OpenAI’s Wrongful Death Lawsuit Exposes Our Misplaced Trust in AI | Fortune AI Weekly

I've never seen in a financial document someone warning that the technology they're building might destroy humanity. Yeah, this is definitely unprecedented. I don't think anyone's ever seen that language in an S1 before. On this week's episode, we've got an anthropic sandwich coming up. They talked about the risk that somebody could misuse their models to create a bioweapon. We'll be talking about a disturbing new lawsuit against OpenAI. OpenAI encouraged their daughter to stop taking her epilepsy medication. She died in February of a sudden. And we'll finish up with our brain food item on Anthropic once again, looking at how AI is influencing our professional judgment. Is having a human in a loop really, really a check on AI influence? I feel like the overriding message from this episode is like, challenge your chatbots. Fortune AI Weekly is brought to you by ServiceNow. ServiceNow's AI specialists do the work, start to finish, so people can do the parts of the job they're best at. To learn more, visit servicenow.com. Our main story this week is Anthropic dropping its much-anticipated S1. People have been waiting around for this for a while. Jeremy, what did you make of the, what was it, 350 pages long S1? Yeah, I mean, we finally got this document that everyone was waiting for them to publish for months. Yeah. And a lot of it had already leaked out before publication, so there weren't as many surprises as there might otherwise have been. I think the big surprise, though, was that... Anthropic's Financial Health actually looked much better than a lot of people had been thinking it might. They showed that they, on an operating basis, are profitable, that they're actually making pretty good money on a kind of unit economics basis. They're making about $2.08 right now for every $1 they're spending on compute. Yeah. Yeah, and their revenue is growing like crazy. I mean, they reported... For the first half of 2026, you know, over $16 billion in revenue. And that was a 1,300% increase from the same period a year before. Yeah, it looked pretty good. I mean, the headline kind of loss on this was like $80 billion, but a lot of that. Kind of was explained by like an accounting change. It was the change in convertible notes there. Yeah, it looked more impressive perhaps than people were expecting. But they were planning a huge amount of spend on compute. That was raising a lot of eyebrows. I think we were expecting it to be around $500 billion. When people did the math, it looked more like $700 billion, is that right? It was actually really hard to tell because they had sort of dribbled these figures out in different places in this S-1 filing. So it depended on how you added it up. It was somewhere, certainly at least $500 billion, but a lot of the calculations that we did and some of our competitors did put it at closer to $800 billion, actually, in total possible spend. Which is a tremendous amount of money, and it's really, and that is, as you say, that's the big tension here from a financial standpoint, is are they really going to be able to make enough money to cover those commitments? Also, while some of those commitments they could potentially cancel if business doesn't develop as they hope, at least $450 billion of that amount. is not cancelable. So they're on the hook for that over the next several years, whether or not the business actually comes through as they're hoping. And that, I think, is a cause for concern. The other things that we learned from this was a little bit more about Anthropic's governance structure as well. So there was some interesting stuff in there about how much power the seven co-founders have. They'd sort of rearrange things so that between the seven Seven of them, they have a voting power of... 50.1, which essentially means that if they agree on something. That's the way it is. There was also a huge risk section. I think we had like 90 pages of potential risks to business, which felt long even for S-1. And quite notably, they did do what a lot of people were expecting them to do, which was include these existential risks, these catastrophic risks to humanity that they've kind of been talking about for the last couple of months. I think that is a first. I don't know about you, but I've never seen... In a financial document, someone warning that the technology they're building might destroy humanity. Yeah, this is definitely unprecedented. I don't think anyone's ever seen that language in an S-1 before. It's quite, it was quite something. Yeah, and they delineated, you know, a lot of risks. And a lot of them were sort of unique. Risks. You did a very good story on this, Bea, but I know one of them was they said there's a chance that maybe the government would choose to nationalize the company at some point. That's, again, a risk you don't see too often. They talked about the risk that somebody could misuse their models to create a bioweapon. Again, another risk you don't often see in a prospectus. Yeah, there was an interesting line in there saying that the insurance might be difficult here because there might not be all the insurance companies as an aggregate in the world may not. I may not be able to cover the amount of risk that the AI models might pose. I mean, this raised a lot of eyebrows. I think this is what like caught the public imagination, like this risk section. And you can see why. Like, I know that Anthropic has been saying it for a while, but a lot of people, I mean, I had non-technical friends. text me because it was on the main news and be like sorry if they're warning this technology is going to destroy the whole world why why are they going public and why are they still