Hi all,
I'm very hesitant about tools like ChatGPT at this stage because from what I've seen, it gives correct and incorrect answers with equal confidence. This is just anecdotal, but I've seen a number of stories where professors have tested ChatGPT by asking it about their work and it just would invent nonexistent papers, books, quotes, etc. The thing to remember about these AI models is that they aren't built to give correct answers, just answers that *sound* correct. Most of the time, the answer that sounds most correct will be the actual correct answer, but if it's struggling to come up with real information it might just output something that uses all the right words and "tone" but isn't actually true rather than throwing an error message.
My personal philosophy on it at this point is to treat AI like Wikipedia: an interesting place to start, but verify everything independently.
Sincerely,
Jack Reidy (he/him)
Research Safety Specialist, Assistant Chemical Hygiene Officer
Environmental Health & Safety
Stanford University
484 Oak Road, Stanford, CA, 94305
Tel: (650) 497-7614
I acknowledge that the land on which I live and work is the ancestral and unceded land of the Muwekma Ohlone Tribe. As an uninvited guest on these lands, I am a beneficiary of the ongoing displacement of the Ohlone people. I pay my respects to the Native peoples, past and present.
-----Original Message-----
From: ACS Division of Chemical Health and Safety <DCHAS-L**At_Symbol_Here**Princeton.EDU> On Behalf Of Craig Merlic
Sent: Monday, May 8, 2023 4:16 PM
To: DCHAS-L**At_Symbol_Here**Princeton.EDU
Subject: Re: [DCHAS-L] What we know − and don't know − about what caused UD's lab evacuation 2 weeks ago
Ralph,
AI vs ACS:
Since you said " I don’t think that AI’s strengths of sifting through word combinations is likely to be helpful in predicting outcomes ", I did a comparison of the Pistoia Alliance Chemical Safety Library vs ChatGPT.
Searching the ACS Chemical Safety Library (https://safescience.cas.org/ ) for acetone or hydrogen peroxide returns the statement: “Formation of acetone peroxides possible. Try to avoid combination or check for peroxides.” There is no entry for TATP.
Asking ChatGPT " Safety of mixture of acetone and hydrogen peroxide" gives:
" Mixing acetone and hydrogen peroxide can potentially be dangerous as it can result in a highly reactive mixture that is prone to spontaneous combustion. This mixture is called acetone peroxide or triacetone triperoxide (TATP) and is a powerful explosive.
Acetone peroxide can be formed when acetone and hydrogen peroxide are mixed in the presence of an acid catalyst or when they are allowed to react at high temperatures or for extended periods. This reaction can occur even with relatively dilute solutions of hydrogen peroxide and acetone.
Therefore, it is important to avoid mixing acetone and hydrogen peroxide unless you are following a specific procedure that requires their combination, and you have the necessary training and safety equipment to handle the resulting mixture safely. If you must mix them, it is crucial to follow proper safety protocols, including working in a well-ventilated area, using protective equipment such as gloves and goggles, and avoiding any sources of heat or sparks.
It is also important to note that acetone and hydrogen peroxide can each pose individual health hazards when not handled properly, such as skin irritation and respiratory problems, so it is important to handle them with care. "
So on this single test, AI seems promising.
Best,
Craig
Craig A. Merlic
Professor of Chemistry, UCLA Department of Chemistry and Biochemistry Executive Director, UC Center for Laboratory Safety http://cls.ucla.edu
On 5/8/23, 1:01 PM, "ACS Division of Chemical Health and Safety on behalf of Ralph Stuart" <DCHAS-L**At_Symbol_Here**Princeton.EDU on behalf of ralph**At_Symbol_Here**RSTUARTCIH.ORG> wrote:
> >How about leveraging AI/ChatGPT, Alexa, Siri, etc. to help predict the outcomes when starting with these reagents and using the described process?
>
> >Has anyone tried using one of these to predict potentially dangerous outcomes?
I know of one effort, although I don’t believe it does involves AI (depending on your definition of AI). The Purdue Reactive Hazard Evaluation Analysis and Compilation Tool can be found at
https://rheact.github.io/#/welcome
Based on the data they ask for and logic they describe there, I don’t think that AI’s strengths of sifting through word combinations is likely to be helpful in predicting outcomes.
More specfically, in addition to SDS’s, the Purdue tools walks the user through these steps:
" • Parsing multiple Safety Data Sheets (SDSs) for chemical and safety information
• Creating an operational chemical hazard matrix
• Generating a pairwise chemical compatibility matrix
• Calculating final reaction temperature after an adiabatic temperature change
• Providing PPE recommendations from SDS and PPE questionnaire
We aim for RHEACT to be a quick preliminary screening tool that alerts users about hazards and pushes them to perform further analysis.”
Given that each of these steps need to be sanity-checked to be used for safety decision-making, the data and logic requirements here don’t appear to be quick or preliminary to me.
- 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
---
For more information about the DCHAS-L e-mail list, contact the Divisional membership chair at membership**At_Symbol_Here**dchas.org
---
For more information about the DCHAS-L e-mail list, contact the Divisional membership chair at membership**At_Symbol_Here**dchas.org