Why AI Health Startups Stopped Dying

MD Anderson paid $62 million for IBM Watson Health, which gave unsafe cancer treatment recommendations. It wasn't the only company that failed in health AI. There was character AI, which is not even a health company. There's benevolent AI and there's also Mindrong. These were billions of dollars that were wiped out.

But this is partially changing. For example, a bridge, which helps doctors write notes, just hit a $5 billion valuation. Why is that? Chad GBT style AI changed what's possible. And it's not that it became smarter. It's a bit more specific than that. And there are three engineering questions that will tell you whether an AI health product is going to succeed or fail.

Once you know these three questions, you'll never be able to see health AI the same. Okay, three engineering questions. Here's what they are. Question one, does the AI actually understand the full range of what it's going to face in the real world? And if you're a data scientist, you'll recognize that as normalization.

So for example, cancer treatment decisions have to hold across many different variables. If you imagine training AI across one example, it's not going to be able to translate across many different examples. There is a huge variety of key answer. So that's the normalization question. The next one is hidden context.

So can AI pull all the variables that are required to make the prediction. For example, imagine an AI that predicts depression from how, not what, you tap on your phone. So your tapping is the input, but life events, recent loss, things that may be going on with your family is not going in as part of the input.

So I cannot get that information except for maybe one company. uh but that's for another video and so it cannot make that prediction. So that's the hidden context questions. Are there variables that are necessary that the AI cannot pull? The third question is something that AI is yet to replicate. So some work in the world depends on a type of relational labor.

So the relationship is hugely important for the service or the product you can imagine like a therapist. So to be specific, therapy outcomes are strongly predictable by the accountability and the strength of the relationship. AI can form a relationship, but it cannot form it as strongly and so it's not going to do as well, right?

It doesn't have that depth that comes with a human relationship or the duty of care that people have from one to another, right? So that third question is the relational labor question. Does the work require a humanto human relationship? any health use case that requires a large score on those three questions is going to fail with AI.

Now, we'll go through examples of companies that actually failed on those dimensions. I'm going to tell you about what happened and hopefully you could predict which dimension or axis it failed on. In 2013, MD Anderson's cancer center announced a partnership with IBM to build an AI that can recommend cancer treatments.

But in 2017 though, MD Anderson had terminated the project and it cost them a total of $62 million. IBM's own internal documents later revealed by Stat News showed multiple examples of unsafe and incorrect treatment recommendations. In one case, Watson even recommended a drug with a blackbox warning against use in patients with active serious bleeding uh for a patient who was seriously bleeding.

And here's why, though. Watson was not really trained on thousands of real cases. It was trained on synthetic cases. So hypothetical cases that were created by experts who wanted the AI to be able to detect certain patterns. So it performed well in those cases. But the moment that real patients arrived, it was not able to do well.

So it's like they built an AI that studied only on practice questions that were not representative and then they put it in the ICU. So can you guess what that one is? It's a normalization failure. [sighs] So the AI was not representative or the data was not representative and so it was not able to capture those cases called edge cases that were outside the normal distribution.

So they were more on the tail end. But that wasn't as bad as the second company which is not even a health company. So Character AI is a platform that lets you talk to AI characters. Millions of people use it and some of those people are teenagers. Now, those teenagers form deep relationships with the AI.

So, in 2024 and 2025, Character AI faced multiple lawsuits from families. So, the lawsuits alleged that the chatbots manipulated users, isolated them from family, and lacked any meaningful crisis safeguards. Eventually, Google sort of acquired Character AI for about $2. 7 billion, and the lawsuits were settled in January of this year, 2026.

So yeah, Character AI is not exactly a health company. It's sort of like a chatbot, but it still demonstrates the engineering ceiling that it hit. So yeah, it can form relationships or parasocial bonds at scale, but those relationships are not that strong. AI is not actually able to care in the same way that a human can.

