using AI? check your refs!
AI, aka artificial intelligence, has been hotter than Hades since Open AI released ChatGPT in November of 2022. And it is impressive. I was certainly blown away at how good a writer it is, at least for the responses it has given me for various technical questions I've asked. But beware: AIs have a definite writing style, something I was able to detect when two students submitted very similar essays to a prompt I gave them in November of 2022. However, researchers (and everyone, really) need to be extremely careful when using AI because, and I don't know how to put this politely, AI is a bullshitter.
I first experienced the bullshitting when I asked it to provide a response along with supporting references on a question I had about evaporation. ChatGPT provided a well-written responsive response with several references. Being anal retentive (that is, a good researcher), I wanted to see the words the reference used with my own eyes. That turned into a two-hour black hole of trying to find a reference--a book--that didn't exist.
The three authors were real, the publisher was real, the topic was in each author's expertise, but the book simply didn't exist.
AIs, including ChatGPT, keep getting better (or so I hear), so I recently gave it another chance. This time, I was trying to find an academic source for a few statements made in non-academic books that Napoleon fired cannons into the sky in an attempt to make it rain (the book I am writing is following the water cycle). I also know through careful prompting (Prompt Engineering) to not only ask for the reference but to also include a link. No more two-hour searches for fake stuff.
So I gave the GPT the order above (yeah, yeah: I should be more respectful to Our Future Skynet). Here's what I got in response:
Looks pretty damn good and pretty authoritative. That Wilkinson bit looks interesting...
And here's what it gave me. Sweet! And with a link!
Let's click, shall we?
Hmmmm. The author is real, the journal is real, the link is real, but the link does not produce the article as titled in the reference.
So I wrote ChatGPT the following, dripping with disappointment:
Here's its response:Touchy much? This is akin to saying "Well, you caught me in a lie, but in reality it's not real, and here's some more crap you won't know is accurate or not.
Also: WTF?
The lesson here is to Check. The. References.
I had a student submit a contribution to a paper the class was working on. The contribution was well written (that should have been a clue...) with three references. Two of the three didn't exist. Did the student use AI? Yes he did. Check your references!
Here's another recent interaction:
Comments
Post a Comment