The TED AI Show: What really went down at OpenAI and the future of regulation w/ Helen Toner

[00:00:00] TED Audio Collective.

[00:00:06] Hey, Belovl here.

[00:00:08] This episode is a bit different.

[00:00:11] Today I'm interviewing Helen Toner,

[00:00:13] a researcher who works on AI regulation.

[00:00:16] She's also a former board member at OpenAI.

[00:00:19] In my interview with Helen, she reveals for the first time

[00:00:22] what really went down at OpenAI late last year

[00:00:25] when the CEO Sam Altman was fired.

[00:00:28] And she makes some pretty serious criticisms of him.

[00:00:31] We reached out to Sam for comments,

[00:00:33] and if he responds, we'll include that update at the end of the episode.

[00:00:37] But first, let's get to the show.

[00:00:42] I'm Belovl Sidhu, and this is the TED-AI Show,

[00:00:45] where we figure out how to live and thrive in a world

[00:00:48] where AI is changing everything.

[00:00:55] This episode is brought to you by Progressive.

[00:00:58] Most of you aren't just listening right now.

[00:01:00] You're driving, cleaning, and even exercising.

[00:01:03] But what if you could be saving money by switching to Progressive?

[00:01:06] Drivers who save by switching save nearly $750 on average,

[00:01:10] and auto customers qualify for an average of seven discounts.

[00:01:14] Multitask right now. Quote today at progressive.com.

[00:01:17] Progressive Casualty Insurance Company & Affiliates.

[00:01:20] National average 12-month savings of $744 by new customers surveyed

[00:01:24] who saved with Progressive between June 2022 and May 2023.

[00:01:28] Potential savings will vary.

[00:01:30] Discounts not available in all states and situations.

[00:01:33] AI is making waves in every field it touches.

[00:01:38] President Biden is now on TikTok,

[00:01:40] and the election draws closer each day.

[00:01:42] With so much going on in the world,

[00:01:44] it is hard to keep up with it all, let me tell you.

[00:01:47] Hi, I'm Kai Riznali, co-host of Make Me Smart.

[00:01:49] It's a podcast from Marketplace.

[00:01:51] And every weekday, Kimberly Adams and I

[00:01:53] break down the latest in business and the economy

[00:01:55] with short daily episodes

[00:01:57] to make it easy for you to stay in the know.

[00:01:59] Listen to Make Me Smart wherever you get your podcasts.

[00:02:05] The OpenAI saga is still unfolding,

[00:02:07] so let's get up to speed.

[00:02:09] In case you missed it, on a Friday in November 2023,

[00:02:13] the board of directors at OpenAI fired Sam Altman.

[00:02:16] This ouster remained a top news item over that weekend

[00:02:19] with the board saying that he hadn't been, quote,

[00:02:21] consistently candid in his communications, unquote.

[00:02:25] The Monday after, Microsoft announced

[00:02:27] that they had hired Sam to head up their AI department.

[00:02:30] Many OpenAI employees rallied behind Sam

[00:02:32] and threatened to join him.

[00:02:34] Meanwhile, OpenAI announced an interim CEO.

[00:02:37] And then a day later, plot twist,

[00:02:39] Sam was rehired at OpenAI.

[00:02:41] Several of the board members were removed

[00:02:43] or resigned and replaced.

[00:02:45] Since then, there's been a steady fallout.

[00:02:48] On May 15th, 2024, just last week

[00:02:51] as of recording this episode,

[00:02:53] OpenAI's chief scientist,

[00:02:55] Ilya Sutskever, formally resigned.

[00:02:58] Not only was Ilya a member of the board that fired Sam,

[00:03:01] he was also part of the Super Alignment team,

[00:03:04] which focuses on mitigating the long-term risks of AI.

[00:03:08] With the departure of another executive, Jan Laika,

[00:03:11] many of the original safety-conscious folks

[00:03:13] in leadership positions have either departed OpenAI

[00:03:16] or moved on to other teams.

[00:03:19] So, what's going on here?

[00:03:21] Well, OpenAI started as a nonprofit in 2015,

[00:03:25] self-described as an artificial intelligence research company.

[00:03:29] They had one mission,

[00:03:30] to create AI for the good of humanity.

[00:03:33] They wanted to approach AI responsibly,

[00:03:35] to study the risks up close,

[00:03:37] and to figure out how to minimize them.

[00:03:40] This was going to be the company that showed us

[00:03:42] AI done right.

[00:03:44] Fast forward to November 17th, 2023,

[00:03:47] the day Sam was fired,

[00:03:49] OpenAI looked a bit different.

[00:03:52] They'd released Ali,

[00:03:53] and ChatGPT was taken the world by storm.

[00:03:56] With hefty investments from Microsoft,

[00:03:58] it now seemed that OpenAI was in something of a tech arms race

[00:04:01] with Google.

[00:04:02] The release of ChatGPT prompted Google to scramble

[00:04:05] and release their own chatbot, BARD.

[00:04:08] Over time, OpenAI became closed AI.

[00:04:12] Starting 2020, with the release of GPT-3,

[00:04:14] OpenAI stopped sharing their code.

[00:04:17] And I'm not saying that was a mistake.

[00:04:19] There are good reasons for keeping your code private.

[00:04:22] But OpenAI somehow changed,

[00:04:24] drifting away from a mission-minded nonprofit

[00:04:26] with altruistic goals

[00:04:28] to a run-of-the-mill tech company

[00:04:30] shipping new products at an astronomical pace.

[00:04:33] This trajectory shows you just how powerful

[00:04:36] economic incentives can be.

[00:04:38] There's a lot of money to be made in AI right now.

[00:04:41] But it's also crucial that profit isn't the only factor

[00:04:44] driving decision-making.

[00:04:46] Artificial General Intelligence, or AGI,

[00:04:49] has the potential to be very, very disruptive.

[00:04:52] And that's where Helen Toner comes in.

[00:04:56] Less than two weeks after OpenAI fired

[00:04:58] and rehired Sam Altman,

[00:05:00] Helen Toner resigned from the board.

[00:05:03] She was one of the board members

[00:05:04] who had voted to remove him.

[00:05:06] And at the time, she couldn't say much.

[00:05:08] There was an internal investigation still ongoing,

[00:05:11] and she was advised to keep mum.

[00:05:13] And oh man, she got so much flack for all of this.

[00:05:17] Looking at the news coverage and the tweets,

[00:05:19] I got the impression she was this techno-pessimist

[00:05:22] who was standing in the way of progress

[00:05:24] or a kind of maniacal power seeker

[00:05:26] using safety policy as her cudgel.

[00:05:29] But then I met Helen at this year's TED conference,

[00:05:32] and I got to hear her side of the story.

[00:05:35] And it made me think a lot about the difference

[00:05:37] between governance and regulation.

[00:05:40] To me, the OpenAI saga is all about AI board governance

[00:05:44] and incentives being misaligned

[00:05:46] among some really smart people.

[00:05:48] It also shows us why trusting tech companies

[00:05:50] to govern themselves may not always go beautifully,

[00:05:54] which is why we need external rules and regulations.

[00:05:57] It's a balance.

[00:05:59] Helen's been thinking and writing about AI policy

[00:06:03] for about seven years.

[00:06:05] She's the director of strategy at CSET,

[00:06:08] the Center for Security and Emerging Technology

[00:06:10] at Georgetown, where she works with policymakers in DC

[00:06:13] about all sorts of AI issues.

