I have written in the past about the “hallucinations” that the various AI chatbots come up with—confidently-reported “facts” that turn out to be pure invention. I reported in detail on such an experience with ChatGPT last September.

Although I have had a good experience using Claude for basic data analysis, I have continued to experience wild hallucinations with chatbots, including Claude.

Those experiences have soured me on chatbots generally—especially where facts are important. I don’t think the technology used by current chatbots, called LLMs (“large language models”), will ever be trustworthy. Even so, I do find the chatbots useful in certain limited situations. Two examples: outlining and brainstorming.

Outlining. I recently did a short presentation for SSAFE (Senior Stewards Acting For the Environment, http://www.ssafe.org) on the subject of “autonomous shuttle systems for retirement communities”. I asked Claude (which seems to be the best of the current chatbots) to help me develop the PowerPoint file. In its response, Claude included non-existent retirement communities and autonomous-vehicle providers that were in unrelated industries or out of business. I had to check each of Claude’s responses by searching for the website of the retirement community or vehicle maker to confirm Claude’s “facts”, and many were wrong. Not good. But Claude did provide a useful framework for organizing the presentation. (And its design skills in PowerPoint are far are better than mine.)

So in that case, Claude was helpful even though its facts were wrong. (In case you’re wondering, I found no verifiable instance of an operating autonomous shuttle system at any retirement community, although there have been several pilots.)

Brainstorming. In the process of my “autonomous shuttle” project, Claude came up with relevant issues around autonomous shuttles that I hadn’t considered.  For example, Claude provided a table listing the factors that distinguish the features needed in a shuttle system for a retirement community from the features a university system needs. Retirees may have higher onboarding needs, may need longer to gain trust in the system, may not be able to use a smartphone for summoning, and so on. That table was a helpful addition that I hadn’t thought of.

In several other projects, I have found that when I ask a chatbot for ideas about a specific topic, it comes up with some good ones that hadn’t occurred to me.

So I find I am happy to use chatbots in situations where they are making suggestions. I can take or leave the suggestions, but often they are helpful. However, I will not believe any “facts” a chatbot tells me until I have independent verification.