> >So on this single test, AI seems promising.

I have run 10 or 15 scenarios with ChatGPT; some have had similarly cogent results to yours. However, for others, I had to re-ask the question several times before the AI admitted that its initial response was wrong. One example was related to the flashpoint of dimethyl formamide, which ChatGPT identified as flammable. Because I knew this response to be incorrect, I re-asked the question several times in different ways, and the AI became progressively less sure of itself in tone. It finally admitted it was mistaken. I don’t know if I would have persisted through that process if I didn’t have reason to know the answer before I started, though.

In the larger picture, ChatGPT does not tell you where it gets its information, so I don’t know how to assess the validity of its responses. There are other chatbots that do provide references for their statements, but their statements are less detailed than those from ChatGPT since their responses are more tightly tied to a specific source.

As was pointed out a day earlier, the ACS Publications web site has a more complete description of TATP concerns in its Analytical Chemistry journal, which is peer-reviewed and has references to help assess its broader statements. The CAS Chemical Safety Library does rely on manual entries. We are working to identify ways to mine the chemistry literature to improve the CSL, but I’m not convinced that a chatbot tool will be part of that process.

In sum, my experience is that AI is as likely to confuse a safety question as it is to answer it. We will continue to have to educate our selves and the Chemistry community in chemical safety literacy, including how to assess chat generated answers that will soon be routinely appearing in the literature we work with.

Good luck to us all! ;)

- Ralph

Ralph Stuart, CIH, CCHO
ralph**At_Symbol_Here**rstuartcih.org

---
For more information about the DCHAS-L e-mail list, contact the Divisional membership chair at membership**At_Symbol_Here**dchas.org