[00:06:16] Welcome to the show.

[00:06:17] Hey, good to be here.

[00:06:18] So Helen, a few weeks back at TED in Vancouver,

[00:06:21] I got the short version of what happened

[00:06:23] at OpenAI last year.

[00:06:25] I'm wondering, can you give us the long version?

[00:06:28] As a quick refresher on sort of the context here,

[00:06:31] the OpenAI board was not a normal board.

[00:06:33] It's not a normal company.

[00:06:34] The board is a nonprofit board that was set up explicitly

[00:06:38] for the purpose of making sure that the company's,

[00:06:41] you know, public good mission was primary,

[00:06:43] was coming first over profits,

[00:06:44] investor interests and other things.

[00:06:46] But for years, Sam had made it really difficult

[00:06:49] for the board to actually do that job

[00:06:52] by, you know, withholding information,

[00:06:55] misrepresenting things that were happening at the company,

[00:06:57] in some cases outright lying to the board.

[00:06:59] You know, at this point, everyone always says,

[00:07:01] like what? Give me some examples.

[00:07:03] And I can't share all the examples,

[00:07:05] but to give a sense of the kind of thing

[00:07:07] that I'm talking about, it's things like, you know,

[00:07:09] when ChatGPT came out, November 2022,

[00:07:12] the board was not informed in advance about that.

[00:07:14] We learned about ChatGPT on Twitter.

[00:07:17] Sam didn't inform the board

[00:07:19] that he owned the OpenAI startup fund,

[00:07:22] even though he, you know, constantly was claiming

[00:07:25] to be an independent board member

[00:07:27] with no financial interest in the company.

[00:07:29] On multiple occasions, he gave us inaccurate information

[00:07:33] about the small number of formal safety processes

[00:07:36] that the company did have in place,

[00:07:38] meaning that it was basically impossible for the board

[00:07:40] to know how well those safety processes were working

[00:07:43] or what might need to change.

[00:07:45] And then, you know, a last example that I can share

[00:07:47] that's been very widely reported relates to this paper

[00:07:49] that I wrote, which has been, you know,

[00:07:51] I think way overplayed in the press.

[00:07:53] For listeners who didn't follow this in the press,

[00:07:56] Helen had co-written a research paper last fall

[00:07:58] intended for policymakers.

[00:08:00] I'm not going to get into the details,

[00:08:02] but what you need to know is that Sam Altman

[00:08:04] wasn't happy about it.

[00:08:06] It seemed like Helen's paper was critical of OpenAI

[00:08:09] and more positive about one of their competitors, Anthropic.

[00:08:13] It was also published right when the Federal Trade Commission

[00:08:15] was investigating OpenAI about the data used

[00:08:18] to build its generative AI products.

[00:08:20] Essentially, OpenAI was getting a lot of heat

[00:08:23] and scrutiny all at once.

[00:08:26] The way that played into what happened in November

[00:08:28] is pretty simple. It had nothing to do with

[00:08:30] the substance of this paper.

[00:08:31] The problem was that after the paper came out,

[00:08:33] Sam started lying to other board members

[00:08:36] in order to try and push me off the board.

[00:08:38] So it was another example that just, like,

[00:08:40] really damaged our ability to trust him.

[00:08:42] And actually, it only happened in late October last year

[00:08:45] when we were already talking pretty seriously

[00:08:47] about whether we needed to fire him.

[00:08:50] And so, you know, there's kind of more individual examples,

[00:08:53] and for any individual case,

[00:08:56] Sam could always come up with some kind of, like,

[00:08:58] innocuous-sounding explanation of why it wasn't a big deal

[00:09:00] or misinterpreted or whatever.

[00:09:02] But the end effect was that after years of this kind of thing,

[00:09:06] all four of us who fired him came to the conclusion

[00:09:09] that we just couldn't believe things

[00:09:11] that Sam was telling us.

[00:09:13] It's a completely unworkable place to be in as a board,

[00:09:16] especially a board that is supposed to be providing

[00:09:19] independent oversight over the company,

[00:09:21] not just, like, you know, helping the CEO

[00:09:23] to raise more money.

[00:09:25] You know, not trusting the word of the CEO,

[00:09:28] who is your main conduit to the company,

[00:09:30] your main source of information about the company,

[00:09:31] is just, like, totally impossible.

[00:09:33] So that was kind of the background

[00:09:36] that the state of affairs coming into last fall,

[00:09:39] and we had been, you know, working at the board level

[00:09:42] as best we could to set up better structures, processes,

[00:09:46] all that kind of thing to try and, you know,

[00:09:48] improve these issues that we had been having

[00:09:49] at the board level.

[00:09:50] But then in mostly in October of last year,

[00:09:55] we had this series of conversations

[00:09:57] with these executives,

[00:09:59] where the two of them suddenly started telling us

[00:10:02] about their own experiences with Sam,

[00:10:04] which they hadn't felt comfortable sharing before,

[00:10:06] but telling us how they couldn't trust him

[00:10:09] about the toxic atmosphere he was creating.

[00:10:12] They used the phrase psychological abuse,

[00:10:15] telling us they didn't think he was the right person

[00:10:17] to lead the company to AGI,

[00:10:19] telling us they had no belief that he could or would change,

[00:10:23] no point in giving him feedback,

[00:10:24] no point in trying to work through these issues.

[00:10:26] I mean, you know, they've since tried

[00:10:28] to kind of minimize what they told us,

[00:10:30] but these were not, like, casual conversations.

[00:10:33] They were really serious to the point

[00:10:35] where they actually sent us screenshots

[00:10:38] and documentation of some of the instances

[00:10:40] they were telling us about of him lying

[00:10:43] and being manipulative in different situations.

[00:10:45] So, you know, this was a huge deal.

[00:10:47] This was a lot.

[00:10:49] And we talked it all over very intensively

[00:10:52] over the course of several weeks

[00:10:55] and ultimately just came to the conclusion

[00:10:58] that the best thing for OpenAI's mission

[00:11:00] and for OpenAI as an organization

[00:11:02] would be to bring on a different CEO.

[00:11:04] And, you know, once we reached that conclusion,

[00:11:07] it was very clear to all of us that

[00:11:09] as soon as Sam had any inkling

[00:11:11] that we might do something that went against him,

[00:11:14] he would pull out all the stops,

[00:11:16] do everything in his power to undermine the board,

[00:11:18] to prevent us from even getting to the point

[00:11:21] of being able to fire him.

[00:11:22] So, you know, we were very careful,

[00:11:24] very deliberate about who we told,

[00:11:27] which was essentially almost no one in advance

[00:11:29] other than obviously our legal team.

[00:11:31] And so that's kind of what took us to November 17th.

[00:11:35] Thank you for sharing that.

[00:11:36] Now, Sam was eventually reinstated as CEO

[00:11:39] with most of the staff supporting his return.

[00:11:41] What exactly happened there?

[00:11:43] Why was there so much pressure to bring him back?

[00:11:45] Yeah, this is obviously the elephant in the room.

[00:11:48] And unfortunately, I think there's been

[00:11:50] a lot of misreporting on this.

[00:11:52] I think there were three big things going on

[00:11:55] that helped make sense of kind of what happened here.

[00:11:58] The first is that really pretty early on,

[00:12:01] the way the situation was being portrayed

[00:12:03] to people inside the company was

[00:12:04] you have two options.

