“Unpopular Opinion: Teacher AI use is already out of control and it's not ok”

My takes:

(1) AIs for "discriminative” tasks of judging “truth” vs judging “value” — AI can be great at TRUTH judgements, that have a “true” or “false” answer — but is HORRIBLE at VALUES judgements, that have a “good” or “bad” answer

E.g. if you give AI an objectively-verifiable list of criteria to judge by, where they are all objectively true or false, with no questions of value (i.e. is this good or bad) — then AI CAN do ok

And this is exactly what I’ve done for all AIs used here — the points awarding AI, the newsworthiness-checking AI, the image-explicitness-checking AI — every single one of those gets handed a list of explicit questions of truth to evaluate

BUT THE MOMENT YOU ASK IT ANY QUESTIONS OF VALUE, e.g. is this a good meme, or is this explicit, etc — absolute disaster, absolutely insanely bad, you cannot even imagine

So, if the teachers are asking questions of value when grading student work, which you can bet they are — absolute insanity

(2) AIs for “generative” tasks of creating some judgable dividing boundary, for “summarizing” vs “surprising” types of generations — AI can be great at “summarizing” types of generations, but horrible at “surprising” types of generations

Basically, as many have realized, and as even PG eventually realized — Good writing has to focus on the surprising, as unsurprising writing is a total waste of time. This extremely applies to teaching as well.

If the AI is just summarizing some writing that already has the surprising stuff figured out — then the AI can nail it. But if you try to get the AI to come up with the surprising stuff itself — Horrible disaster

Which is a lot of what this guy is talking about, with AI generating repetitive, useless teaching content — AI totally lacks the ability to identify where is are the surprising parts, that can shortcut readers into learning much faster. At least that’s a huge part of it. FYI jokes/memes are another type of content that critically rely on an element of surprise.

Here, in the upcoming AI systems, again sidestepped this problem, by ONLY having AIs doing the descriminative part, curating and pruning content — NOT the generative part, which again ultimately is left in the hands of humans, for now

— So, at least for today’s off-the-shelf-LLMs:

+ Descriminative AI should only be used for questions of truth (true or false), never for questions of value (good or bad), with all value questions rephrased as questions of truth

+ Generative AI should only be used for non-surprising generations (write code fitting these specs, rephrase this already-good teaching material), never for generation of content that requires an understanding of surprise (generating good teaching materials, generating good jokes) — or at least you cannot directly ask today’s AIs to generate these things

These are where these teachers went horribly wrong,

Lessons we’d already figured out long ago

BTW reminder that teachers often have among the lowest IQs of professional fields, especially teachers of young kids

🄳🄾🄾🄼🄿🄾🅂🅃🄸🄽🄶