OpenAI researcher speaks on why she resigned | ABC NEWS (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:02) Anthropic's artificial intelligent (00:00:05) assistant known as Claude is pitched as (00:00:07) a kind of safe AI thinking partner. But (00:00:11) this month, the company's head of trust (00:00:13) and safety sensationally quit, posting (00:00:16) on social media that the world is in (00:00:18) peril and that we're approaching a (00:00:20) threshold where our wisdom must grow in (00:00:23) equal measure to our capacity to affect (00:00:25) the world, lest we face consequences. (00:00:29) He's now off to study poetry. All this (00:00:32) as Anthropic negotiates with the US (00:00:34) Pentagon over the use of its AI on (00:00:37) classified systems. H and other AI (00:00:41) companies toy with the idea of (00:00:43) introducing targeted advertising on (00:00:46) their platforms. New businesses often (00:00:48) struggle with cash flow. So try Quick (00:00:50) Dash Payday Loans because girl bosses (00:00:52) need CEO money quick. (00:00:55) What (00:00:57) >> would you like to make a quick credit (00:00:58) check? (00:01:01) >> That ad is not so subtle a dig at (00:01:04) Anthropic's rival Open AI, which is (00:01:07) pushing ahead with advertising plans and (00:01:10) is now at least in part the reason (00:01:12) behind Open AI researcher Zoe Hitig's (00:01:16) resignation. (00:01:17) And Zoe Hitzik joins us now from New (00:01:20) York. A very good morning to you Zoe. (00:01:21) Thank you for your time. (00:01:23) >> Thank you. Great to be here. (00:01:25) >> Zoe, can you talk me through the reasons (00:01:27) about why this industry has made you so (00:01:30) nervous, so nervous that you've in fact (00:01:31) left your job? (00:01:34) >> So, what what I think about this (00:01:36) industry right now is that it's just (00:01:38) moving extraordinarily fast and we can't (00:01:41) keep up with the changes that it's (00:01:43) making to our society. You know, 800 (00:01:46) million people log into ChatGpt every (00:01:49) week. I don't know the numbers for (00:01:51) anthropic. I think something like 700 (00:01:53) million people every month log into (00:01:56) Gemini, Google's product. And this is (00:01:58) all coming at us out of nowhere almost. (00:02:02) And what I'm really concerned about is (00:02:04) that we don't understand this new type (00:02:07) of interaction that is forming between (00:02:09) humans and chat bots. So what do you (00:02:12) think the impact of this might be on (00:02:15) this unknown social interactions that (00:02:18) humanity is having with this tech? (00:02:21) >> Well, one of the things that concerns me (00:02:24) and one of the reasons why advertising (00:02:26) is a concern for us to be bringing to (00:02:30) everyone's attention right now is that (00:02:32) advertising creates a direct translation (00:02:35) between time spent on the platform and (00:02:39) dollars for the company. And what I'm (00:02:41) really concerned about is that that (00:02:43) creates an incentive for these companies (00:02:45) to keep people on the platforms to keep (00:02:49) people engaged to keep people coming (00:02:51) back and we really don't know what (00:02:53) dependence on these platforms will mean (00:02:56) for people socially and psychologically. (00:03:00) >> Don't these companies like Open AI where (00:03:02) you worked adhere to strong and strict (00:03:06) principles around running (00:03:08) advertisements? (00:03:10) Certainly the company OpenAI put out a (00:03:14) set of principles that they plan to (00:03:17) adhere to and I feel rather confident (00:03:20) that in the very first iteration of (00:03:23) advertising they will probably adhere to (00:03:25) those principles pretty strongly. But (00:03:28) what I'm concerned about is that when (00:03:30) you strap this giant billion billion (00:03:34) dollar economic engine on top of a (00:03:37) platform and have these incentives to (00:03:39) override those principles, I'm just not (00:03:42) sure that they have earned our trust. M. (00:03:47) So the issue then, Zoe, not so much (00:03:49) necessarily about ads right now, but (00:03:52) more about what it