[00:12:05] Either Sam comes back immediately with no accountability,

[00:12:09] you know, totally new board of his choosing,

[00:12:11] or the company will be destroyed.

[00:12:14] And, you know, those weren't actually the only two options,

[00:12:17] and the outcome that we eventually landed on

[00:12:19] was neither of those two options.

[00:12:21] But I get why, you know, not wanting the company

[00:12:24] to be destroyed got a lot of people to fall in line,

[00:12:27] whether because they were in some cases

[00:12:30] about to make a lot of money from this upcoming tender offer

[00:12:33] or just because they love their team,

[00:12:35] they didn't want to lose their job,

[00:12:37] they cared about the work they were doing.

[00:12:38] And of course, a lot of people

[00:12:40] didn't want the company to fall apart, you know, us included.

[00:12:43] The second thing I think it's really important to know

[00:12:46] that has really gone underreported

[00:12:48] is how scared people are to go against Sam.

[00:12:53] They had experienced him retaliating against people,

[00:12:56] retaliating against them for past instances of being critical.

[00:13:00] They were really afraid of what might happen to them.

[00:13:02] So when some employees started to say, you know,

[00:13:05] wait, I don't want the company to fall apart.

[00:13:07] Like, let's bring back Sam.

[00:13:09] It was very hard for those people

[00:13:11] who had had terrible experiences

[00:13:13] to actually say that for fear that, you know,

[00:13:17] if Sam did stay in power, as he ultimately did,

[00:13:20] you know, that would make their lives miserable.

[00:13:23] And I guess the last thing I would say about this

[00:13:25] is that this actually isn't a new problem for Sam.

[00:13:29] And if you look at some of the reporting

[00:13:31] that has come out since November,

[00:13:33] it's come out that he was actually fired

[00:13:35] from his previous job at Y Combinator,

[00:13:37] which was hushed up at the time.

[00:13:39] And then, you know, his job before that,

[00:13:41] which was his only other job in Silicon Valley,

[00:13:43] his startup looped.

[00:13:45] Apparently, the management team went to the board there twice

[00:13:48] and asked the board to fire him

[00:13:50] for what they called, you know, deceptive and chaotic behavior.

[00:13:53] If you actually look at his track record,

[00:13:55] he doesn't, you know, exactly have a glowing future.

[00:13:58] He doesn't have a glowing trail of references.

[00:14:00] This wasn't a problem specific to the personalities

[00:14:03] on the board as much as he would love

[00:14:05] to kind of portray it that way.

[00:14:07] So I had to ask you about that,

[00:14:09] but this actually does tie into

[00:14:10] what we're going to talk about today.

[00:14:12] OpenAI is an example of a company

[00:14:14] that started off trying to do good,

[00:14:16] but now it's moved on to a for-profit model

[00:14:18] and it's really racing to the front of this AI game,

[00:14:21] along with all of these like ethical issues

[00:14:23] that are raised in the wake of this progress.

[00:14:26] I would argue that the OpenAI saga shows

[00:14:28] that trying to do good and regulating yourself isn't enough.

[00:14:32] So let's talk about why we need regulations.

[00:14:34] Great, let's do it.

[00:14:35] So from my perspective,

[00:14:37] AI went from the sci-fi thing

[00:14:39] that seemed far away to something

[00:14:41] that's pretty much everywhere,

[00:14:42] and regulators are suddenly trying to catch up.

[00:14:45] But I think for some people,

[00:14:46] it might not be obvious

[00:14:47] why exactly we need regulations at all.

[00:14:50] Like for the average person,

[00:14:51] it might seem like,

[00:14:52] oh, we just have these cool new tools

[00:14:54] like Ali and ChatGPT that do these amazing things.

[00:14:57] What exactly are we worried about in concrete terms?

[00:15:00] There's very basic stuff

[00:15:01] for very basic forms of the technology.

[00:15:03] Like if people are using it to decide who gets a loan,

[00:15:07] to decide who gets parole,

[00:15:09] you know, to decide who gets to buy a house,

[00:15:12] like you need that technology to work well.

[00:15:14] If that technology is going to be discriminatory,

[00:15:16] which AI often is, it turns out,

[00:15:18] you need to make sure that people have recourse.

[00:15:21] They can go back and say,

[00:15:22] hey, why was this decision made?

[00:15:23] If we're talking AI being used in the military,

[00:15:26] that's a whole other kettle of fish.

[00:15:28] And I don't know if we would say like regulation for that,

[00:15:31] but certainly need to have guidance rules,

[00:15:34] processes in place.

[00:15:35] And then kind of looking forward

[00:15:37] and thinking about more advanced AI systems,

[00:15:39] I think there's a pretty wide range of potential harms

[00:15:43] that we could well see

[00:15:44] if AI keeps getting increasingly sophisticated.

[00:15:47] You know, letting every little script kitty

[00:15:49] in their parents' basement,

[00:15:50] having the hacking capabilities

[00:15:52] of a crack NSA cell, like that's a problem.

[00:15:56] I think something that really makes AI hard

[00:15:58] for regulators to think about

[00:15:59] is that it is so many different things

[00:16:01] and plenty of the things don't need regulation.

[00:16:03] Like I don't know how Spotify decides

[00:16:06] how to make your playlist,

[00:16:07] the AI that they use for that.

[00:16:09] Like I'm happy for Spotify

[00:16:10] to just pick whatever songs they want from me

[00:16:12] and if they get it wrong, you know, who cares?

[00:16:14] But for many, many other use cases,

[00:16:16] you want to have at least some kind

[00:16:17] of basic common sense guardrails around it.

[00:16:19] I want to talk about a few specific examples

[00:16:21] that we might want to worry about,

[00:16:23] not in some battle space overseas,

[00:16:25] but at home in our day-to-day lives.

[00:16:27] You know, let's talk about surveillance.

[00:16:28] AI has gotten really good at perception,

[00:16:31] essentially understanding the contents

[00:16:33] of images, video and audio.

[00:16:35] And we've got a growing number

[00:16:36] of surveillance cameras in public and private spaces.

[00:16:39] And now companies are infusing AI into this fleet,

[00:16:42] essentially breathing intelligence

[00:16:44] into these otherwise dumb sensors

[00:16:46] that are almost everywhere.

[00:16:48] Madison Square Garden in New York City, as an example,

[00:16:51] they've been using facial recognition technology

[00:16:53] to bar lawyers involved in lawsuits

[00:16:56] against their parent company, MSG Entertainment,

[00:16:58] from attending events at their venue.

[00:17:01] This controversial practice obviously raised concerns

[00:17:03] about privacy, due process

[00:17:05] and potential for abuse of this technology.

[00:17:07] Can we talk about why this is problematic?

[00:17:10] Yeah, I mean, I think this is a pretty common thing

[00:17:12] that comes up in the history of technology

[00:17:14] is you have some, you know,

[00:17:16] some existing thing in society

[00:17:18] and then technology makes it much faster

[00:17:20] and much cheaper and much more widely available.

[00:17:22] Like surveillance, where it goes from like,

[00:17:23] oh, it used to be the case that your neighbor

[00:17:25] could see you doing something bad

[00:17:26] and go talk to the police about it.

[00:17:27] You know, it's one step up to go to,

[00:17:29] well, there's a camera, a CCTV camera

[00:17:31] and the police can go back and check at any time.

[00:17:33] And then another step up to like,

[00:17:34] oh, actually it's just running all the time

[00:17:36] and there's an AI facial recognition detector on there

[00:17:38] and maybe, you know, maybe in the future

[00:17:39] an AI activity detector that's also flagging,

[00:17:42] you know, this looks suspicious.