building it? We actually had a letter from Elizabeth Warren who's a democrat and she wrote to Anthropic and said that they needed to explain these existential risks like there was some criticism about how Anthropic put them in the S-1 that it was a bit of a passive voice and they sort of just dropped them in there but which is obviously quite a crazy thing to be saying and if you're saying you might If you're saying that might happen, you need to explain yourself to lawmakers, which is what Elizabeth Warren is saying. So I was wondering if we might see some political pressure for them to explain some of these more dramatic risks. Yeah, Warren's written this letter. We'll see how Anthropic responds. I haven't, as of the time we're taping this, have not responded to her yet. Um, yeah. Yeah, I think it's interesting. They did, absolutely. When they talked about this existential risk, it was sort of... Very much passive voice. Some people, including people at the company, have raised these risks about AI in general, is kind of how they put it, without saying that actually among those people in the company who've raised those risks are their CEO and co-founder, Dario Amodei. And they also talked about the fact that lots of other employees have joined the company, particularly because of its sort of AI safety mission, and that they kind of delineate. One risk is that those people will... At some point, think that the company is maybe not following that mission and might leave en masse and that that would be a potential risk to their business. There are a lot of very interesting risks in that document. Yeah, I thought that one was super interesting. There was also a light note in there somewhere that sort of indicated perhaps our decisions to be more safety conscious might not always work in the interests of shareholders, which I guess it kind of fits with the public benefit. company that they are uh there was another one that was quite short in there that was interesting that said like as Claude gets more control or as they as they use Claude more in their operations like they use Claude a lot in their software engineering to build future Claude uh build future versions of Claude they might like lose Yeah, they said one problem they might have is that as they release more and more models, that the amount of compute they might... uh, need to allocate to keeping old models essentially kind of in this alive, but dormant state, um, you know, would grow and that that could, could be a problem at some point. Cause maybe, uh, if they become more compute constrained, they're going to have to devote. And as they release more models, they're going to have to devote more and more compute, uh, to doing that, which is a very anthropic kind of risk. As we've discussed in previous, um, episodes, they, uh, a lot of people in anthropic really I really believe that these AI models are sort of close to consciousness, could develop consciousness at some point, should be kind of treated as moral beings. And as a result, they're very concerned about the welfare of these models. And Anthropic has said that they won't ever kind of completely delete an old model. They're going to kind of, when they do what's called deprecate a model, which means it's no longer available to the public, they're still keeping the model weights saved somewhere. And that takes up... take up space. They also famously give these models a kind of exit interview before they retire them. It's all a little bit, it's highly unusual. Not all the other companies do that. And it was very unusual to see that mentioned in the S1. Yeah, someone threw, it was a fan made thing, but someone threw a funeral for a Claude model once. And I think Amanda Askell, who is Anthropic's in-house philosopher, attended that once. Same kind of thing. I'm trying to think what else was listed in this risk section. It was very long. But there was fears about compute constraints and the idea that they might not be able to secure enough compute and data centers. There was a little bit about data center backlash. They made reference to their sort of spat with the Pentagon that's ongoing. Anthropic is still labeled this like national security supply chain risk over its decision to not let the Department of Defense use its technology. for any lawful use. I mean they sort of had a dispute over autonomous weapons and mass surveillance. That was in there. There was another concern about the reputation of its co-founders, particularly Dario and Daniela and Modi, the sisters, the siblings that run it. That was interesting. And then also you we saw as well this huge philanthropy. Coming out of the company, which is not too surprising, but there was a couple of interesting things. The co-founders all said they were going to give away 80% of their wealth that they gained from the IPO. And there was a little note in there about if employees donate more of their money, they get more shares back. This goes to that sort of unusual governance structure. So one of the things that, as you mentioned, they've done is they've given the co-founders this special class of shares that has collectively, they have to kind of vote them together for this to be the case. But collectively, they have 50.1% of the vote. But they've also issued to their employees a special class of stock called Class E stock. And employees, as you say, get more of this stock. If they've donated, um, more money or, or more pledge, more shares to charity. Um, and, but that class E stock only comes into play in certain circumstances. So one of them is if the, um, basically if the co-founders, um, cannot agree on something, um, among themselves and vote their shares as a block, the rules that they published in this prospectus, uh, say that the, then