I continue to be fascinated by the topic of AI confabulation. The other day I read the Ernest Hemingway short story “Fifty Grand.” It’s about a boxer who fights a championship bout. I liked the story but was confused by a couple of details in the end. So, I turned to my favorite AI, Bing, which proceeded to hallucinate a whole other version for me.
Of course, I’ve seen AIs make up other scenes from famous literary works before. Bard even confabulated a “woke” version of the poet E.E. Cummings. So, Bing’s summarization of the Hemingway story was not a shock. But it’s worth writing about because of the nature of Bing in particular and the other similar AIs more generally.
Confabulating Parts of Fifty Grand
“Fifty Grand” is a story that hinges on a couple of boxing-related bets: one known, one unknown. Because the unknown bet isn’t made clear, the reader isn’t sure of what happened until the end of the story. To help clarify the ending, I asked Bing about it.
Bing’s summary of the story left out a critical part when a couple of “wise guys” (by the name of Morgan and Steinfelt) come to visit the boxer, Jack, and make some sort of proposal to which the reader isn’t privy.
Bing’s summary includes a number of events that never happened, such a meeting between Jack and his wife, her accusation that he was going to throw the fight, Jack’s broken jaw, and Jack’s trip to the hospital.
Bing didn’t accurately relate events in the story even though it had access to summaries such as the Wikipedia entry about it (to which Bing linked) as well as to the story itself (though in an unsearchable PDF format, which may have been an issue).
Wikipedia Gets It Right
So, did Bing get it wrong because Wikipedia had it wrong? No. Wikipedia summarizes the fight as follows:
The fight begins. Doyle (the narrator) reports, “There wasn’t anybody ever boxed better than Jack,” and the fight goes well for Jack for several rounds as his left fist repeatedly connects with Walcott’s face. By the seventh round, Jack’s left arm gets heavy and Walcott begins to gain the upper hand as he pummels Jack’s torso. After the eleventh round John Collins tells Jack the fight will go to Walcott, but Jack says he thinks he can last through the twelfth round and goes to meet Walcott and “finish it off right to please himself.” Walcott backs Jack up against the ropes and delivers a very low blow; Jack, in obvious pain, stays standing–”if he went down, there go fifty thousand bucks” because he would win on a foul–and tells the referee not to call the foul because it must have been an accident. He walks toward the dumbfounded Walcott and begins swinging, landing several body punches before hitting his opponent twice below the belt. Walcott falls and wins the match on a foul.Back in the dressing room, Jack comments, “It’s funny how fast you can think when it means that much money.”
In a couple of sentences, Wikipedia clarifies why Jack would have lost the money he bet on Walcott (the other fighter): because Walcott’s blow would have been deemed a foul and Jack would have won the fight on a technicality.
This suggests that the shady characters who talked to Jack earlier in the story wanted Jack to win the fight. At the end of the story, the narrator tells us that Walcott was favored to beat Jack, which is why Morgan and Steinfelt wanted to bet and win on Jack (that is, they’d win more money that way).
But it appears that Jack’s agreement with them is that he would lose the fight. That’s why toward the end of the story, Jack’s manager says, “They certainly tried a nice double-cross” and Jack responds with sarcasm, “Your friends Morgan and Steinfelt. You got nice friends.”
So, Morgan and Steinfelt wanted Jack (and most other people) to bet against Jack’s victory so they would make more money when Jack won. In essence, Jack turned the tables on them by making sure he lost the fight even while getting revenge on Walcott for his dirty boxing and treachery.
What Can We Learn About Today’s Neural Networks?
I certainly don’t “blame” Bing for getting a nuanced story wrong. I know that the confabulations boil down to how the algorithms work, as explained in another post. In fact, unlike the other AIs on the market, Bing pointed me to references that, if I hadn’t already read the story, would have allowed me to verify it was giving me the wrong information. That’s the beauty of Bing.
Not Quite Plagiarism
The famous intellectual Noam Chomsky has claimed that the generative AIs are just a form of “high-tech plagiarism.” But that’s not quite right. I don’t know if the story “Fifty Grand” was part of the data on which the Bing model (based on ChatGPT4) was trained. If so, then it wasn’t able to properly parse, compress and “plagiarize” that nuanced information in such as way that it could be accurately related after model training.
But we do know that Bing was able to access (or at least point to) the Wikipedia article as well as an “enotes” summary of the story, so it knew where to find the right plot summary and interpretation. The fact that it still confabulated things indicates that the makers and users of these technologies have some serious issues to address before we can trust whatever the AIs are telling us.
Will Hallucinations Ever Go Away?
There’s some debate about whether the confabulations and hallucinations will ever go away. On one hand are people such as Emily Bender, a linguistics professor and director of the University of Washington’s Computational Linguistics Laboratory, who has said, “This isn’t fixable. It’s inherent in the mismatch between the technology and the proposed use cases.”
On the other hand are those who think the problems are indeed fixable. Microsoft co-founder Bill Gates said, “I’m optimistic that, over time, AI models can be taught to distinguish fact from fiction.”
Maybe APIs Will Help Fix the Issue
Some think they can address the confabulation problem, at least in part, by better use of APIs (that is, application programming interfaces). By interfacing with other types of programs via APIs, the large language models (LLMs) can develop capabilities that they themselves do not have. It’s like when a human being uses some tool, such as calculator, to solve problems that they would not easily be able to solve by themselves.
That is, in fact, part of the hope associated with Gorilla, a LLaMA-7B model designed specifically for API calls. This particular LLM is a joint project developed researchers from UC Berkeley and Microsoft, and there is now an open-source version available.
So, if Gorilla can more dependably access APIs, it can reduce the hallucination problem.
At least, that’s the hope.
We’ll see over time.