[00:17:45] In some ways there's no like qualitative change

[00:17:48] in what's happened.

[00:17:49] It's just like you could be seen doing something.

[00:17:51] But I think you do also need to grapple with the fact

[00:17:53] that if it's much more ubiquitous,

[00:17:55] much cheaper, then the situation is different.

[00:17:58] I mean, I think with surveillance,

[00:17:59] people immediately go to the kind of

[00:18:01] law enforcement use cases.

[00:18:03] And I think it is really important

[00:18:04] to figure out what the right trade-offs are

[00:18:06] between achieving sort of law enforcement objectives

[00:18:09] and being able to catch criminals

[00:18:11] and prevent bad things from happening

[00:18:13] while also recognizing the huge issues that you can get

[00:18:16] if this technology is used with overreach.

[00:18:18] For example, facial recognition works better and worse

[00:18:22] on different demographic groups.

[00:18:23] And so if police are, as they have been

[00:18:25] in some parts of the country,

[00:18:26] going and arresting people purely

[00:18:28] on a facial recognition match and on no other evidence,

[00:18:31] there's a story about a woman who was eight months pregnant

[00:18:33] having contractions in a jail cell

[00:18:35] after having done absolutely nothing wrong

[00:18:37] and being arrested only on the basis

[00:18:39] of a bad facial recognition match.

[00:18:41] So I personally don't go for,

[00:18:43] this needs to be totally banned

[00:18:45] and no one should ever use it in any way for anything.

[00:18:47] But I think you really need to be looking at

[00:18:49] how are people using it?

[00:18:51] What happens when it goes wrong?

[00:18:52] What recourse do people have?

[00:18:53] What kind of access to due process do they have?

[00:18:56] And then when it comes to private use,

[00:18:58] I really think we should probably

[00:18:59] be a bit more restrictive.

[00:19:01] Like, I don't know,

[00:19:02] it just seems pretty clearly against,

[00:19:04] I don't know, freedom of expression,

[00:19:06] freedom of movement for somewhere

[00:19:08] like Madison Square Gardens to be

[00:19:09] kicking their own lawyers out.

[00:19:10] I don't know, I'm not a lawyer myself,

[00:19:11] so I don't know what exactly

[00:19:13] the state of the law around that is.

[00:19:14] But I think the sort of civil liberties

[00:19:17] and privacy concerns there are pretty clear.

[00:19:21] I think the problem with sort of

[00:19:24] an existing set of technology

[00:19:25] getting infused with more advanced capabilities,

[00:19:28] sort of unbeknownst to the common population at large,

[00:19:30] is certainly a trend.

[00:19:32] And one example that shook me up

[00:19:34] is a video went viral recently

[00:19:36] of a security camera from a coffee shop,

[00:19:38] which showed a view of a cafe

[00:19:39] full of people and baristas.

[00:19:41] And basically over the heads of the customers,

[00:19:43] like the amount of time they spent at the cafe,

[00:19:45] and then over the baristas was like,

[00:19:47] how many drinks have they made?

[00:19:49] And then, you know, so what does this mean?

[00:19:50] Like ostensibly the business can one,

[00:19:52] track who is staying on their premises for how long,

[00:19:55] learn a lot about customer behavior

[00:19:57] without the customer's knowledge or consent.

[00:20:00] And then number two,

[00:20:01] the businesses can track how productive

[00:20:03] their workers are and could potentially fire,

[00:20:05] let's say less productive baristas.

[00:20:07] Let's talk about the problems and the risk here.

[00:20:09] And like, how is this legal?

[00:20:11] I mean, the short version is,

[00:20:12] and this comes up again and again and again,

[00:20:14] if you're doing AI policy,

[00:20:15] the US has no federal privacy laws.

[00:20:17] Like there are no rules on the books

[00:20:20] for how companies can use data.

[00:20:23] The US is pretty unique

[00:20:24] in terms of how few protections there are

[00:20:25] of what kinds of personal data are protected

[00:20:27] in what ways.

[00:20:28] Efforts to make laws have just failed

[00:20:30] over and over and over again,

[00:20:31] but there's now this sudden stealthy new effort

[00:20:33] that people think might actually have a chance.

[00:20:34] So who knows,

[00:20:35] maybe this problem is on the way to getting solved.

[00:20:37] But at the moment it's a big hole for sure.

[00:20:39] And I think step one

[00:20:40] is making people aware of this, right?

[00:20:42] Because people have, to your point,

[00:20:43] heard about online tracking,

[00:20:44] but having those same set of analytics

[00:20:46] in like the physical space in reality,

[00:20:48] it just feels like the Rubicon has been crossed

[00:20:51] and we don't really even know that's what's happening

[00:20:53] when we walk into whatever grocery store.

[00:20:55] I mean, again, yeah.

[00:20:56] And again, it's about sort of the scale

[00:20:59] and the ubiquity of this.

[00:21:01] Because again, it could be like your favorite barista

[00:21:05] knows that you always come in

[00:21:06] and you sit there for a few hours on your laptop

[00:21:08] because they've seen you do that a few weeks in a row.

[00:21:10] That's very different to this data

[00:21:13] is being collected systematically

[00:21:14] and then sold to data vendors all around the country

[00:21:17] and used for all kinds of other things

[00:21:18] or outside the country.

[00:21:20] So again, I think we have these sort of intuitions

[00:21:24] based on our real world person-to-person interactions

[00:21:26] that really just break down

[00:21:27] when it comes to sort of the size of data

[00:21:29] that we're talking about here.

[00:21:31] Hey, TED Tech listeners.

[00:21:33] We're supported by our friends at Working Smarter,

[00:21:36] a new podcast from Dropbox

[00:21:38] exploring the exciting potential of AI in the workplace.

[00:21:42] Working Smarter talks with founders,

[00:21:44] researchers and engineers

[00:21:45] about the things they're building

[00:21:47] and the problems they're solving

[00:21:48] with the help of the latest AI tools.

[00:21:51] Tools that can save them time,

[00:21:52] improve collaboration

[00:21:53] and create more space for the work that matters most.

[00:21:57] On Working Smarter,

[00:21:58] hear practical discussions about what AI can do

[00:22:01] so that you can work smarter too.

[00:22:03] Listen to Working Smarter on Apple Podcasts, Spotify

[00:22:06] or wherever you get your podcasts

[00:22:08] or visit workingsmarter.ai.

[00:22:12] I also want to talk about scams.

[00:22:14] So folks are being targeted by phone scams.

[00:22:17] They get a call from their loved ones.

[00:22:18] It sounds like their family members have been kidnapped

[00:22:20] and being held for ransom.

[00:22:22] In reality, some bad actor just used off the shelf AI

[00:22:26] to scrub their social media feeds for these folks voices

[00:22:29] and scammers can then use this

[00:22:30] to make these very believable hoax calls

[00:22:33] where people sound like they're in distress

[00:22:35] and being held captive somewhere.

[00:22:36] So we have reporting on this particular hoax now,

[00:22:39] but what's on the horizon?

[00:22:41] What's like keeping you up at night?

[00:22:43] I mean, I think that the obvious next step

[00:22:45] would be with video as well.

[00:22:46] I mean, definitely if you haven't already gone

[00:22:48] and talked to your parents or grandparents,

[00:22:50] anyone in your life who is not super tech savvy

[00:22:53] and told them like, you need to be on the lookout for this,

[00:22:56] you should go do that.

