Med-PaLM 2 is the latest creation from Google, targeted at helping doctors analyzing cases.
Is it possible to replace doctors with artificial intelligence (AI)? Many people ask themselves this question as Large Language Models (LLMs) have started being used in professional fields.
In their latest AI development attempt, Google introduced Med-PaLM 2, an AI software trained on medical experts’ demonstrations and scientific literature. Its purpose is to help doctors analyze cases and provide useful advice.
Med-PaLM 2 is proven to work better than ChatGPT, Bard and the Bing chatbot concerning medical matters.
These are the three main AI softwares that have emerged in recent months. They all use LLMs to work, meaning they are trained on all the human-generated text present on the internet.
LLMs "predict" which words they are going to use next using a sophisticated algorithm. They learn how human language works simply through experience, by reading every possible string of text generated by humans.
LLMs, however, do not exactly "understand" text per se. They simply predict word for word what is going to come next. This way, they can also change the style and the details of the answer when prompted by users.
ChatGPT was the first text-based AI software. Developed by OpenAI, it was later acquired by Microsoft for a $10 billion deal.
Google immediately tried to jump on the AI bandwagon with their chatbot Bard. After a few months, however,it was clear Google would not catch up with Microsoft, therefore switching to business-oriented AI.
Does Med-PaLM 2 work?
The short answer to this question is... yes, more or less. Google research shows that Med-PaLM 2 understands the questions and can provide scientifically sound answers. It can back it up with academic consensus and usually shows signs of correct comprehension.
There is, however, a large problem with Med-PaLM 2, which is the same issue for every other AI software: accuracy.
Google’s software, just like ChatGPT, Bing chatbot and Bard, does not provide accurate answers. Often, it confidently replies with completely wrong assumptions, though with such confidence as to sound right.
Obviously, a similar tool cannot be used on a large scale right now. However, research continues as Med-PaLM 2 starts testing at the Mayo Clinic research center.
If it gets better, Med-PaLM 2 could be of serious help to doctors and other medical professionals. It is extremely unlikely, however, that it will replace one of them for good.
For the foreseeable future, AI softwares need human intervention and double-checking. They cannot, and probably will never, replace human reasoning and ingenuity.
