Boxes and boxes, why doesn't it fit in all my little boxes

Every time I read a post about Artificial Intelligence, I have to first figure out what they mean by AI. We created a big box of "AI" and we all have our own definitions of what goes in and what is kept out. Usually, the poster is talking about generative AI, and they've built something that helped them observe something. Yeah, me too, because now that I've finally started using AI effectively, I updated my website and got back to stacking boxes in Cascading Style Sheets, frustratingly mumbling "why don't you just fit in the box" to a line of content. Is it satirical or derivative? Depends on how you've defined your boxes. But really, we're lacking nuance. And I know, I'm late to this critique, that we are all talking about AI without taking the time for nuance or to see what our boxes have in common.

Back in the mid-2000s, when I was an eager and self-assured philosophy undergrad, debating the definition of artificial intelligence was just a fun way to pass some time. I wasn't wondering if spell check was AI, I wanted to know what it meant when a machine was naming brand new things. Kicking around the consequences of an unmanned Mars expedition looking for new minerals, that if the probe/rover did find something new, it would have to name it. And to name it, would it try to cram the discovery into as many programmed boxes as it could? Would it follow a strict protocol (rules are a box)? Or would it spit out a "random" string of characters and send it to humans to classify? Would that naming be a genuine meaning-conferring act, or just mechanical token-generation (geez that feels familiar)?

No one assigned the Mars version specifically, but it was the kind of question our Philosophy of Science course kept circling. From Kripke and Putnam's causal theory of reference (roughly, that a name gets fixed to a thing by an initial act of dubbing, then passed down the chain of everyone who uses it) to Harnad's symbol grounding problem (whether a machine's symbols can mean anything on their own, or only ever mean what we decide they mean), there was real thought being spent on what naming even is. So when the rover dubs its new mineral, is it performing one of Kripke's "initial baptisms," and if it is, does that baptism actually confer meaning, or is it just generating a token? Just the idea of a robot naming something was enough to get philosophers and computer scientists into proper dust ups, fervently defending the boxes defined for us. And meanwhile, ChatGPT will name your variables without even pausing for your input.

And because of our love of boxes and putting things in boxes, we build things like typologies to classify things cleanly into new boxes, or nested boxes, connected boxes, adjacent boxes, etc. It's turtles in boxes all the way down. I could build you a quick set of boxes: "there are three types of AI commentators: the techno-utopian (pro-tech, pro-singularity), the Luddite (pro-torches and pitchforks, pro-clean water and air), and the fence-sitter (just waiting to see who is going to win; pro-Betamax and VHS, pro-employment)," throw it up on the right subreddit, and just wait for some version of Cunningham's law to illustrate the problems with my model. And the problem isn't just that I've leaned on 200 years of slander against Ned Ludd and textile workers who wanted fair wages and some guarantee their jobs were safe (geez, that feels familiar).

Our love of boxes leaves little room for nuance, and once some nuance has been carved out, it's usually the size and shape of a new box. Whether we are moving the goal posts about what it means to be artificially intelligent, or making assumptions about how colleagues actually feel about GenAI tools and the impact of LLMs, we are classifying the conversation in ways that are constructive and effective for a number of purposes, but leaving out a lot of the theory-to-practice in between.

One of the most common struggles I experience when reading about practical AI use, is that one of GenAI's greatest strengths can make my outputs completely ungeneralizable. If one chooses, GenAI can customize and individualize to a very personal level. My AI workflow and support systems are amazing and have unlocked a new level of scholar and critical thinker for me personally, but no one would ever need the .md files from my Git repo. Why would they?

I've renovated this house so specifically, so completely to my own shape, that it's unsellable. Every wall I moved made it more mine and less anyone else's.

I built my boxes that fit me perfectly, as many GenAI users have, but what good would my standing_instructions.md be to someone else? They don't fit in my boxes.

All of our experiences with GenAI are so specific that on some level, none of us are talking about the same thing. And the speed of everything, even without AI, can penalize us for stopping to actually listen. Slow down long enough to hear your peer out, and it turns out their worry about the climate footprint of data centers isn't anti-AI at all. They're just trying to decide whether they can use Gemini to write a letter to their senator. And that excited friend who is using AI to help organize their life isn't really saying they have handed their agency to an LLM, but that they developed a project in Claude Cowork that helps them manage some of their neurodivergent habits via offloading.

Have your boxes. They help keep things organized, right? But you'll never have enough of them, because the world keeps handing you things that don't fit. Your context window isn't that big.

References

Harnad, S. (1990). The symbol grounding problem. Physica D: Nonlinear Phenomena, 42(1–3), 335–346. https://doi.org/10.1016/0167-2789(90)90087-6

Kripke, S. A. (1980). Naming and necessity. Harvard University Press.

Putnam, H. (1975). The meaning of "meaning." Minnesota Studies in the Philosophy of Science, 7, 131–193.

← All essays · Also published on Substack.