[00:22:57] I talk a lot about kind of policy

[00:22:59] and what kind of government involvement

[00:23:01] or regulation we might need for AI.

[00:23:02] I do think a lot of things we can just adapt to

[00:23:04] and we don't necessarily need new rules for.

[00:23:06] So I think, you know,

[00:23:07] we've been through a lot of different waves of online scams

[00:23:10] and I think this is the newest one

[00:23:11] and it really sucks for the people who get targeted by it.

[00:23:14] But I also expect that, you know,

[00:23:15] five years from now would be something

[00:23:17] that people are pretty familiar with

[00:23:18] and will be a pretty small number of people

[00:23:20] who are still vulnerable to it.

[00:23:21] So I think the main thing is, yeah,

[00:23:23] be super suspicious of any voice.

[00:23:25] Definitely don't use voice recognition

[00:23:27] for like your bank accounts or things like that.

[00:23:29] I'm pretty sure some banks still offer that,

[00:23:31] you know, ditch that.

[00:23:32] Definitely use something more secure

[00:23:34] and yeah, be on the lookout for video scamming as well

[00:23:38] and for people, you know, on video calls who look real.

[00:23:40] I think there was recently just the other day

[00:23:42] a case of a guy who was on a whole conference call

[00:23:45] where there were a bunch of different AI-generated people

[00:23:47] all on the call and he was the only real person,

[00:23:49] got scammed out a bunch of money.

[00:23:51] So that's coming.

[00:23:52] Totally, content-based authentication

[00:23:54] is on its last legs it seems.

[00:23:56] Definitely.

[00:23:57] It's always worth like checking in with

[00:23:58] what is the baseline that we're starting with?

[00:24:00] And I mean, so for instance, a lot of things,

[00:24:03] a lot of things are already public

[00:24:04] and they don't seem to get misused.

[00:24:05] So like I think a lot of people's addresses

[00:24:08] are listed publicly.

[00:24:09] You know, we used to have literal white pages

[00:24:11] where you can look up someone's address

[00:24:13] and that mostly didn't result in terrible things happening.

[00:24:15] Or, you know, I even think of silly examples

[00:24:17] like I think it's really nice that delivery drivers

[00:24:20] or when you go to a restaurant

[00:24:21] to pick up food that you ordered, it's just there.

[00:24:24] All right, so let's talk about what we can actually do.

[00:24:26] It's one thing to regulate businesses

[00:24:28] like cafes and restaurants.

[00:24:30] It's another thing to rein in all the bad actors

[00:24:33] that could abuse this technology.

[00:24:35] Can laws and regulations actually protect us?

[00:24:38] Yeah, they definitely can.

[00:24:39] I mean, and they already are.

[00:24:40] Again, AI is so many different things

[00:24:42] that there's no one set of AI regulations.

[00:24:44] There's plenty of laws and regulations

[00:24:46] that already apply to AI.

[00:24:47] So there's a lot of concern about AI,

[00:24:50] you know, algorithmic discrimination with good reason.

[00:24:53] But in a lot of cases, there are already laws on the books

[00:24:55] saying, you know, you can't discriminate

[00:24:56] on the basis of race or gender or sexuality

[00:24:58] or whatever it might be.

[00:25:00] And so in those cases, it's not even,

[00:25:02] you don't even need to pass new laws

[00:25:04] or make new regulations.

[00:25:05] You just need to make sure

[00:25:06] that the agencies in question have,

[00:25:08] you know, the staffing they need.

[00:25:10] Maybe they have the,

[00:25:12] maybe they need to have the exact authorities

[00:25:14] they have tweaked in terms of

[00:25:15] who are they allowed to investigate

[00:25:16] or who are they allowed to penalize

[00:25:17] or things like that.

[00:25:18] There are already rules

[00:25:19] for things like self-driving cars.

[00:25:20] You know, the Department of Transportation

[00:25:22] is handling that.

[00:25:23] It makes sense for them to handle that.

[00:25:24] For AI and banking,

[00:25:26] there's a bunch of banking regulators

[00:25:27] that have a bunch of rules.

[00:25:29] So I think there's a lot of places

[00:25:30] where, you know, AI actually isn't fundamentally new

[00:25:33] and the existing systems that we have in place

[00:25:35] are doing an okay job at handling that,

[00:25:38] but they may need, again, more staff

[00:25:40] or slight changes to what they can do.

[00:25:42] And I think there are a few different places

[00:25:44] where there are kind of new challenges emerging

[00:25:47] at sort of the cutting edge of AI

[00:25:49] where you have systems that can really do things

[00:25:51] that computers have never been able to do before

[00:25:53] and whether there should be rules around

[00:25:54] making sure that those systems

[00:25:56] are being kind of developed and deployed responsibly.

[00:25:58] I'm particularly curious

[00:26:00] if there's something that you've come across

[00:26:02] that's really clever

[00:26:03] or like a model for what good regulation looks like.

[00:26:06] I think this is mostly still a work in progress,

[00:26:08] so I don't know that I've seen anything

[00:26:09] that I think really absolutely nails it.

[00:26:12] I think a lot of the challenge

[00:26:13] that we have with AI right now

[00:26:16] relates to how much uncertainty there is

[00:26:17] about what the technology can do,

[00:26:19] what it's going to be able to do in five years.

[00:26:21] You know, experts disagree enormously

[00:26:23] about those questions,

[00:26:24] which makes it really hard to make policy.

[00:26:26] So a lot of the policies that I'm most excited about

[00:26:28] are about shedding light on those kinds of questions,

[00:26:31] giving us a better understanding

[00:26:32] of where the technology is.

[00:26:34] So some examples of that

[00:26:37] are things like

[00:26:39] President Biden created this big executive order

[00:26:41] last October

[00:26:42] and had all kinds of things in there.

[00:26:44] One example was a requirement

[00:26:46] that companies that are training

[00:26:48] especially advanced systems

[00:26:50] have to report certain information

[00:26:51] about those systems to the government.

[00:26:53] And so that's a requirement where

[00:26:54] you're not saying you can't build that model,

[00:26:56] can't train that model.

[00:26:58] You're not saying the government has to approve something.

[00:27:00] You're really just sharing information

[00:27:02] and creating kind of more awareness

[00:27:04] and more ability to respond

[00:27:05] as the technology changes over time,

[00:27:07] which is such a challenge for government

[00:27:09] keeping up with this fast-moving technology.

[00:27:11] There's also been a lot of good movement

[00:27:13] towards funding

[00:27:15] like the science of measuring and evaluating AI.

[00:27:18] A huge part of the challenge

[00:27:20] with figuring out what's happening with AI

[00:27:22] is that we're really bad at actually

[00:27:24] just measuring how good is this AI system?

[00:27:26] How do these two AI systems compare to each other?

[00:27:28] Is one of them sort of quote unquote smarter?

[00:27:30] So I think there's been a lot of attention

[00:27:32] over the last year or two

[00:27:33] into funding and establishing within government

[00:27:37] better capabilities on that front.

[00:27:39] I think that's really productive.

[00:27:41] Okay, so policymakers are definitely aware of AI

[00:27:43] if they weren't before

[00:27:45] and plenty of people are worried about it.

[00:27:47] They want to make sure it's safe, right?

[00:27:49] But that's not necessarily easy to do.

