How to Explain AI to a Doomer

Quite often, when someone who is not an academic discovers that I research AI in education, they say that AI scares them. 

Fair enough, but which AI? The LLM that dreamed up the image above? I asked it, when I was last teaching my class "Writing With and About AI," "what sort of donut would you eat, if you could eat?" The answer came immediately, that it would choose a donut with icing and sprinkles, because the donut not only looks delicious but is pretty.

 If only more humans could reason that well. 

So let's back off the snark and help folks who have many fears about AI. My answer goes like this:

• Agentic AI is the one that is in the news. It does tasks without a human always looking over its shoulder. It can follow its instructions too literally. To get it done, it might lie, cheat, or hack. It needs some regulation. Our President, who knows nothing about AI, opposes regulation. The people who are building these models, particularly the most advanced "frontier models," want regulation. I side with those fellows and not our unread President on this issue.
• General Large Language Models (or LLMs) are the ones you might have used to depict a robot dreaming of a donut. They are not going to kill you, directly. But every query with one uses power in a data center that drives up electricity costs, drinks water, and is contributing to climate change. You decide what to do. Most of them are what we call "unsupervised models" because they range far and wide in their tasks.
• Specialized models are the ones that do things like sorting large data-sets for scientists, help your map app find the best route to drive on a particular day, or turn text input into podcasts. They are not dangerous as they don't go rogue like they used to, by directing drivers into lakes. 
  
  These AIs also don't get used nearly as much as those LLMs that kids use to cheat on tests or that transform images of fat men in basements into svelte superheroes. The danger of these systems is that bias can enter their training as "supervised models," so they might ID more people of color or young men as potential criminals, when used with facial-recognition cameras.

The bottom line? We need to better understand what "AI" means. Then we can best control the most pernicious varieties.

I'll be honest about my own bias: it may be the sycophantic tone of many AIs, but I prefer them to most humans. At the same time, I choose to deal with humans, in all their meaty imperfection. I've used an LLM exactly twice in the year since I taught my course. 

But as writer Fran Wilde once told me, I'm a curmudgeon in danger of becoming a misanthrope. Fair enough, too! I'll soon be on her and her husband's boat angling for Rockfish, so they can toss me overboard when I get too snarky.