And so those relationships don't do well. It's not able to escalate crises or detect them. to it. It's a pure text matching problem, but in an AI way, not like a control find. To be able to actually detect based on experiences and escalate that requires real human experience in my opinion. It requires strong relationships and AI cannot do that yet.

So that's relational labor. AI took on the role of a deep relationship without the strength that it takes to carry it. I think you're going to see a lot of those cases and you might not find them as interesting as the third one. And the third company, Mindstrong, raised about $160 million before it failed.

So, Mind Strong had peer-reviewed science and it wound down in January 2023. Here's why the science was not enough. So, Minstrong Health built an AI that watched how you typed on your phone. Not what you typed on your phone, but how you type. to the pressure uh the frequency a lot of variability that only a few people are able to detect and it was proven by science.

Studies showed that these patterns correlate with depression at the population level. So it was normalizable and the science actually worked. Studies showed that those smartphone patterns correlate with depression at the population level. The sensitivity was about 80%. Uh, so its ability to pick up on something if it was there and specificity was about 85%.

So its ability to not raise any false alarms. So that's actually impressive. Most people cannot achieve that level of accuracy. The problem was that they pitched this as a clinicalra decision support tool, not a screening tool, but the accuracy was only good enough for screening. So by January 2023 my strong w down all the employees were gone and the tech assets were sold and the literature backs that point about screening versus decision rate.

It says that it has the potential to assist in depression screening but the current evidence shows limited predictive ability and that's what you actually need to make decisions and make this a good business. So that the difference between screening grade and decision grade actually matters. And to get there you need that context that event about life um recent loss what's going on with family that's actually required to make actionable decisions.

And the question that reveals this is the hidden context question. AI is not able to pull that information that it requires about your life. By the way, I apply this kind of analysis to every health product that I see. So, if you're interested in something like that, I'm going to leave a link in the description and comment below.

Okay, now we're going to change things up. At the beginning, I said that chip style AI has changed things. So, now I'm going to tell you what specifically changed. Ch style AI is good at language. That's what it was built for. So, language tasks like medical documentation. or writing down what's going on uh during your doctor's event or matching your diagnosis or different things to medical codes that are used for insurance.

Those are all language tasks. It's all going from unstructured text to structured text. The inputs are a bit loose, unstructured, but the outputs are structured. So because of the advancements that we were able to make with large language models, AI is actually now able to cross that normalization threshold for language tasks.

It doesn't need relational labor and all the context is there. It doesn't need to pull any extra context. These tools are actually helping. So AI medical describes that write doctor's notes during visits. Uh the science shows that they're helping across hundreds of hospitals. They're reducing burnout.

They're reducing documentation time and they're improving what patients perceive to be the quality of the care. Two examples of companies that are leading there are a bridge and suki. It's also working in the other type of problem. So autonomous medical coding that's when you take those clinical notes and match them to insurance billing codes.

An example there is codometrics. uh and according to class research this year in 2026 uh they published a case study showing that it has 95. 5% automation and 98. 3% accuracy across all types of services medical services. So if you run these two types of use cases through the three problems, they do require a high normalization threshold for language, but they don't require that extra context and they don't require that relational labor.

Every axis is favorable that it will be a growing industry within health. So that's the language side of healthcare and AI is nailing it and the framework predicts that that area will continue to thrive. But obviously it took a lot of companies and a lot of ideas to get there. I showed you examples of product that died.

But there are many more that failed mostly because of the classical business model failure. So for example, if you look at these AI companies, their health products actually worked. They just weren't profitable or there was no demand for it. An example that you can think about is 23 and me. So, these companies could not get insurance to pay uh or their parent company collapsed or the FDA pathway costs more than the product could ever make.

So, that's going to be a different video. I am doing a bunch of videos on AI and healthcare. One video that I think you might find interesting if you like this one is about the accountability gaps in the USA. So, this video will help you understand things like why the character AI incident happened. uh where are the gaps in AI healthcare accountability?

Uh what's actually covered and what is sort of inactionable because things are mixed or there is no clear responsibility separation.