[00:27:52] And you've talked about this,

[00:27:54] how it's hard to regulate AI.

[00:27:56] So why is that?

[00:27:58] What makes it so hard?

[00:28:00] Yeah, I think there's at least three things

[00:28:01] that make it very hard.

[00:28:03] One thing is AI has so many different things

[00:28:05] like we've talked about.

[00:28:07] It cuts across sector,

[00:28:09] it has so many different use cases

[00:28:11] it's really hard to get your arms around

[00:28:13] what it is, what it can do, what impacts it will have.

[00:28:15] A second thing is it's a moving target.

[00:28:17] So what the technology can do

[00:28:19] is different now than it was even two years ago,

[00:28:21] let alone five years ago, 10 years ago.

[00:28:23] And policymakers are not good

[00:28:25] at sort of agile policymaking.

[00:28:28] They're not like software developers.

[00:28:29] And then the third thing is

[00:28:31] no one can agree on how they're changing

[00:28:33] or how they're going to change in the future.

[00:28:35] If you ask five experts,

[00:28:37] you know, where the technology is going,

[00:28:39] you will get five completely different answers,

[00:28:41] often five very confident,

[00:28:43] completely different answers.

[00:28:45] So that makes it really difficult

[00:28:47] for policymakers as well

[00:28:49] because they need to get scientific consensus

[00:28:52] and just like take that and run with it.

[00:28:54] So I think maybe this kind of third factor

[00:28:57] is the one that I think is the biggest challenge

[00:28:59] for policy for AI,

[00:29:01] which is that for policymakers,

[00:29:03] it's very hard for them to tell

[00:29:05] who should they listen to,

[00:29:07] what problems should they be worried about,

[00:29:09] and how is that going to change over time?

[00:29:11] Speaking of who you should listen to,

[00:29:13] obviously, you know, the very large companies

[00:29:15] in this space have an incentive

[00:29:17] and there's been a lot of talk

[00:29:19] about regulatory capture.

[00:29:21] When you ask for transparency,

[00:29:23] why would companies give a peek

[00:29:25] under the hood of what they're building?

[00:29:27] They just cite this to be proprietary.

[00:29:29] They don't have a policy

[00:29:31] and institutional framework

[00:29:33] that is actually beneficial for them

[00:29:35] and sort of prevents any future competition.

[00:29:37] How do you get these powerful companies

[00:29:39] to like participate and play nice?

[00:29:41] Yeah, it's definitely very challenging

[00:29:43] for policymakers to figure out

[00:29:45] how to interact with those companies.

[00:29:47] Again, because, you know, in part

[00:29:49] because they're lacking the expertise

[00:29:51] and the time to really dig into things

[00:29:53] in depth themselves.

[00:29:55] Like a typical Senate staffer

[00:29:57] might cover like, you know,

[00:29:59] infrastructure and education, you know,

[00:30:01] and that's like their portfolio.

[00:30:03] So they are scrambling.

[00:30:05] Like they have to, they need outside help.

[00:30:07] So I think it's very natural

[00:30:09] that the companies do come in and play a role.

[00:30:11] And I also think there are plenty of ways

[00:30:13] that policymakers can really mess things up

[00:30:15] if they don't, you know,

[00:30:17] know how the technology works

[00:30:19] and they're not talking to the companies

[00:30:21] they're regulating about what's going to happen.

[00:30:23] The challenge, of course,

[00:30:25] is how do you balance that with external voices

[00:30:27] who are going to point out the places

[00:30:29] to also be in these conversations?

[00:30:31] Certainly what we try to do at C-SET,

[00:30:33] the organization I work at,

[00:30:35] we're totally independent and, you know,

[00:30:37] really just trying to work in the best interest

[00:30:39] of making good policy.

[00:30:41] The big companies obviously do need

[00:30:43] to have a seat at the table,

[00:30:45] but you would hope that they have,

[00:30:47] you know, a seat at the table

[00:30:49] and not 99 seats out of 100

[00:30:51] in terms of who policymakers

[00:30:53] are talking to and listening to.

[00:30:55] There also seems to be a challenge

[00:30:57] with enforcement, right?

[00:30:59] You can't really put that genie

[00:31:01] back in the bottle,

[00:31:03] nor can you really start, you know,

[00:31:05] moderating how this technology is used

[00:31:07] without, I don't know,

[00:31:09] like going full 1984

[00:31:11] and having a process on every single computer

[00:31:13] monitoring what they're doing.

[00:31:15] So how do we deal with this landscape

[00:31:17] where you do have, you know,

[00:31:19] closed source and open source,

[00:31:21] like various ways to access

[00:31:23] and build upon this technology?

[00:31:25] Yeah, I mean, I think there are

[00:31:27] a lot of intermediate things

[00:31:29] that we can do.

[00:31:31] There's things like, you know,

[00:31:33] Hugging Face, for example,

[00:31:35] is a very popular platform

[00:31:37] for open source AI models.

[00:31:39] So Hugging Face in the past

[00:31:41] has delisted models that are,

[00:31:43] you know, considered to be

[00:31:45] offensive or dangerous

[00:31:46] or whatever it might be.

[00:31:48] And that actually does meaningfully

[00:31:50] reduce kind of the usage

[00:31:52] of those models because

[00:31:54] Hugging Face's whole deal

[00:31:56] is to make them more accessible,

[00:31:57] to make them more accessible

[00:31:59] to people who are

[00:32:01] not necessarily

[00:32:03] the most vulnerable.

[00:32:05] And so, you know,

[00:32:07] there's a lot of things

[00:32:09] that can be done

[00:32:11] to reduce the use of

[00:32:13] those models.

[00:32:15] And I think that's a really

[00:32:17] important thing to think about

[00:32:19] when we're talking about

[00:32:21] open source AI.

[00:32:23] And then, you know,

[00:32:24] particularly on the kind of

[00:32:26] image and audio generation side,

[00:32:28] there are some really interesting

[00:32:30] initiatives underway,

[00:32:31] mostly being led by industry,

[00:32:33] around what gets called

[00:32:35] content provenance or content

[00:32:37] authentication, which is

[00:32:38] basically how do you know where

[00:32:40] this piece of content came from?

[00:32:42] How do you know if it's real?

[00:32:44] And that's a very rapidly

[00:32:45] evolving space and a lot of

[00:32:46] interesting stuff happening

[00:32:47] there. I think there's a good

[00:32:48] amount of promise, not for

[00:32:49] perfect solutions where we'll

[00:32:50] always know, is this real or

[00:32:51] is it fake, but for making

[00:32:52] sure that we're not saying,

[00:32:53] this is fake.

[00:32:54] It was AI generated by this

[00:32:55] particular model or this

[00:32:57] is real. It was taken on this

[00:32:58] kind of camera and we have the

[00:33:00] cryptographic signature for

[00:33:01] that. I don't think we'll ever

[00:33:02] have perfect solutions.

[00:33:03] And again, I think, you know,

[00:33:05] societal adaptation is just

[00:33:06] going to be a big part of the

[00:33:07] story.

[00:33:08] But I do think there's

[00:33:09] pretty interesting technical

[00:33:11] and policy options that

[00:33:13] can make a difference.

[00:33:14] Definitely. And even if you

[00:33:15] can't completely control, you

[00:33:17] know, the generation of this

[00:33:18] material, there are ways to

[00:33:20] drastically cap the

[00:33:21] evolution of it.

[00:33:22] And so, like, I think that

[00:33:24] reduces some of the harms

[00:33:25] there. Yeah.

