I wanted to see what all the fuss was about so I registered an account at OpenAI and have been initiating conversations with ChatGPT about a wide range of subjects. The past two evenings I have been asking the chatbot questions about boxing, something about which I know a lot. A lot more than ChatGPT, apparently. In this blog, I share screenshots of where our conversations go off the rails. It looks like ChatGPT bluffs and then when you call it out on it, it will apologize and attempt provide the correct information. As you will see, it bluffed me several times. Check it out.
The exchange is revealing because of the faulty logic the program used to rule out a fight with Armstrong. In fact, Garcia had at least two fights with Armstrong, Armstrong defeating Garcia over 15 rounds in a welterweight title match, then the draw for the middleweight title. I could have challenged ChatGPT on another error, as the middleweight fight with Armstrong was for 10 rounds, not 15.
I laughed out loud when it dropped Luis Firpo’s name. But there is so much more that’s wrong with this. Paul Berlenbach and Tommy Loughran were light-heavyweight champions. Berlenbach, who beat Mike McTigue for the title, did not lose his title to Loughran but Jack Delaney. To my knowledge Berlenbach never met Harry Grebs in the ring. For sure he did not ever hold the middleweight title. Loughran won the light-heavyweight title from Jimmy Slattery.
That was Wednesday night. On Thursday evening, ChatGPT’s performance was even more dismal. The first exchange concerned the lightweight great Benny Leonard:
I then moved on to a series of queries about the mighty Henry Armstrong.
I knew it was 18, but I wanted to see if ChatGPT would go with 14. Here’s some more Armstrong questions. The bot gets so much wrong in this exchange. You cannot trust the accuracy of the language model.
Robinson and Armstrong met in 1943, not 1945. ChatGPT corrected itself to say they never fought. So I corrected the bot (again):
These wild inaccuracies do not recommend this language model for use for anything. Those who talked it up did not spend much time with the language model or properly interrogate it. God help us when they use this in telemedicine.
Finally, I asked about the Marvin Hager. Watch what happens when I ask about his fight with Sugar Ray Leonard.
There is a lot of concern that high school and college students will use ChatGPT to write their essays for them. Somebody should tell them that ChatGPT can be highly inaccurate. Moreover, the writing is mediocre and formulaic. Ask for an essay on any thinker and his significant work and you get an introduction, three repetitive paragraphs in support of the thesis, and a conclusion—all paragraphs of approximately the same length. The conclusion invariably begins, “In summary, …”
Granted, this is not the best AI has to offer us, but maybe we are a little more than seven years away from singularity.