I have been thinking about this for awhile. How physicians, and people in general, can approach decision making has been changing for some time. It started with the explosion of information, new studies, new drugs, and new technologies. Then came things like Google. The idea that with the huge influx of information, we could use these tools to help us find answers. Google achieved the rare honor of becoming a verb. And it really has been pretty amazing. I recently needed a copy of the form that accompanies an honorable discharge. I searched for it, found it, and got it sent to me with little effort. Amazing? I could find information about any topic a group might be arguing about at the dinner table. So much fun to be right and put people in their place! But there is the little problem. Our friend “the Google Machine” will give you an answer no matter what. Sometimes you can get yourself in trouble. Suddenly the verb “google” was also something to ridicule in the sense of “I found this ridiculous thing on the internet machine”.
Of course, things have continued to evolve and now we have AI. Lots and lots of AI. Once again, plenty to love, but also the same problem. Ask a difficult question, and your AI tool will happily give you an answer (with a nice dose of sycophantic congratulations for asking such an insightful question). Usually, with certainty. The idea of trying to separate out junk and harmful information is getting harder. How are we to know when information given to us by an AI tool isn’t correct? Especially if it is in an area that you may have little of your own expertise to evaluate the response you are getting. This is a question that I think should bother most everyone, but in a healthcare setting uncritical use of what new AI tools yield could be harmful. So, what’s to be done?
First. I think it is worth celebrating that these new tools are available for us to use on the behalf of ourselves and our patients. But AI tools, at least for now are still things that have sucked in huge amounts of data and then are able to organize it to answer questions. But we need to recognize that in sucking up all that information, some of that data is incorrect.
How do we go forward? While I think of this topic as something for students and learners, I believe that anyone who is making their way in this new environment can benefit from finding methods to safely incorporate AI into the way they engage with the questions that they ask with the information and aid they receive from AI tools.
A few thoughts to start – AI systems can serve as “cognitive accelerators,” helping learners brainstorm, draft, analyze, and potentially simulate ideas. However, as mentioned, Large Language Models (LLMs) lack true understanding and can generate confidently incorrect information that may embed biases derived from training data. Further, they can be very sycophantic. Students must therefore be trained not only how to use AI—but how to think about AI.
Effective AI literacy requires:
– Foundational domain knowledge – this means we have to understand what that foundational knowledge really is
– Understanding AI limitations
– Structured verification of the information supplied
– Ethical reasoning and transparency
In the future AI may be able to critically sift through data in a way that allows it to discount false, old, uncertain, or incorrect data. But until that time, we need to prepare future physicians with the skills to use AI as a partner, but also develop the intellect and critical decision making skills to know when that partner with so many facts, might be giving advice that needs to be critically reexamined and potentially rejected.
More to come.