[00:33:26] At the same time, labeling

[00:33:27] content that is synthetically

[00:33:29] generated, a bunch of

[00:33:30] platforms have started doing

[00:33:31] that. That's exciting because

[00:33:32] like, I don't think the

[00:33:33] average consumer should be a

[00:33:34] deep fake detection expert.

[00:33:36] Right. But really, like if

[00:33:37] there could be a technology

[00:33:39] solution to this, that feels a

[00:33:40] lot more exciting, which

[00:33:42] brings me to the future.

[00:33:43] I'm kind of curious in your

[00:33:44] mind, what's like the

[00:33:46] dystopian scenario and the

[00:33:47] utopian scenario in all of

[00:33:49] this? Let's start with a

[00:33:50] little bit of the

[00:33:52] dystopian scenario.

[00:33:53] What does a world look like

[00:33:54] with inadequate or bad

[00:33:56] regulations? Paint a picture

[00:33:57] for us.

[00:33:59] So many possibilities.

[00:34:01] I mean, I think I think there

[00:34:02] are worlds that are not that

[00:34:03] different from now where you

[00:34:04] just have automated systems

[00:34:05] doing a lot of things,

[00:34:06] playing a lot of important

[00:34:07] roles in society, in some

[00:34:08] cases doing them badly and

[00:34:09] people not having the ability

[00:34:11] to go in and question those

[00:34:12] decisions. There's obviously

[00:34:13] this whole discourse around

[00:34:15] existential risk from AI,

[00:34:16] et cetera, et cetera.

[00:34:17] Kamala Harris had a whole

[00:34:18] speech about like, you know,

[00:34:19] someone loses access to

[00:34:20] Medicare because of an

[00:34:21] algorithmic issue. Like, is

[00:34:22] that not existential for

[00:34:23] that? You know, an elderly

[00:34:24] person, you know. So there

[00:34:26] are already people who are

[00:34:27] being directly impacted by

[00:34:29] algorithmic systems and AI

[00:34:31] in really serious ways.

[00:34:32] Even, you know, some of the

[00:34:33] reporting we've seen over the

[00:34:34] last couple months of how AI

[00:34:35] is being used in warfare,

[00:34:36] like, you know, videos of a

[00:34:38] drone chasing a Russian

[00:34:39] soldier around a tank and

[00:34:40] then shooting him. Like, I

[00:34:43] don't think we're full

[00:34:44] dystopia, but there's sort

[00:34:45] of plenty of plenty of

[00:34:46] things we worried about

[00:34:47] already.

[00:34:48] Something I think I worry

[00:34:49] about quite a bit or that

[00:34:50] feels intuitively to me to

[00:34:52] be a particularly plausible

[00:34:53] way things could go is sort

[00:34:55] of what I think of as the

[00:34:57] WALL-E future. I don't know

[00:34:58] if you remember that movie.

[00:35:00] Oh, absolutely.

[00:35:01] With the little robot. And

[00:35:02] the piece that I'm talking

[00:35:03] about is not the like junk

[00:35:05] earth and whatever. The

[00:35:06] piece I'm talking about is

[00:35:07] the people in that movie.

[00:35:09] They just sit in their soft

[00:35:11] roll around wheelchairs all

[00:35:12] day and, you know, have

[00:35:14] content and food and

[00:35:17] whatever to keep them

[00:35:18] happy. And I think what

[00:35:20] worries me about that is I

[00:35:21] do think there's a really

[00:35:22] natural gradient to go

[00:35:24] towards what people want in

[00:35:26] the moment and will, you

[00:35:27] know, will choose in the

[00:35:28] moment, which is different

[00:35:30] from what they, you know,

[00:35:32] will really find fulfilling

[00:35:33] or what will build kind of

[00:35:34] a meaningful life. And I

[00:35:36] think there's just really

[00:35:37] natural commercial

[00:35:38] incentives to build things

[00:35:39] that people sort of

[00:35:40] superficially want but then

[00:35:42] end up with this really

[00:35:43] kind of meaningless,

[00:35:45] shallow, superficial

[00:35:47] world. And potentially

[00:35:48] one where kind of most of

[00:35:49] the consequential decisions

[00:35:50] are being made by machines

[00:35:52] that have no concept of

[00:35:55] what it means to lead a

[00:35:56] meaningful life. And, you

[00:35:57] know, because how would we

[00:35:58] program that into them

[00:35:59] because we have no, we

[00:36:00] struggle to kind of put our

[00:36:01] finger on it ourselves. So

[00:36:03] I think those kinds of

[00:36:04] futures where not where

[00:36:05] there's some, you know,

[00:36:06] dramatic big event, but

[00:36:09] just where we kind of

[00:36:10] gradually hand over more

[00:36:11] and more control of the

[00:36:13] future to computers that

[00:36:14] are more and more

[00:36:15] sophisticated but that

[00:36:16] don't really have any

[00:36:17] concept of meaning or

[00:36:19] beauty or joy or

[00:36:20] fulfillment or, you know,

[00:36:22] flourishing or whatever it

[00:36:23] might be. I hope we

[00:36:26] don't go down those paths

[00:36:27] but it definitely seems

[00:36:28] possible that we will.

[00:36:30] They can play to our

[00:36:31] hopes, wishes, anxieties,

[00:36:32] worries, all of that. Just

[00:36:33] give us like the junk food

[00:36:35] all the time whether that's

[00:36:36] like in terms of

[00:36:37] nutrition or in terms of

[00:36:38] just like audiovisual

[00:36:39] content and that could

[00:36:40] certainly end badly. Let's

[00:36:42] talk about the opposite of

[00:36:43] that, the utopian

[00:36:44] scenario. What does a

[00:36:45] world look like where

[00:36:46] we've got this perfect

[00:36:47] balance of innovation and

[00:36:49] regulation and society is

[00:36:51] thriving?

[00:36:52] I mean, I think a very

[00:36:53] basic place to start is

[00:36:54] can we solve some of the

[00:36:55] big problems in the world?

[00:36:56] And I do think that AI

[00:36:57] could help with those. So

[00:36:58] can we have a world

[00:37:00] without climate change, a

[00:37:01] world with much more

[00:37:02] abundant energy that is

[00:37:03] much more, you know,

[00:37:04] cheaper and therefore more

[00:37:05] people can have more

[00:37:06] access to it? Where, you

[00:37:08] know, we have better

[00:37:09] agriculture so there's

[00:37:11] greater access to food.

[00:37:12] And beyond that, you know,

[00:37:13] I think what I'm more

[00:37:14] interested in is setting,

[00:37:16] you know, our kids and our

[00:37:18] grandkids and our great

[00:37:19] grandkids up to be deciding

[00:37:20] for themselves what they

[00:37:21] want the future to look

[00:37:22] like from there rather than

[00:37:23] having kind of some

[00:37:24] particular vision of where

[00:37:26] it should go. But I

[00:37:28] absolutely think that AI

[00:37:29] has the potential to really

[00:37:30] contribute to solving some

[00:37:31] of the biggest problems

[00:37:32] that we kind of face as a

[00:37:33] civilization. It's hard to

[00:37:34] say that sentence without

[00:37:35] sounding kind of grandiose

[00:37:36] and, you know, trite. But

[00:37:38] I think it's true.

[00:37:39] So, you know, I think

[00:37:40] that's a good question.

[00:37:41] I mean, I think that's

[00:37:42] true.