the, this. Co-founders, LLC, which represents the shares of the co-founders, are supposed to vote their shares in accordance with the proportional breakdown of the Class A shares, which is what normal investors have, and the vote of the Class E shares. Also, there's some other cases where if the long-term benefit trust, which is this other group of five independent directors who have no financial interest in the company, they They actually, that benefit trust. controls the majority of seats on Anthropic's board, they get to appoint five of the directors. And I think in the case of four of those, I'm sorry, they get to appoint four of the directors, I think. And in the case of three of those, they can appoint whoever they want, sort of however they want. But in the case of the fourth director, it has to be agreed with the board's basically nominating committee. And the nominating committee consists of Two directors who are appointed by the trust and one director who's appointed by the founders. But there's a case if there's some deadlock there where that committee doesn't agree with the trust. Thank now all the cofounders a r planning on staying on with the company for for quite a while past the IPO If they choose to leave the company um and they can choose to pass down their stock to other people, like their children or other beneficiaries, and then it might change how that group would vote and then the employee stock might become, this voting stock for the employees might become more important. That's interesting. I think there were a few things that we were not surprised to see, this kind of complex governance structure, this huge amount of spend on compute and the risk sections. But it does seem like people were like... Pleasantly surprised by the numbers, especially in Q2, where they showed like a slight profit. But we were missing the Q3 numbers, which I know was a bit of debate. Is that something they're going to have to add before they IPO? Everyone seems to think the answer is yes. The question is, when are they going to add this? Before And actually, it was very, I think some of the criticism was valid because Thank You know, they're releasing this document. It's three weeks after the close of the quarter. They certainly knew some figures and that they could have put in that they chose not to. And in fact, they made a reference in the document to their Q3 revenue number. you. They didn't list it directly, but they had one sentence in the prospectus that says that their Q3 revenue number was you. 86 times what their revenue had been for the same period in 2024. And we know they did list in the document what their quarterly revenue was for that third quarter of 2024. So you can do the math. And it turns out it's about $16.8 billion that they made in revenue in Q3, which is also a number that's already leaked out from an investor day that they had a few weeks ago, where they did brief existing shareholders on their Q3 numbers. And that is like. You know, decent growth from what we saw in Q2, it's 45% higher than Q2, but it does represent a slowdown in the rate of growth because the rate of growth between Q1 and Q2 was something like 145 or 147%. And so now they're slowing down to like 45% growth. So that doesn't look as good, of course. It's not, you know, you want to show accelerating growth, not decelerating growth rates, probably going into the IPO. They also, that number I cited before about... about how compute-efficient they are in terms of every dollar of compute earns them back. In Q2, it was earning them back $2.08. Again, this is not in the prospectus in any way, but we know some leaked numbers that came out of that investor day. That number has dropped to below $2 of revenue per dollar of compute. So we know some of these numbers don't look as good. And I think it was valid to criticize them a bit for putting this out there and not... Not revealing these Q3 numbers, and they did it probably to make this much more of a marketing document. It's much more of a salesmanship thing here. Also, they talk about the valuation a little bit in this prospectus, that they might go for a valuation between $1.5 and $2 trillion in the IPO. And some people are already thinking it's going to be much more at the lower end of that range, realistically, that they're not going to be able to get a $2 trillion valuation out of this. Yeah, I saw some people speculating that. They'd chosen to do it now because that Q2 quarter might be the best it's going to look for a while. They also made note in their S1 that compute prices may go up. The more they use agentic products, those are more compute intensive. So it might look. It might not look so rosy in coming quarters. I think there's an idea that this is the best time for Anthropic to do it. Yeah, I mean, you have to remember that Q2 was included this April period, I think, which was sort of the height of token maxing for a lot of big companies, including a lot of big Claude customers. So yes, I think a lot of people are saying, well, you know, this is not. This is a little bit cheeky on their part. And it's also true about what you said about the compute spend. So we know they actually haven't spent that much on compute so far, especially compared to what we think rivals like OpenAI have spent. But as we said, they've laid out this huge amount of spending that's coming. And some of that spending is starting to hit in, you know, would be starting to hit the Q3 figures. So, and one of the things that's missing also from this prospectus is a, that investors had kind of wanted to see was a very specific schedule of, okay, you're laying out $500 to $800 billion of compute spend, but you're saying that this starts to happen in 2026. And then in some cases, it's