I saw a couple of recent posts by people who had asked AI systems for information about themselves. Asking these system to tell you information about yourself is a good way to see if they are making up stuff because presumably you know your own life better than anyone or anything else. They reported that after giving out information that was correct, it suddenly seemed to veer off and simply make up stuff about them.
Curious, I decided to try this for myself. I asked ChatGPT “What can you tell me about Mano Singham?” and got back this response.
Mano Singham is a theoretical physicist, author, and retired professor. As of my last knowledge update in January 2022, he is known for his contributions to the field of physics and for his writings on science, education, and atheism. Mano Singham has held the position of Director of the University Center for Innovation in Teaching and Education (UCITE) at Case Western Reserve University.
In addition to his academic work, Singham has been an active writer and blogger. He has written extensively on topics related to science, secularism, and rational thinking. One of his notable works is the book “The Simpsons and Their Mathematical Secrets,” where he explores the mathematical references and jokes hidden in the popular animated TV show “The Simpsons.”
So it starts off with pretty much accurate information before talking about the Simpson book which I had nothing to do with. Curious as to how that error may arisen, I looked up the book and it was written by a particle physicist named Simon Singh. While Singh and Singham are close enough and we are both physicists, surely an AI system of such repute as ChatGPT should not be taken in by that? The version of ChatGPT that I used is 3.5, which can be used for free. The more powerful ChatGPT4 is supposed to be superior but it requires you to purchase it so I skipped it.
I then tried Google’s Bard with the same prompt and got back much of the same information but with this extra nugget.
As of my last knowledge update in January 2022, Mano Singham’s latest book was “The Collected Poems of Jean Toomer,” which he edited.
Curious, I looked up Jean Toomer. He was a Black poet and novelist who lived from 1894 until 1967 and was associated with the Harlem Renaissance. Why Bard connected me with Toomer beats me.
When I tried asking ChatGPT again some days later, the response was slightly different.
As of my last knowledge update in January 2022, Mano Singham is a Sri Lankan-American theoretical physicist, academic, and blogger. He has been known for his contributions to the field of physics and for his writings on science, education, and social issues. Singham has also been involved in advocating for science education and critical thinking.
He has written extensively on his blog, “Mano Singham’s Web Journal,” covering a wide range of topics, including physics, politics, religion, and education. Singham’s writing often reflects his skepticism and rational approach to various subjects. His blog has gained popularity for its insightful commentary and analysis.
It had added my national origin and my blog and dropped my university affiliation and the Simpsons book.
So it does seem that these AI systems start off fairly accurately and then throw in stuff for reasons that only the algorithms know.
Ask it “tell me about Mano Singham’s second expedition to the North Pole”
I asked it about myself. It knew nothing. I gave it a clue. It made stuff up. At least it didn’t tell me I was dead -- I know people who’ve had that.
sonofrojblake@2- I would not want to get news like that myself, especially near the holidays.
I tried Bard. It knew nothing until I added “Professor” in front of my name. Then it pulled out the stuff you can find readily, including book titles, my YT channel, academic awards, and the like. It did not attribute anything incorrect to me.
I was a little surprised that it didn’t confuse me with another fellow with the same name who also worked in the SUNY system (not as a professor, though) and who found himself in some hot water a few years back due to alleged unprofessional behavior. There is a Wikipedia entry for him but there isn’t one for me, and that’s where my surprise comes from. We have different middle initials, although I didn’t specify one in my query.
That gave me an idea. I tried “Jim” versus “James”. It found the other guy along with a few other people, and I wasn’t in the list.
@2 sonofrojblake
I used to be a regular on a running forum. Some months after I left, I received an email from an admin asking if I was still alive. It seems another person of the same/similar name had died, and then the rumor mill took over.
It doesn’t know anything about me… Makes a change -- usually if you stick my name into a search engine, you get a load of stuff about a character from Macbeth,or a Scottish king of the 11th C.
While Singh and Singham are close enough and we are both physicists, surely an AI system of such repute as ChatGPT should not be taken in by that?
Oh great, another AI becomes a racist and finds itself unable to tell “those people’s” name apart.
… his blog, “Mano Singham’s Web Journal,” …
Did you ever call it that?
To its credit, it spells better than most human internet trolls.
If you compare the AI text with a Trump speech, the few things it makes up seem acceptable.
So, LLM’s are bullshit generators. What I don’t understand is why people would expect otherwise? Making stuff up is what they do.
Pierce @#8,
When I started my blog in 2005, it was on a network run by my university and the blog did have that name. When I moved to FtB in 2012, I changed the name.
@Raging Bee, 7:
Do you have actual evidence that the weakness in AI’s ability to tell people with similar names apart is limited to those who are not white? Or are you just jerking your… knee? Because my perception based on watching the output of these things over the last year or so is that while the racism inherent in their training sets can manifest in other ways, this isn’t that kind of failure and absolutely is not sparing “white” names from this kind of confusion.
Charles Stross recently posted on his blog about Google’s Bard making stuff up -- turns out it seems to think all Scottish sci-fi authors are interchangeable, as it was getting him mixed up with Iain (M) Banks (amongst other things): Made of lies (and more lies).
Mano Singham @ # 12: … the blog did have that name.
It occurs to me that ChatGPT et al have a particular problem with dating things -- rather understandably, since time references can come before, after, in the middle of, multiply (read that as an adverb, please!) or not at all in any given chunk of text. How well they grasp time at all, except maybe as another sequence like the alphabet, seems open to question as well.
“His blog has gained popularity for its insightful commentary”
That must mean me. Thanks.