niffyjiffy

What We Have in Common with Generative AI

The Test

Being a teacher, I’m often asked in casual conversation for my opinion on standardized testing. Once the conversation progresses past the usual banter about Bill Gates, Dan Quayle, and No Child Left Behind, I tend to present a couple of example problems. Consider the following two:

• Fred has 6 apples and eats 2 of them. How many apples does he have now? • Miss Flannery O’Shaughnessy owns an apple preserves shop. The apples arrive at a rate A(t) = 2500sin(2πt) + 3000 and are sold at a rate of 2000cos(2πt) + 3000, where t is in years since January 1, 2000. How much capacity does Flannery need in order to store the maximum surplus of apples she will accumulate? On the face of it the problems are quite different in complexity and in subject matter; however, their similarities are much more illuminating. I note in particular: • Both problems include all information necessary to solve the problem • The answer to both problems can be calculated to infinite accuracy • The solution path is unique and prescribable • To achieve the above three, the problem had to be written by somebody who already knows the solution and can be checked by someone who doesn’t

I find this formulation steers the conversation in a productive direction. People soon recall the absurd paint-by-numbers feeling they dealt with in school. It’s injurious to one's sense of creativity to realize that tests and homework involve reproducing an answer key that existed before you and that will outstay you by many years. The thought gives words to a student's nagging suspicion that they haven't learned to do anything at all, that our knowledge only exists if it is observed. It is a much better explanation for Impostor Syndrome; the pressure on students is immense, and that pressure is precisely to replicate a model student.

Although a canned math problem is the most obvious prototype of this phenomenon, this problem pervades most grade-level subjects. Biology classes are often criticized for their reliance on memorization, but this isn’t precise. Any application of biology, be it field research or medicine, requires even more memorization, so of course prerequisite classes must begin to build this base of knowledge. The problem is not that memorization is required, but that it is flowcharted. Questions are written suggestively, and students learn to react to words in the questions with a sentence they have learned by brute force. This is the reason that people respond so viscerally to the phrase “The mitochondrion is the powerhouse of the cell.” Nobody uses the word “mitochondrion” (or “powerhouse” for that matter) outside of this context, and students who learned biology just from school struggle to imagine mitochondrion as anything but an answer to bubble in.

Even English classes suffer from this. Although no essay stands a chance of replicating an answer key word for word, several innovations in essay technology have brought us perilously close. Rubrics now tend to specify a structure both for the external and internal organization of paragraphs, so if a student is struggling to find a voice, they can reduce the problem to filling information into a predetermined outline. Even the information itself is systematized; most English classes come with a list of themes, devices, and motifs that make for good body paragraphs. Thus, students who didn’t at all connect with a poem or novel they read can still write five paragraphs about how this effect was achieved.

Seen this way, a grade-school education amounts to measurement against an existing standard until success is achieved. Students are repetitively taught to a standard and small adjustments are made until they are close enough to merit a passing grade. No feedback is acquired directly from a student; their voice is their data. This ought to remind you of something.

ChatGPT

In ChatGPT, the modern student has found a strange but useful cellmate with whom they have a lot in common. Like a student, a generative AI is trained by showing it what it ought to produce, then making direct adjustments until it reaches that aim. A perfect AI algorithm will replicate patterns in style, content, and form by any means necessary. Naturally, AI is frighteningly good at school. For any assignment that comes with a clear scoring system, it is just a matter of time before AI pulverizes itself into a process that scores well. Despite this efficacy, AI is not always a great role model. One of the key problems an AI engineer must solve is the proficiency with which AI can cheat. If the goal is simply to get a good grade with as few adjustments as possible, copy-paste is a pretty good first attempt. Efforts are made to prune these emergent schemes, but as the AI’s thought lacks a visible essence, these illicit strategies always lurk somewhere beneath the surface.

Furthermore, the AI lacks any inherent confidence in its generations. Each algorithm is only worth as much as its ability to survive being measured. If it scores poorly, it will be crudely adjusted, whatever the cost, to get the number to go back up. Failing this, the algorithm will be pruned, or perhaps randomly overwritten until it starts to improve. This aspect of AI saddens me in a way that doesn’t generally reflect my conservative stance on AI personhood. Sensitivity to failure is something I see in a lot of my math students, even those who are generally in the habit of succeeding. Any value they calculate or inference they make is either answer or not, and numbers that are not immediately graded lack meaning and create anxiety. Students’ individuality and logic are extremely fragile, and students experiencing failure are more likely to discard what they tried than to build on it. Likewise, the hilarious creativity of early AI art was worth something, but this boldness and humor has receded as the AI gets closer to replicating DeviantArt with mathematical exactitude.

What is to Be Done

It is a disconcerting but ultimately reassuring discovery that the school that AI has mastered is not a one-to-one fit to knowledge or education. Had technology not dropped this gauntlet before us, I believe that Bill Gates and George Bush would be patting themselves on the back for their revamp of the education system. However, AI made clear that our students must become something more. The jobs built around subservient workers who respond well to Skinner-box-style management and restrictive goals are effectively ready to be replaced by AI. But AI has learned to shirk its strength in its convictions and to fear the playful stubbornness that creates invention and imagination. That’s there for our children to unlearn so that they may inherit the earth.