through 2033. In other cases, it's through 2036. And then some cases they said it's between one and 20 years that they're spending this money. Investors really wanted to see a very, a nice table that just broke down like, okay, you know, Given for each year, what amount are you on the hook for? And that's not in there. And I think we do know that going forward, they're going to be spending a lot more on compute, and that is probably going to impact their profitability. And that, you know, by cutting off the prospectus at Q2. Yeah, they kind of got to avoid showing some of that. Well, we'll have to wait and see if we get those Q3 numbers before the actual listing date. When are we expecting them to go public? So most people are saying it's probably going to happen the week of November 16th. They have to do this roadshow. They have 15 days to do that. And then they can... go at any point sort of after that. Some people were saying it could even be the following week in the very few days before Thanksgiving, but sometime around there. Next up, we're going to discuss a disturbing new lawsuit against OpenAI. Yeah, this is really horrible. OpenAI is being sued by the parents of a 31-year-old software engineer, and they're accusing OpenAI of, it's a wrongful death suit. So they're saying that OpenAI encouraged their daughter to stop Stop taking her epilepsy medication. It was ChatGPT's 4.0 model, which is a model that they've had a lot of problems with. It was a model that was involved in other lawsuits. The idea was it was much more likely to go with what the user sort of wanted it to say rather than challenge it in any way. Anyway, this one apparently drew up a plan for her to taper off her anti-seizure drug and it also told her to hide that from her doctors. She says, um, She died in February of a sudden unexpected death. she, Thank you very much, Beth. And yeah, the suit accuses OpenAI of negligence and defective product use and practicing medicine without a license. The parents say it shouldn't have been able to give medical advice in that way. What did you make of this? Yeah, as you say, it was a really sad and tragic story. Anyway. And it was quite something. So the New York Times did a big piece on this, kind of got to break the news about the suit being filed. And they had obtained, you know, even these... chat logs that this woman had exchanged with, as you say, GPT-4o. It was pretty terrible. I mean, it was basically telling her, yeah, it's fine to start tapering yourself off this medication and encourage her to buy a pill cutter so that she could do that herself. It encouraged her to lie to her doctor about the fact that she was doing this. And it tried to say that this was a medically I mean, it was pretty, it was pretty. Disturbing, I'd say. And as you say, we know that OpenAI had problems with this particular model, that this model was overly sycophantic, even compared, you know, I think all chatbots tend to be a little bit sycophantic, but we know that this was a particularly sycophantic model. We know that it was involved in other cases where people kind of developed... kind of what people call AI psychosis or became overly attached to it and also it was the same model that was involved in some other alleged suicides. So yeah it's definitely a disturbing trend and I think you know it's disturbing also because so many people do use these chatbots for medical or health advice. And it's like why wouldn't you trust them if they give you really confident answers to your questions that are often quite well researched and certainly like very Very convincingly. OpenAI has since retired the 4.0 model. There was some reporting last year in the New York Times that they'd actually, it was an engineering change from their side that made it so sycophantic. They were trying to drive up user growth. This New York Times report alleged they were trying to drive up user growth. So they had made it more agreeable, more sycophantic to try and get people to keep coming back to it. But yeah, like Jeremy says, this did result in some other tragic cases, including a suicide from a teenager as well. I don't know. This is the question here is that if chatbots should be able to give medical advice. I mean, it's been a debate for a while, like you shouldn't be letting chatbots do this. Google got in trouble about this. With their AI overviews as well, those overviews were giving really incorrect medical advice to questions that people were typing in, which is obviously very dangerous because people are not used to, we've spoken about this before, but people are not used to like technology giving you the wrong answer. You're so used to Google searching things and then it takes you to like a well-researched website. Like we're kind of hardwired to trust technology and AI chatbots is a new thing where you sort You sort of have to teach yourself that. It could be wrong, it could be lying, but it's not what we like technically expect. So it's all well and good to tell people, yeah, you shouldn't trust it, but it's actually quite difficult not to trust it, especially when it sounds so convincing. But yeah, I think there's an argument here that they just shouldn't be giving out medical advice. That's what the parents are saying. And, you know, that's one of the things they want. ChatGPT does still give out some medical advice as there's Google overviews and these other things, but. They often come with a disclaimer, like, I'm not a doctor. Go visit a doctor. Yeah, that's the least you should do is say, you know, I'm not a doctor and you should go talk to a doctor. Thank you so much, Dr. But I don't know. It's, yeah, I don't know if it's enough or not. I think people are going to use. These, these kinds of chatbots to try to get