means in the future (00:03:55) and potentially manipulating people. (00:03:59) >> Exactly. And I think about social media. (00:04:02) I think that, you know, social media (00:04:04) started out in a way that felt very (00:04:06) innocuous. Facebook also had principles (00:04:10) about how it would and wouldn't use user (00:04:12) data and it all looked very fine and (00:04:15) good. But over time once they built that (00:04:17) advertising engine once that engine (00:04:20) started bringing in billions of dollars (00:04:22) of profit you know the company really (00:04:25) changed and the consequences for society (00:04:28) and for people's lives I believe were (00:04:30) pretty dire. You know, I know that in (00:04:33) Australia this is top of mind, but of (00:04:35) course there are, you know, issues for (00:04:38) kids, teenage eating disorders, (00:04:40) political polarization, (00:04:42) fake news. I mean, I don't have to I (00:04:45) don't have to tell you all what we went (00:04:46) through. (00:04:47) >> Yeah. So, is there any are there any (00:04:49) solutions here to get on top of this (00:04:51) before the horse bolts? Are there any (00:04:53) mitigation strategies that can make it (00:04:55) safer in your opinion? (00:04:58) >> Absolutely. I think that one of my big (00:05:00) frustrations is that the industry as a (00:05:03) whole is just being deeply uncreative. I (00:05:06) think that you know these are executives (00:05:08) who are promising to transform the world (00:05:10) entirely and cure cancer and build super (00:05:13) intelligence and transform the economy. (00:05:15) I think they can bring a little bit more (00:05:17) creativity to funding AI in a way that (00:05:21) promotes access while not descending (00:05:24) into a form of manipulation and (00:05:26) surveillance and control. And one idea, (00:05:28) for example, is to think about ways of (00:05:31) cross-subsidizing (00:05:32) these uh these uses so that people who (00:05:35) can't pay are subsidized by people who (00:05:38) are benefiting from these tools and (00:05:41) potentially contributing to layoffs in (00:05:44) the economy and so forth. H you would (00:05:46) have heard I'm sure uh the anthropic CEO (00:05:49) uh speaking uh this week in India and he (00:05:52) was saying only it's only a number of (00:05:54) years for AI models to surpass cognitive (00:05:57) abilities for most human things. That's (00:06:00) a pretty dire thing to say. I think if (00:06:03) you're a human he also had more to say. (00:06:06) >> We're increasingly close to what I've (00:06:08) called a country of geniuses in a data (00:06:10) center. a set of AI agents that are more (00:06:13) capable than most humans at most things (00:06:15) and can coordinate at superhuman speed. (00:06:20) >> Zoe, do you think that's where we're (00:06:22) headed? (00:06:24) >> I think that AI tools are enormously (00:06:27) capable and I don't know that we get so (00:06:30) much from fear-mongering or even (00:06:32) necessarily that we get so much from (00:06:34) comparing them to humans. I think the (00:06:37) most valuable thing for us to do as a (00:06:39) society is understand that this is a (00:06:42) powerful new technology, a powerful new (00:06:44) force that has vast uh vast consequences (00:06:48) for production and the economy and also (00:06:51) potentially vast consequences on our (00:06:54) social fabric and on our minds. And I (00:06:58) believe that we need to reckon with that (00:07:00) together and try to understand what we (00:07:04) want from this technology before we (00:07:06) sleepwalk into a situation that we'll (00:07:08) regret. (00:07:09) >> Uh Zoe, really quickly before I let you (00:07:11) go, where to now for you? (00:07:15) >> Well, you know, I also write poetry. (00:07:18) I'll work on some of that. And I want to (00:07:20) work on getting some new ideas, some (00:07:22) creative and imaginative ideas into the (00:07:26) public so that we can figure out a way (00:07:28) to live with AI in a way that works for (00:07:30) everyone. (00:07:31) >> We look forward to your next steps very (00:07:32) much, Zoe Hitik. Many, thanks for your (00:07:34) time. (00:07:35) >> Thank you.