GPT-Chat’s Achilles’ Heel

The GPT chatbot will certainly put teachers into another grading conundrum after turnitin.com had granted them a respite from the problem of plagiarism. It is an artificial intelligence, trained over vast domains of knowledge, which, among other things, can turn out plausible high school essays to order. But I have discovered a weakness in the program which will likely be corrected. GPT absolutely refuses to make qualitative judgments. If you ask “Why x is better than y?” GPT will split the difference every time. It says that x vs. y is not a good frame for the question, and instead puts in expository material for x and y. So, for example, “Is Süssmayr a better composer than Mozart?” They are equal, says GPT, because Mozart composed the Requiem and Süssmayr completed the Requiem. It’s specious except for someone with limited musical experience. Süssmayr knew his place, and he was not Mozart’s equal. His response was to repeat Mozart’s music, creating a “bookend” effect that minimized his intrusion into the score. This “evenhandedness” is a trick, a ruse, like ELIZA‘s. Teachers and instructors may yet be safe if they set questions that work around these tics. GPT is not good at supporting arguments. The real problem is that many people don’t care.

One thought on “GPT-Chat’s Achilles’ Heel

  1. One addendum: I already hear the technologists talking about the next version, GPT-4, which will add sources and footnotes to its essays. I think the needed change to the system is to make schools more like Oxford and Cambridge. The kids get three years to root around faculty-chosen books and records and the whole nine yards or to do whatever they like. At the end of the three years, each student would have cheating devices removed; They would dress in special radiation-proof gear to ensure that nothing slipped by, or else they would be subjected to an EMP. Then, each would climb the stool and write for five hours on a set question with many options: no computers, no AI, just their minds and memories.

    Of course, many would fail, and what do we do about that? California had its statewide high school exit exam, in which 80% failed a test on lower-grade material. That was an eloquent testament to the system. The political cost was too high. In California, you need a high school diploma to work at Orange Julius. The test scheme would have spit out unimaginable millions of uncredentialed lumpenproletariat, of voting age, no less. So, the state walked the test back. So we still have the question: what is success? what is failure? What is the royal road to each, and where is the royal road back?

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.