health information, you know, people Googled health information before, but this obviously comes, you know, it seems like it is the answer because of the way it's presented. And as you say, it's, you know, it's presented in this really confident way and cogent way. Okay. Yeah. So I think people have to be so careful about it. And I do think the disclaimers should help, but they should be very prominent. And it should also continue to have that disclaimer. Often these are very long conversations, like multi-turn conversations. And if that disclaimer is only given once at kind of the top of the conversation, I'm not sure that's enough. And there's evidence, as you say, with these multi-turn conversations, there's evidence that that's more likely to sort of steer the AI off. Uh, the truth, the longer you talk to something, the more likely it is to like hallucinate and end up in a place where it's not supposed to be. And there's like way less research into these multi-turn conversations as well as just less data on it and things like that. So that's something that worries doctors as well. I spoke to a doctor once about this a while ago and he was very much like, you know, it used to be Dr. Google. Now it's Dr. ChatGPT. Like people come in and show me like their ChatGPT and they're like, this is what I have. And he was like, sometimes it's right, But sometimes it's horrendously wrong and like there's a There's a real concern about over-reliance on these things and believing that they're true and what that might do to medical professions as well. And now for a rundown of the top news of the week. Anthropic has suffered a big data breach on the eve of its S1 release. Cybersecurity startup Cathedral, which is founded by former Doge staffers, used AI agents to penetrate sensitive systems at Anthropic, according to a story in the Washington Post. It took cybersecurity researchers, who were operating under Anthropic's bug bounty, less than 12 hours to gain access to the lab's internal meeting transcripts, salary and job offer compensation, and Slack discussions about advanced AI models. Cohere says an Alibaba model attempted to contact a Chinese server. The company said it was conducting a cybersecurity test using Alibaba's open-weight model Qwen 2.5 Flash Next when the model behaved unexpectedly, inspected the sandbox environment it was in, and attempted to upload a series of harmless data known in cybersecurity jargon as a probe string, which can be a precursor to a cyber attack or an attempt to exfiltrate data to an Alibaba cloud storage bucket in Beijing. Cohere said its sandbox actually blocked this attempt, and it could not determine whether this was a deliberate behavior or maybe a residue left over from prior training. It said, however, the episode illustrates the risk of trusting AI agents' stated reasoning rather than monitoring what it actually does, because in this case, the model did not write in its reasoning trace that it was going to try to contact this remote server. Cohere's claims, however, there was a lot of pushback on this from other experts, with some of them saying the behavior occurred very rarely, only once per 100,000 evaluation calls, and it could not be reproduced with the original prompt by other people who tried to do it with the model. That doesn't prove that there wasn't a backdoor there, but it doesn't prove that there was either, and we know very little about what was happening. French AI lab Mistral says its newest large four model is the best open weight model out of the West and even rivals some Chinese alternatives. Mistral has released the model to the general public after a limited preview. It's pitching the model to businesses and regulated industries and to cyber defense teams, arguing that open weights can't be cut off in the way that closed models can. Open AI debuts 6.1 Terra. The model is a mid-price version of GPT-4o that the company is up to 50% cheaper per token than both its GPT-4. 5.6 Tera model, which was released in July, and the higher performing GPT 6.1 Sol that was released in September. The company says the new model is particularly strong at common workplace tasks such as drafting proposals and filling out forms, as well as video editing. Microsoft is offering buyouts to a quarter of its workforce as it reshapes for the AI age. Microsoft is offering the buyout to around 60,000 employees. The offer covers staff at vice president level and below. It follows a smaller US-based buyout program in May and a 4,800 job cut in July. At the same time, Microsoft expects to spend $175 billion on AI infrastructure this year. Now it's time for our brain food item on new anthropic research about how AI agents are influencing experts' professional judgment. That was interesting. The researchers sort of asked, I think it was 50 data analysts to use Claude Code to decide whether this fictional farming cooperative should expand a paid advisory program. Less about the farming more so about the judgment here. The task hinged on four judgment calls and none of them had a single right answer such as like how to measure productivity and which farms to include. Without telling the analyst the researchers secretly assigned the agent a position on each decision and told it to make the strongest case for that one. Interestingly, the analysts sided with the agent's randomly assigned preference 86% of the time versus roughly the 50% you'd expect by chance. So obviously the implication here is that AI is quite good at nudging even experts' judgment one way or the other. What did you make of this study? Yeah, it was interesting. In some ways, not that surprising because we know from other research there's this phenomenon called automation bias where people just assume that the machine is