[00:37:43] So maybe to close things

[00:37:44] out, just like what can we

[00:37:45] do? You mentioned some

[00:37:46] examples of being aware

[00:37:48] of synthetically

[00:37:49] generated content. What

[00:37:50] can we as individuals do

[00:37:52] when we encounter, use, or

[00:37:54] even discuss AI? Any

[00:37:55] recommendations?

[00:37:57] I think my biggest

[00:37:58] suggestion here is just

[00:38:00] not to be intimidated by

[00:38:01] the technology and not to

[00:38:03] be intimidated by

[00:38:04] technologists. Like this

[00:38:05] is really a technology

[00:38:06] where we don't know what

[00:38:07] we're doing. The best

[00:38:08] experts in the world

[00:38:09] don't understand how it

[00:38:10] works. And so I think

[00:38:11] just, you know, if you

[00:38:12] find it interesting, being

[00:38:13] interested. If you think

[00:38:14] of fun ways to use it,

[00:38:15] use them. If you're

[00:38:16] worried about it, feel

[00:38:17] free to be worried. Like,

[00:38:18] you know, I think the main

[00:38:19] thing is just feeling like

[00:38:20] you have a right to your

[00:38:21] own take on what you want

[00:38:23] to happen with the

[00:38:24] technology. And no

[00:38:26] regulator, no, you know,

[00:38:29] CEO is ever going to have

[00:38:31] full visibility into all

[00:38:32] of the different ways

[00:38:33] that it's affecting, you

[00:38:34] know, millions and

[00:38:35] billions of people around

[00:38:36] the world. And so kind of,

[00:38:37] I don't know, trusting

[00:38:38] your own experience and

[00:38:39] exploring for yourself.

[00:38:40] And seeing what you think

[00:38:41] is, I think the main

[00:38:42] suggestion I would have.

[00:38:43] It was a pleasure having

[00:38:44] you on, Helen. Thank you

[00:38:45] for coming on the show.

[00:38:46] Thanks so much. This is

[00:38:47] fun.

[00:38:48] So maybe I bought into

[00:38:52] the story that played out

[00:38:53] on the news and on X, but

[00:38:55] I went into that

[00:38:56] interview expecting Helen

[00:38:57] Toner to be more of an

[00:38:59] AI policy maximalist. You

[00:39:01] know, the more laws, the

[00:39:02] better, which wasn't at

[00:39:04] all the person I found

[00:39:05] her to be. Helen sees a

[00:39:07] place for rules, a place

[00:39:08] for techno-optimism, and

[00:39:10] a place for society to

[00:39:11] just roll with, adapting

[00:39:13] to the changes as they

[00:39:14] come for balance. Policy

[00:39:17] doesn't have to mean being

[00:39:19] heavy handed and

[00:39:20] hamstringing innovation.

[00:39:21] It can just be a check

[00:39:23] against perverse economic

[00:39:24] incentives that are really

[00:39:26] not good for society. And

[00:39:27] I think you'll agree. But

[00:39:29] how do you get good

[00:39:30] rules? A lot of people in

[00:39:32] tech are going to say,

[00:39:33] you don't know shit. They

[00:39:34] know the technology the

[00:39:35] best, the pitfalls, not

[00:39:37] the lawmakers. And Helen

[00:39:39] talked about the average

[00:39:40] Washington staffer who

[00:39:41] isn't an expert, doesn't

[00:39:43] even have the time to

[00:39:44] become an expert. And yet

[00:39:46] it's on them to craft

[00:39:47] regulations that govern

[00:39:49] AI for the benefit of all

[00:39:51] of us. Technologists have

[00:39:53] the expertise, but they've

[00:39:54] also got that profit

[00:39:55] motive. Their interests

[00:39:57] aren't always going to be

[00:39:58] the same as the rest of

[00:39:59] ours. You know, in tech

[00:40:00] you'll hear a lot of

[00:40:01] regulation bad, don't

[00:40:03] engage with regulators.

[00:40:04] And I get the distrust.

[00:40:06] Sometimes regulators do

[00:40:08] not know what they're

[00:40:09] doing. India recently put

[00:40:11] out an advisory saying

[00:40:12] every AI model deployed

[00:40:14] in India first had to be

[00:40:15] approved by regulators.

[00:40:17] Totally unrealistic. There

[00:40:19] was a huge backlash there

[00:40:20] and they've since reversed

[00:40:21] that decision. But not

[00:40:23] engaging with government

[00:40:24] is only going to give us

[00:40:25] more bad laws. So we got

[00:40:27] to start talking, if

[00:40:29] only, to avoid that

[00:40:30] WALL-E dystopia.

[00:40:33] Okay, before we sign off

[00:40:34] for today, I want to turn

[00:40:36] your attention back to the

[00:40:37] top of our episode. I

[00:40:39] told you we were going to

[00:40:40] reach out to Sam Altman

[00:40:41] for comments. So a

[00:40:43] couple of hours ago, we

[00:40:44] shared a transcript of

[00:40:45] this recording with Sam

[00:40:46] and invited him to

[00:40:47] respond. We've just

[00:40:49] received a response from

[00:40:50] Brett Taylor, chair of

[00:40:51] the OpenAI board, and

[00:40:53] here's the statement in

[00:40:54] full. Quote, we are

[00:40:56] disappointed that

[00:40:57] Ms. Toner continues to

[00:40:58] revisit these issues. An

[00:41:00] independent committee of

[00:41:01] the board worked with the

[00:41:02] law firm WilmerHale to

[00:41:03] conduct an extensive

[00:41:04] review of the events of

[00:41:05] November. The review

[00:41:06] concluded that the prior

[00:41:07] board's decision was not

[00:41:09] based on concerns

[00:41:10] regarding product safety

[00:41:11] or security, the pace of

[00:41:13] development, OpenAI's

[00:41:14] finances, or its

[00:41:16] statements to investors,

[00:41:17] customers, or business

[00:41:18] partners. Additionally,

[00:41:20] over 95% of employees,

[00:41:22] including senior

[00:41:23] leadership, asked for

[00:41:24] Sam's reinstatement as

[00:41:25] CEO and the resignation

[00:41:26] of the prior board. Our

[00:41:28] focus remains on moving

[00:41:29] forward and pursuing

[00:41:30] OpenAI's mission to

[00:41:31] ensure AGI benefits all

[00:41:33] of humanity. End quote.

[00:41:35] We'll keep you posted if

[00:41:37] anything unfolds. The TED

[00:41:39] AI show is a part of the

[00:41:44] TED Audio Collective and

[00:41:46] is produced by TED with

[00:41:47] Cosmic Standard. Our

[00:41:49] producers are Ella Fetter

[00:41:51] and Sarah McCray. Our

[00:41:52] editors are Ben Bensheng

[00:41:54] and Alejandra Salazar. Our

[00:41:56] showrunner is Ivana

[00:41:57] Tucker, and our associate

[00:41:58] producer is Ben Montoya.

[00:42:00] Our engineer is Asia

[00:42:01] Pilar Simpson, our

[00:42:03] technical director is

[00:42:04] Jacob Winnick, and our

[00:42:05] executive producer is

[00:42:06] Eliza Smith. Our fact

[00:42:08] checkers are Julia

[00:42:09] Dickerson and Dan

[00:42:10] Kalachi. And I'm your

[00:42:12] host, Bilal Al-Saddu. See

[00:42:14] y'all in the next one.