How AI Is Cracking the Case on Medical Mysteries

[00:00:03] Welcome to Tech News Briefing. It's Tuesday, August 18th. I'm Belle Lin, a reporter for the Wall Street Journal Leadership Institute. A new AI scandal seems to be engulfing the book publishing industry every month. It's become harder than ever for everyone involved to figure out where and when it's okay to use AI, and it's raising questions over whether the book industry will even survive.

[00:00:29] Then, patients with rare diseases can spend years searching for a diagnosis. But now, AI is helping some of them get answers faster. We delve into how it's letting patients, doctors, and nurses crack the case on medical mysteries.

[00:00:45] But first, not long after scoring a multi-million dollar book deal, One Novelist agents suddenly pulled the plug. Why? Because they couldn't verify that the book had been entirely written by their client, and without the help of AI. That's becoming an increasingly common scenario as AI scandals and AI slop books sweep across the literary world.

[00:01:15] WSJ reporter Anna Silman is here now to explain what's going on in book publishing, and why AI may be an inevitable part of its future. Anna, what are slop books, and why are there so many of them out there?

[00:01:29] So, slop books refers to online books that have been written largely by an LLM, and often they contain material that is either false or largely pulled from, say, Wikipedia, and they are designed to look like real books to the people buying them. You know, there are companies and authors that just churn out hundreds of books designed to kind of game the algorithm on any given topic.

[00:01:59] They might come up with a title that looks similar to a title that's already in publication to try and deceive you to clicking it and buying it. But writers are using AI in all sorts of ways. Some of them are using it for help with research, for help with outlining, for help with coming up with inspiration. So, it's a gradient of use, although it can be a bit of a slippery slope.

[00:02:25] And usually when an author writes something, an agent or someone like an agent would get involved. What are agents doing about AI in writing? I spoke to a number of agents for this piece, and they said they're having a tough time. They're seeing a real uptick in the number of queries that they're getting, which they attribute to writers being able to mass spam multiple agents using AI. They are also receiving manuscripts that have clear hallmarks of AI use.

[00:02:55] You know, M dashes and the it's not X, it's Y construction. There are other things as well, specific adjectives and nouns that come up more commonly, specific sentence constructions. But the technology is evolving really fast, and the things that we can spot as hallmarks of AI use now might change very well with the next models that come out. So, agents, they're trying to act as a sort of filtration system, but it's not always easy.

[00:03:25] And every agent seems to approach it differently. Some editors use tools like Pangram to try and detect AI use. Others are telling their authors, please be honest with me, and are drawing a hard line and saying any AI use is not something that I want to work with. And others just aren't asking at all. So it's really still something that's being figured out as everyone goes along.

[00:03:49] When it comes to drawing that hard line, have any of the big five publishers like Penguin Random House or HarperCollins actually done that? Have they completely rejected AI use? The big five publishers have, by and large, been reluctant to make sweeping statements about AI use. They want works that are original authorship, meaning that the text is written by a human.

[00:04:13] But they've stopped short of putting any restrictions on how AI could be used in other elements of the writing process. And I think that's largely because they are waiting to see how the norms around this technology evolve and things are changing so fast. What about authors governing bodies or any other groups out there that might want to try and put a stop to this or put some rules around it? Has that started to happen?

[00:04:40] Yeah, the Authors Guild, a professional organization that lobbies and advocates on behalf of writers, has created a human authored certification. So you can go online and sign an attestation that you wrote your book wholly without the use of AI. And then you can get this trademark sticker that says human authored and put it on your book. It hasn't been super widely adopted yet.

[00:05:04] And some of the people I spoke to felt that using anything that requires an honor system doesn't necessarily have enough teeth to combat a problem of this scale. And this also makes me wonder, what about the readers? Do we want to read AI-generated books? What do we know about that? I think the consensus right now is that we don't want to read AI-generated books.

[00:05:27] There's been a lot of public outcry and backlash when authors have been alleged to have used AI in the writing process. However, that is also changing. And there are some genres like fantasy and sci-fi where readers read a lot of books a year. They can be quite formulaic. And there are some companies and some writers experimenting with using AI to help generate some of those stories.

[00:05:54] So that is a norm that could evolve as well. That was WSJ reporter Anna Silman. What do you think AI will change about the book publishing industry? Would you read a book written by AI? If you're a listener on Spotify, leave us a comment with your thoughts. Coming up, AI is offering new hope to patients with rare diseases. We'll get into how the technology is being used to help identify undiagnosed medical conditions after the break.

