

21:57And sometimes they don't. But how is it that somehow this is all going to work such that when they're using it in school, it's protected. And when they're at home and they go on some other. I mean, the thing that I can't figure out about this is it's great that this service will have the privacy on school machines. But it doesn't change anything if if all the other machines don't are not using the same thing. Well, what if we actually, from what we've now negotiated or forged, what if we actually could get that into law? Which part? The part about ensuring privacy of data. We already have FERPA and HIPAA and all these laws that are supposed to protect student data, but they're outdated.

22:48They never envisioned AI. So what if we could actually use this standard to update our laws? We have forged something that could create that basis instead of lawmakers saying, oh, I don't know how to do it. I also think, Andrew, the way we've designed this is to think about what we each do well and what we don't do well. Look, I don't think anybody wants a tech company to tell teachers how to teach. I mean, let's be serious. But they do want us, I believe, and we can and should and are, stepping up with the kinds of guardrails that ensure, for example, that student data isn't used to train an AI model.

23:26And whether the student is using that application at school or home, we have a contractual obligation. We'll put it in place with every school district in America so that regardless of where the student is, that obligation persists. And if you go through the contract, the AFT is going to publish it this afternoon. Everybody will be able to read it. And I think what you'll find is these are the types of protections that our kids need. And they should really. Can I just go to Becky's question? Because we kind of glossed over it. So the calculator. I remember HP was going to. No one was ever doing long division again.

24:05Do kids learn long division? Do kids learn calculus? Does the muscle memory to do those things that no one can do anymore, did it matter that they don't have that now? And is this 10 times worse what AI could? Look, it's 10 times. Let me do this about social studies and English as opposed to about math. Just about math. OK, just because of. So if you what makes A.I. both more powerful and more and scarier is that if you can, by the touch of your prompt, get an answer, you are not learning how to think. It was math, though, but this is different. Yeah. But but what happens is it's so the the first calculators could help you do long division.

25:03They could not help you do calculus right now. Yes, I said the first ones. But if you don't have the muscle to think, if you don't have the muscle to discern fact from fiction, If you don't have the muscle to think about what is fact and what is opinion, if you can't ask the follow up question, this is why now projects based instruction is so important. If you can't look at each other like we're doing right now. Yeah, I gave me one more question. But I want to. But the second there's three pieces to this agreement. One is the actual data protection that we need to use, that the tech companies are the ones who can do that.

25:49But the second is making sure that teachers have more control over their classrooms. And that is something that we have to do with school districts. Meaning teachers actually have to control the tech. Teachers should be in charge. The tech should not be in charge. Yeah, they took away our kids' phones in New Jersey this year, and I'm all in favor of that. But you have to understand the tool in order to be in charge. When you say control the tech, what do you mean? Meaning? You're taking the phones away or you're actually doing something to the actual tech? That's a physical. Physically putting it in a different room is it.

26:22I think it's a combination. You're doing things to the actual tech. The first part of the agreement, for example, our technology will be designed and provided in a manner that teachers make the grading decisions. They don't delegate that to AI. If you have an agentic feature, it needs to be a feature that a teacher can turn off. Now it's getting more interesting. So you know a whole bunch of teachers are now doing report cards and letters to parents and other things using AI. Should they? Teachers should not be doing grading. You can do a first pass in terms of an evaluation of a paper. You can do a first pass in terms of AI.

27:03Personally, I would not do that. But you cannot give a grade based upon AI. It has to be teacher judgment. Make it more complicated. Wait, wait, wait. Should a teacher be building their lesson plan with AI? Should a teacher be building the actual classroom discussion with AI? I mean, this is to me where it gets much more interesting. So this is, so Andrew, you can use AI. So I use AI sometimes for a first pass on things. I use AI because I know the material. I just taught at Cornell last week. I used AI to actually say I did my lesson plan, but I used AI to say, what am I missing? Am I missing something?

27:45But I know the work. So AI as an aid, AI as an assist, like AI as the Encyclopedia Britannica used to be, using that that way, I think is fine. Making it the decider is not fine. And so on letters, recommendation letters, if you're a guidance counselor and you have 100 to write in a weekend, if you use it as the first pass and then you personalize it, that's OK. But that is different than as a kid not knowing anything. Say you don't know anything about the War of 1812 and you use AI to do your paper and you're and all of a sudden AI is wrong and you're talking about the Civil War, not the War of 1812.