You Ask, I Answer: How Not To Use Generative AI In Healthcare?

In today's episode, Amy asks, what advice do you have about how not to use generative AI, particularly for uh concerns of privacy and authenticity? There's so many ways to answer this question. Okay, first, don't use language models for tasks that are not language. That one you would think it'd be obvious, but it isn't because people the general public does not understand that language models are good at language, but they're not good at not language. People have a tendency to think of AI as this all knowing, all seeing oracle.

And a lot of that can be blamed on pop culture. A lot of that can be blamed on Hollywood and on Terminators and WALLI and Short Circuit and all those films that and TV shows that we grew up with where machines had these magical capabilities, commander data from Star Trek. There is no way that uh that system that we watched growing up would actually exist in that form with how today's AI works. There's a whole other tangent to go on, by the way. But we're gonna we're gonna give that a miss.

So use generative AI for what it's good at. So, for example, these tools are not great at generation, believe it or not. They need detailed prompts, lots of examples to to do a really good job. So you definitely want to not use them to just crank out generic content. And that's that's pretty easy.

You don't want to use them to try and do math. They're bad at math, they can't count. A language model under the hood is a word prediction machine. That's what it does. It predicts words.

And so if you're trying to get to pr to predict things that are not words, it's not gonna do a very good job. So the workaround for that is to have the tools write code, right? They because writing code, code is language, and then the code can do math. So that would be another thing. Don't use tools.

It's not that you shouldn't use AI for this. You should not use AI in an unsupervised manner for anything high risk. Right. So what do I mean by that? These tools are very good at things like image analysis.

They could take an image, an X-ray or a CT scan, and provide an analysis of it. You would not, under any sane circumstances, just hand that to a patient and say, hey, here's what the AI spit out. You've got this. It might be right. It might not be right, but that is a very high-risk situation where you want human review.

And this is a part of uh generative AI that I don't think people give enough thought to. Yes, it is capable of doing a lot of tasks very quickly and at very high quality. But for tasks where you need to where you have a level of risk, you need human review. So there may be fewer writers writing, but you may have more reviewers reviewing, right? Those writers may become reviewers doing QA on what the models put out because they can hallucinate, they can make things up, they can just go off the rails, and you absolutely positively need to have human beings fact-checking anything as high value.

Things that are not as risky would be things like summarization, and even there they can screw up, but they screw up less. Things like drafting commodity emails, like, hey, rescheduling this meeting for next week. Is this okay? Right? That's that's a a lower risk transaction than here's your medical diagnosis, right?

In SEO, there's this term that Google uses called your money or your life. And essentially, Google treats in SEO, Google treats any page content that is around finance and health with added scrutiny. That is a really good rule of thumb. That is a really good benchmark for AI, your money or your life. Are you telling people things?

There's a model telling people things that could have financial or or health care impacts. Not that you shouldn't use AI, but you should never let it be unsupervised. You or another human being who has subject matter expertise should be supervising what that model does at all times. And it should never be able to go directly to the public. Other ways to not use AI.

A big one is data privacy. Here's the golden rule. And this is something I say in our generative AI for marketers course, which you can get at TrustInsights.ai slash AI courses. If you are not paying, you are giving away your data. If you're not paying with money, you're paying with data.

And in healthcare particular, that's bad. Because if you're putting protected healthcare information that is other people's health information into a third party, you are violating so many laws that's not even funny. So that would be an example of how not to use AI. You would want to use a system where that was governed by your overall healthcare information technology policies. You would want to use systems, maybe that maybe there's some data you don't even want in the hands of third party contract or no contract, right?

Because there's always the the probability that you work with a third party and that third party gets compromised somehow. And then you got to send out that whole you know paper mail saying, Oh, hey, by the way, your your information was leaked or hacked or whatever. You may in those situations want to run AI locally on servers under your control, behind your firewalls, supervised by your IT team to protect that information. And you know, it that would then be as as secure as the rest of your IT infrastructure. But that be that's another area that again, people don't think of.

If you're not paying money, you're paying with data. And in healthcare, that's not allowed in in pretty much every place on the planet. Uh even in the US, where you know business regulations are notoriously lax for everything else. So those are the the how not to use AI things in healthcare in particular. The other thing I would say, it's not that you again, you don't want to not use AI.

You want to disclose. You want to disclose the use of AI everywhere. Everywhere that you use AI, disclose that, hey, we used AI for this. The terminology I Microsoft did this at their Microsoft Ignite, and I really like this language. Uh, for content they made with AI, uh, and then you know a human being uh supervised and edited.

It always said, this content made in partnership with AI using the whatever model. I really like that language because it is a partnership in many ways, and it's not that you're just letting the machines do things and you know you're you're like Homer Simpson just asleep at the wheel. Um no, you are an active partner too. So machines are doing stuff, you're doing stuff, and the final product should be the best of both worlds. It should be the speed of AI with the quality that the quality of human review.

That's a good way to approach AI and a good way to approach disclosure, the transparency, and say this is this is made in partnership with AI. So, um, hopefully that helps. Thanks for tuning in. I'll talk to you next time. If you enjoyed this video, please hit the like button.

Subscribe to my channel if you haven't already, and if you want to know when new videos are available, hit the bell button to be notified as soon as new content is live.