[00:06:32] AI has given people a sudden wealth of medical information, for better or worse. But it turns out that one thing AI is particularly good at is flagging potential rare and hard-to-diagnose diseases, which may otherwise go undetected for years because doctors don't often see them. WSJ reporter Alex Janin joins us now to talk about what makes AI capable of spotting rare diseases

[00:07:00] and the risks involved in trusting chatbots with medical data. Alex, I want to start exactly where you do in your story, and that's with a woman named Rachel Hinken. Tell me about her. So Rachel, she had wondered for a long time why her son Oliver seemed to be missing key growth milestones as he was growing up. He walked and talked a little bit later than most kids do,

[00:07:27] and she took him to a geneticist between the ages of two and five. She felt like she was largely shrugged off by doctors who said, Oliver seems fine. He'll catch up. He seems okay. And she said, you know what? I'm just going to investigate this on my own. So she uploaded his photo into an AI-powered app called Face2Gene.

[00:07:52] The app uses facial recognition technology to identify potential rare genetic conditions. It came up with a strong match for a condition that goes by the acronym TERPS. And she did genetic testing, and that confirmed the AI's diagnosis. And that helped explain some of the milestones that he had been missing. Wow, that's really powerful. There's research that indicates this isn't just limited to Rachel's story.

[00:08:21] Can you walk me through that? Yeah, that's right. There's a growing body of research. Some of the studies that are out there are funded by companies that have financial stakes in some of these AI tools. But there was one really interesting one that caught my eye, in which researchers took 90 complex rare disease cases that had already been solved and diagnosed and ran them through AI chatbots and tested their ability to suggest the right diagnosis.

[00:08:50] And the chatbots got the diagnoses right in 13% of cases for one of the chatbots and 10% of cases for the other. And compared to human doctors that had reviewed those cases by looking at patients' medical records, those doctors only got it right in about 5.6% of cases. So AI actually did a little better there. We don't always see that same data across the board,

[00:09:16] but there is a growing body of evidence that suggests that these tools may be uniquely well-suited in cases of rare disease diagnosis specifically. So that makes me wonder what it is about AI that makes it good at solving these medical mysteries. Apparently, these tools are good at what researchers and clinicians call pattern matching.

[00:09:41] So looking for associations between both physical features like Rachel did when she uploaded a picture of her son's face or even physical features in a photo of medical imaging or a picture of a tumor and making associations between those and then words in medical literature or case reports and drawing from all of that data that exists out there.

[00:10:05] And it makes sense, right, because in rare disease, clinicians may never have even seen a case of one of these conditions. And so they may not remember from medical school that one obscure condition that they learned about. It is important to caveat that because we're talking about rare disease, there's often less underlying data for the AI to use to draw from. And so there are experts who are working on digitizing cases related to rare disease

[00:10:33] to make sure that the AI has reliable data and enough data to actually pull from. Right. That brings me to my next question, which is around the limitations that these tools might have, including, as you mentioned, the sort of shortage of data to make them really robust. What are the other problems or issues these tools might have? Experts always caution that hallucinations are still possible and they see them all the time. One doctor I spoke to said that even though she finds face-to-gene really helpful in practice,

[00:11:04] sometimes when she's tired and she uploads her own picture into face-to-gene, it will say that she has a likelihood of having a rare genetic disorder that's life-threatening. And she definitely doesn't have that disorder. So just because it's making associations doesn't mean it's always right. You know, AI can seem very authoritative and it is common for patients and even clinicians to see AI come up with something and just immediately trust that.

[00:11:33] And maybe that sends them on a rabbit hole. They might end up spending a lot of time, maybe even a lot of money getting tested for conditions they don't have. And what about the risks to me as an individual to put my personal health data into an AI tool? Experts have cautioned that when we upload our medical data into most AI chatbots, we are forfeiting our HIPAA rights. So that is an important privacy consideration.

[00:12:04] That was WSJ reporter Alex Jannan. And that's it for Tech News Briefing. If you're a listener on Spotify, be sure to leave us a comment. Today's show was produced by Julie Chang with supervising producer Katie Ferguson and deputy editor Chris Sinsley. Logging off, I'm Belle Lin, a reporter for the Wall Street Journal Leadership Institute. We'll be back later this morning with TNB Tech Minute. Thanks for listening.