What does ChatGPT know about California workers’ comp? I decided I’d ask it, and during 2 sessions I posed about 45 questions on a wide range of California workers’ comp topics. I’ve included both discussions with ChatGPT in pdf format at the end of this post.
In case you’re just returning from an isolated island, Chat GPT is the artificial intelligence chatbot recently unveiled by San Francisco’s OpenAI, with major investment by Microsoft.
My questions to ChatGPT cover a range of California workers’ comp topics: workers’ comp terminology, workers’ comp law, pending legislation, reliable sources for information on workers’ comp, historical information about reforms, events under different governors, current comp system problems and trends, etc.
System stakeholders would do well to pour over the ChatGPT sit-down. Chat GPT will continue to evolve, but warts and all, the future is here. You can draw your own conclusions about how this will affect the workers’ comp community going forward.
My conclusions? ChatGPT gets much right, but makes many critical errors. Its algorithm clearly scrapes information but has trouble making fine distinctions. Some of the errors are minor, but others egregious. Sometimes it injects incorrect facts or concepts into answers. An injured worker, employer or policymaker who relied on these answers might come to erroneous conclusions about the comp system.
For example, I had to correct ChatGPT which asserted that California injured workers who can not return to work at their employer have a right to $16,000 in vocational retraining benefits. Reforms replaced that with a much smaller retraining voucher years ago.
It listed a bunch of incorrect names when asked about DWC, DIR and CAAA leadership. It thought the SIBTF had something to do with behavioral health.
It did direct me to what it considers reliable sources such as the Department of Industrial Relations DWC website, the CWCI website, Lexis Nexis etc. It directed me to some noteworthy blogs, including my own blog. Sometimes it seemed confident about answers but other times seemed to hedge and provide cautionary warnings.
ChatGPT did better explaining system basics than in analyzing some of the thorny issues in California workers’ comp. But answers on apportionment, the QME process, the UR process, and the politics of workers’ comp reform often included a mixture of correct and incorrect information.
I didn’t take the time to explicate all the errors to the Chatbot, lest the discussion get bogged down.
But the chatbot was apologetic when I did point out errors. It promised me it would try to do better.
Maybe it will. But if it doesn’t, then what?
I suggest you read my first and second ChatGPT discussions for yourself (note that the ChatGPT answers to my questions are in the black text boxes)
This is the first chat:
The is the second chat:
To learn more about ChatGPT as of the date of the blog, see this Wikipedia entry: