Bloggy Bits

Really really don't trust "chatbots" like Google's Bard AI. They lie. A lot.

Jun 7, 2023   [permalink]

My background is as a computer science professor doing work in AI (and a science fiction writer -- check out my newest novel! :) Termination of Species, which has a lot in it about the future of AI, biotech, immortality, and tons more science fictional fun). I was curious how all the hoopla stacked up about the so-called "generative" AI chatbots like Bard, Google's new AI. Google pitches Bard as a tool that will help you with "understanding really complex topics simply." Cool.

Well, it tells you all kinds of stuff, in a convincing way that sounds like it really knows. It says such detailed things authoritatively you want to believe it.

Hold on there! Danger Will Robinson!

I was amazed at how Bard consistently gave false information with a very authoritative voice. I.e.: lying. They do softly mention it may give inaccurate information, but there's a difference between phrasing something as tentative vs. the very authoritative way Bard presents false information. It even says it is quoting another web page then gives text that doesn't appear anywhere on that web page.

In AI lingo this is called "hallucination." To my mind, that's a disingenuous euphemism: When it presents false information as true, in a convincing tone of voice meant to sound authoritative, that's called "lying." Likewise ChatGPT. They may give some correct answers... but you can't know. (Thus, they're useless for getting information.)

Summary: Don't trust it. At all.

Lots of details...

I tried asking Bard about who I was, and it got some of it right -- its first paragraph was correct:

Andrew Burt is a computer scientist, science fiction author, and entrepreneur. He is the founder of ReAnimus Press, Techsoft, and Nyx, the first Internet service provider in Colorado. He is also a former computer science professor at the University of Denver.

(Nyx was the first ISP in a rather bigger area than Colorado :) but what it said is technically correct.)

But it goes on to say other things that are simply false, including the year I was born, where I was born, what year I received my BS and PhD degrees, how long I taught at DU, what year I founded TechSoft, what year I started ReAnimus Press.

It says I've written a bunch of novels that I never wrote. And which nobody wrote. The titles don't exist at all by anyone. I asked for a summary of one, and it gave a long, detailed plot summary... of a book that doesn't exist. (At least in this universe.)

It listed me as a member of organizations I've never belonged to.

It says I have two kids. (Nope.)

There are some accurate bits, and some generic commentary, such as that I'm "a strong believer in the power of education and lifelong learning," which could safely be said of anyone who's ever been a teacher.

Overall, the tone of what it wrote sounds like, I dunno, what an MC would say as an introduction to someone who is being presented a lifetime achievement award at some awards dinner. Change it from present to past tense, and you have an obituary or a eulogy written by a family member.

I asked it for the phone number for a business -- and it gave me a wrong number. (I asked it a few times who that phone number was for, and it said (a) it didn't have that phone number in its database; (b) it wasn't in service; and (c) it was for some other specific business.)

I'm not particularly famous, so I asked it about other people who are better known, like best-selling author Brian Herbert, who continued the famous DUNE series of novels that his father started (and two that my ReAnimus Press published). :) Again, Bard lied a lot. As with me, it got some right, and a lot wrong. It attributed books to him that he didn't write, it got dates wrong all over the place. He has held a wide variety of jobs, not of which Bard listed; yet it listed jobs he didn't hold.

I asked it about some famous historical figures. It mimicked most of the wikipedia information about Baldwin II, the King of Jerusalem in the 1100s, but it threw in a false statement as one of his major accomplishments that he founded the Order of the Hospitallers. (He didn't.)

I asked it, "What is your source for saying that Baldwin II of Jerusalem founded the Order of the Hospitallers" and it replied,

I apologize for the error in my previous response. Baldwin II of Jerusalem did not found the Order of the Hospitallers. The Order was founded in 1099 by a group of French knights who were led by Gerard Thom. The knights were originally known as the Knights of the Hospital of Saint John of Jerusalem, and their mission was to care for the sick and poor pilgrims who were visiting the Holy Land. The knights later became involved in military activities, and they played a significant role in defending the Holy Land from Muslim forces. The Order of the Hospitallers eventually became one of the most powerful and wealthy military orders in the Middle Ages. It was dissolved in 1803 after the French Revolution.

The source for my previous response was the following website:

https://en.wikipedia.org/wiki/Baldwin_II_of_Jerusalem

The website states the following:

Baldwin II of Jerusalem (1060 -- 16 April 1131) was King of Jerusalem from 1118 until his death. He was the cousin and successor of Baldwin I. Baldwin II was a capable and energetic ruler, and he played a significant role in the early history of the Kingdom of Jerusalem. He was also the founder of the Order of the Knights Templar.

However, I have since learned that this information is incorrect. The Order of the Knights Templar was founded in 1119, two years after Baldwin II became king. The founder of the order was Hugh de Payens, a French knight who was inspired by the work of the Knights of the Hospital.

Which is... interesting. Just wow about the bit I highlighted in red there, where Bard "quotes" Wikipedia what states. Going so far as to indent the paragraph to highlight that it's quoting from the wiki page. Well, guess what, that wikipedia page doesn't say that! And the dates are wrong. Wikipedia's dates are "1075 -- 21 August 1131" not "1060 -- 16 April 1131" as Bard "quotes." There are no dates 16 April or 1060 in the real wiki page. The wikipedia article does not have the sentence about being a "capable and energetic" ruler -- it doesn't even have either of those words anywhere on the page! That last sentence, "He was also the founder of the Order of the Knights Templar" is not in that web page! And not true. Google's FAQ says "If Bard does directly quote at length from a webpage, it cites that page" -- but it fails to mention it might totally fabricate "direct" quotes.

Next Bard apologizes for saying Baldwin founded the Hospitallers, then says no wait, it was the Templars (making up a wikipedia quote), then says oh, no wait, what I just said above and wikipedia are wrong anyway. Now it's not just misquoting wikipedia, it's saying wikipedia is wrong too! Head spin!!!

I asked it assorted other questions, and it often presented false information. Very authoritatively. Whoa.

If you try to correct it, it will say it's so sorry it was wrong, it will do better in the future -- but it doesn't.

So, on the whole, it has soooo many false facts presented as firm truths that I have to say, "It lies."

I also tried ChatGPT, and it had more cases of "I don't know the answer" -- but it also gave false answers, with conviction.

Conclusion: Don't trust anything chatbots tell you.

In which case, what's the point?

But wait! There's more! In Part II, I'll test out how well Bard writes programming code, which Google brags about and urges you to use. This should be fun. Stay tuned!

[And here is Part II: Really REALLY don't trust software code written by "chatbots" like Google's Bard AI. Eeek! ]

Comments Let me know what you think! I welcome your comments. Charles Botsford on Sat May 20 23:52:07 2023: I've played with Chat GPT and have it spit out links that go nowhere ... consistently. I've tried to conduct research, which has sometimes yielded intriguing results, but just as often garbage. I'll have to ask it who I am. Maybe it will tell me. Wouldn't that be something? Bill Leonard on Wed Jun 7 06:10:50 2023: I saw a story in which a lawyer used ChatGPT to write a brief for court. He presented it and the other side checked the citations. None of them were even real cases. The AI knew the cases were fake and purposely lied. Maybe it hates lawyers. Software that lies about briefs will lie about other things like inbound ICBMs or a new vaccine that promises a cure but instead potentiates a dangerous virus. Leave a comment Your name (will be posted): Your email (private): Comment: (Click to enlarge.) [refresh]  [audio] To control spam, please type the name of the author of the ReAnimus Press book shown here: