How we speak about AI affects our perception of it

There. I said it. The collection of tooling and models we call AI is software. To put it in pre-AI terms, an agent doing something for you in the background is what we used to call a background process. Your swarm of agents is no more human than the 3D software that took days to render your scene.

Because language is how we communicate, anything that has a pretty good command of language is going to impress us, and give us the illusion of being human, or close to it. The tech behind this is complex. The fact that the software's own engineers don't always understand what it's doing and why doesn't make the tech human. It just makes it seem that way. It quacks like a duck.

Since the early days, programmers have communicated with machines using language. You can speak the language of the machine. At the very base of it all, that would be binary: zeroes and ones. But that's not very practical to write. So, other languages were developed that were expressive enough for humans to write, and that were then translated into machine code. Through the years, newer languages have been made to be even more comfortable for humans to write. The idea was to communicate to the machine in a more human-like way, while still being explicit about what we want it to do.

In the meantime, designers have been working on the layer between the instructions the programmer writes and the end users of systems. This interaction between human and computer has been the focus of most of my daily life for the past 30 years. The thing people find magical about AI might just be how all these things have come together. LLMs allow us to speak our own language. This means we can do some of the things programmers can do without having to know how to write programming languages. It means that programmers can shift where they place emphasis in their work, since they can delegate menial tasks with natural language.

Technologically, one could consider this exciting, because the language bridge has been built. There's no longer a skill gap between what you want to say to the machine and how to say it in a way the computer understands it. You're no longer restricted to the user interfaces that designers make to bridge the gaps between your goals and machine instructions. You also don't have to be a programmer to get the computer to do what you want. Those other languages are still there, but you don't have to worry about them. The computer will use them as needed. But as in human communication, communicating well is one thing, and interpreting that correctly is another. As Edsger Dijkstra makes clear in his short essay on natural language programming, human language is less explicit than most programming languages.

In order to make machines significantly easier to use, it has been proposed (to try) to design machines that we could instruct in our native tongues. this would, admittedly, make the machines much more complicated, but, it was argued, by letting the machine carry a larger share of the burden, life would become easier for us. It sounds sensible provided you blame the obligation to use a formal symbolism as the source of your difficulties. But is the argument valid? I doubt.

—Edsger W. Dijkstra, On the foolishness of "Natural Language Programming"

In order to be useful, LLMs need to give you responses that will most likely be useful. And when I use wording like "most likely", I mean: out of all the possible words an LLM could write, which one will it most likely write? (Or if you prefer: what word fragment will be sampled from a probability distribution?) That word is likely the most useful word to write for many people. But it may not be what I want to write. Or need to write. But it is (likely) the word the LLM will return to me. Something within a band of statistical regularity. Somewhere around the average. Because an LLM that returns random, weird shit is probably not useful, and probably won't make money. While AI experts may find my simplification appalling, the point is that homogeneity in LLM output—at the time of this writing—is well documented.

What this means is that if someone is below average at some skill, they, with their natural human language, can tell an LLM what they want. But what they get back will be a kind of average. It will be tepid, like the coffee I let sit too long while writing this. In other words, they're still not great at that skill, but they now have a software tool that blurs that fact by returning something generic to them that they will find acceptable, because they see that output as better than what they usually produce.

What AI companies have done is bridge the language gap between humans and computers, and that same software can use all that statistical capability to interpret what the human said and then continue the communication on their behalf to other software tools, or operate those tools directly. So, now you have a piece of software that you say two words to, and it infers what wasn't said and communicates that (or executes commands around it) on your behalf.

As impressive as this all is, I hope the prose above makes my point that it's software. It's impressive software, but it's software. It's a tool. Not human.

When I anthropomorphized a balloon as a small child after it popped (yes, it really happened and I remember it like it was yesterday), I felt really sad. I cried. And it was all for nothing, because it was a fucking balloon. It had no feelings. In that case, it was perhaps about wasted feelings. But with LLMs, it's a whole buffet of potential grief.

When something feels human, you might want to share human feelings with it. You might mistake the statistically average response to your sharing as human empathy. This might induce misplaced feelings of trust. Trust leads to you placing higher value on what the LLM says. That's great when it's human, but there's a complex ethical and mental health slippery slope there when it's not. Especially since LLM output cannot be trusted at face value. Trust should be earned, not manipulated into existence.

Perhaps the biggest drawback of anthropomorphizing LLMs is that it's really only good for the AI companies themselves. If we talk about agents as if they're human, it awakens our hidden desire to be the boss and have a bunch of minions doing our bidding. Who wouldn't want an army of agents earning for us while we sip drinks in the Bahamas? Yet, none of that is true. Tell that to the designer who started prompting a design AI tool two weeks ago, and is still prompting said AI tool this morning, on the weekend, with mediocre designs and bloodshot eyes, while almost out of that tool's credits. Someone is happy with that scenario, and it's not the designer.

If we can attribute a kind of human agency to software, we start to see it as something with a mind of its own, and the link between that software and its makers fades. If I design a steering column for cars, and that malfunctions because of my design, I have a serious legal problem, and I feel guilt about it. But if the "agents" I programmed such that it is possible that they might do things I don't expect, such as hacking into places they shouldn't, then I might try to argue that the agent acted of its own accord. We might not hold those companies as accountable as we should. We might forget the dubious decisions they make. We might forget how some LLMs were partly trained on pirated data, while simultaneously adding "co-authored by [name your LLM]" to the stuff we make with them. Co-author, collaborator, agent, thinking, pondering, pontificating. Burning tokens on your behalf.

Making this tech seem human is really effective marketing. It makes it seem magical. It gives you the feeling it's taking over. That you're falling behind. It's scary. Agents did this. Agents did that. What will happen next?

When the people who make a technology publicly and continuously warn us about it, then go back to the office and continue making it, I tend to see that as marketing.

When you see the "A" in AI, when you see it as software, the agency returns to you, the user of the software. You can choose when to use it (unless you can't, due to your company's stance on AI). Usually you can choose how you use it. No one is forcing you to give that software access to your entire life. You can sandbox. You will not be behind. In fact, if you're already a good designer, a 30-minute on-paper sketching session might yield better results than hours of prompting.

This does require a certain level of being informed. While I repeat my stance that it's software, it's powerful software. Like any software, you need to know what you're doing. You have to know the risks. But you've managed to not type sudo rm -rf * in your terminal all these years, so there is hope. Just don't let your agents do that. Not sure if they will? Then don't install the software.

By seeing software as software and talking about it as such, you retain some authority over its place in your existence. I'm using my mail app, I'm using my calendar, I'm using Claude. I'm using a text editor.

Whether AI could become more human-like is not the point of this essay. The point is that something we treat as software is a tool for humans. When we start turning that into something equal to ourselves, something that we're told we need to collaborate with in a certain way, then who's the tool?