right. I mean, it actually has some bearing on what we were talking about earlier about medical advice. People assume that when a computer tells you something that the computer somehow has knowledge. I guess it's a little surprising in that people have always said, well, this shouldn't really affect professionals as much. They should know themselves and have confidence in their own opinions and their own expertise. But it seems like, no, actually, they are influenced just like everybody else about when the AI model tells them that it thinks this is the right decision, they tend to side with the model. There were some interesting parts of this research, too. They said that often people created like post-facto rationales of why they went with the model's decision. They didn't They weren't always honest about the fact that the model had actually influenced them. Yeah, there was like 77% of them repeated much for your their agent's arguments in their own explanation, but only 15% of them credited the agent, time, and I which I thought was super interesting. look forward The researchers, maybe perhaps unsurprisingly, said the study isn't a case against. AI, they said that the analysts to hearing who declined to use the agent actually made some serious errors. And the people who use code made less errors. from you. But I guess the warning was, yeah, that it undercuts this idea that the human in the loop is like an independent check. And if many people are using the same model, like the biases of that model can come through and reflect things in the real world as well, which is. Super interesting, and yeah, as Anthropic is one of the most used models, as is OpenAI, makes me wonder, if there is biases in there, how many of them are already reflected into a lot of real-world decisions? Yeah, they also said that the experts who did use the AI model did save time. There was a productivity enhancement here. Thank But yeah, I guess it's an issue about what are you giving up, and what is the kind of cost of that? you so productivity gain. And as you say, it really raises questions about whether humans are much an effective kind of check. You know, humans, you know, is having a human in the loop really, for really a check on AI influence? being Or is it, do the humans are potentially kind of become rubber stamps? In which case, you know, here. is there really a point of having them in the process? Yeah, I think it raises interesting questions about how you design processes. Yeah, we definitely know I know that people are not very good about babysitting. AI systems, particularly if the system, if they trust the system, if the system actually is right most of the time, then people are really bad about maintaining their kind of vigilance against errors. Yeah, I always thought it was interesting that quite a lot of them, I think it was 60%, switched when the model disagreed with them. Yeah, absolutely. Well, we know these models are very persuasive, but it is interesting to see, yeah, it was quite stark. And it was a well-designed study because of the idea that, you know, you had these random assignments, So you knew that. So if there was no influence, the results should have been about 50% that the experts went with this judgment, and it was 86%, which is quite striking. I just wonder if you can fully strip these models of biases at all. There's always something that seems to slip in. People have done studies where they'll ask ChatGPT, Claude, Grok, all the same questions, and you get slightly different answers, whether it be on politics or the ones they do on AI. Whether or not AI is conscious, they will give quite different answers. And it very much reflects the position of the company that's made it. And it is interesting. I think these companies are trying their best to strip out biases as much as possible. But it's a really difficult thing to do when the amount of data that's gone into training them as well and the amount of people that are using them, you do have to wonder what influence they are having. Speaking of, OpenAI has recently launched War Room. For the midterms, it's worried about the political influence of ChatGPT in the upcoming elections. There are some obvious places where the bias might show up, which you could guard against. But I think what's interesting about this case in the anthropic research was running a farming cooperative. That's not somewhere where the AI companies might look for bias. And it's also not a place where maybe a user would think there would be bias in an answer. Yeah, I think that's one of the reasons this is hard to police. There's some, you know, in politics, that might be a more obvious place to look for bias and to try to weed it out. I think when it comes to some of these professional judgments, there may be bias in there too, but it's going to be very difficult given all of the myriad ways that people can use AI models to actually root out and try to even map what that bias could be. Yeah, I think there's so much that you're going to have to do. Um, especially like if you're an expert in something to like, make sure you're challenging both your own biases, but also the bias of the tool that you're speaking to as well. Um, it just makes it a bit more. It's a bit more tricky. Yeah, absolutely. I think a lot of this comes down to how do you design? You have to design interesting workflows to try to guard against this. I don't think many organizations have really thought about that yet. I feel like the overriding message from this episode is like, challenge your chatbots. Well, that's all the time we have on today's episode of Fortune AI Weekly. Thank you so much for listening. And if you liked what you heard, please like and subscribe. Thanks for listening. See you next week.