A Chatbot Treats a Forum Post Like Top Science

Avishai Abrahami, the founder and chief executive of Wix, was working through where artificial intelligence helps beyond writing software code. He argued that law is a natural fit because, like code, it is mostly words, precedents, and written records. Then he offered the exception that unsettles him.

He was talking about large language models, the systems behind chatbots like ChatGPT. “Every time I see a doctor trying to do research by using LLMs, okay, I get horrified because that's the wrong tool.”

What horrifies him, he said, is how flat the machine makes the world. Something it picked up on Reddit, the sprawling online forum where anyone can post anything, is presented as equivalent to the best research published in Nature, one of the world's most selective science journals. There is no difference in weight, he argued, and most of the time the model has not actually read the research it claims to summarize, even in its most advanced mode.

His interviewer pushed back with his own view that ChatGPT is already better than most doctors, that health had become among the most popular uses of chat, and that K Health, a health startup he backs, had shown AI can diagnose better than most doctors.

Abrahami drew a sharp line between that kind of diagnosis and research. He conceded the point on everyday diagnosis, the familiar visit that ends with a prescription, rest, and recovery in a few days. That kind of pattern matching, he said, is totally different from doing research.

For work built on text, like sorting legal documents and past cases, he thinks the technology is really good. For work that creates new medical or scientific knowledge, he thinks it is the wrong algorithm.