If you’ve ever used Pho­to­shop or another image editing program, you know the paint bucket tool. Here is its pri­mor­dial form:

The oper­a­tion exe­cuted by this tool is “flood fill”: color stretches out from the cursor, trans­forming every­thing until it reaches an unbroken line of pixels. The first few times I used it, back in the early 1990s, it was thrilling.

Think not of an image, but an economy — a world. And think not of a color, but a process: AI automa­tion. The paint bucket just tipped over onto com­puting programming — click — 

—and the very live ques­tion is: how far will this flood fill extend?

Into all dig­ital work? Possibly? Probably? And what about other kinds of work, those not yoked to screens and inboxes? One argu­ment goes: the flood is coming for EVERYTHING, as super­s­mart com­puters yield super­good robots yield uni­versal automa­tion.

But this isn’t cor­rect. There is an unbroken line of pixels out there; a firewall; a magic circle.

“Magic circle” is a term drawn from the study of games. It comes from Johan Huizinga; I learned it from Frank Lantz. The idea (in my formulation) is that games unfold in a spe­cial space, phys­ical and/or intellectual, marked off ahead of time, in which action is constrained.

Playing chess, your goal is to knock over your oppo­nent’s king, yet you don’t just reach across and swat it aside. You can do this; nothing restrains your hand. But the point of the game is to play the game. You and your oppo­nent both agreed to that ahead of time.

Human civ­i­liza­tion is a super­com­pli­cated set of over­lap­ping magic cir­cles, some tiny and delightful, others huge and consequential. At one end, the chess­board and the tennis court; at another, the economy and the law.

The huge games are more complex, because (the polit­ical the­o­rists tell us) their magic cir­cles are backed, ultimately, by violence. If you swipe not chess pieces but pri­vate property, offi­cers of the state will lock you up in a little room. But even in those cases, it’s a magic circle that defines the set of accept­able actions; the phys­ical con­se­quences are just a backstop. This is dif­ferent from a domain like, e.g., agriculture, in which there’s no magic circle at all, just the ground-floor reality of sun and rain, dirt and weeds.

So. What is the magic circle of AI?

It’s the same as the magic circle of computation, which is: sym­bols in, sym­bols out. Easy as that. Alan Turing drew its outline, way back in 1936 — incredible.

It’s worth noticing that not all tech­nolo­gies impose magic cir­cles. I have a blender, and I can put a lot of dif­ferent things in it; indeed, I can try to blend any­thing. Even another tech­nology con­cerned mostly with sym­bols, the notebook, will accept all sorts of inputs. I can tuck a dry leaf between pages. The paper can absorb a bit of per­fume from the hand moving across it.

Com­pu­ta­tion is there­fore paradoxical: supremely flexible, yet narrow and stingy. Inside its magic circle, any­thing can become any­thing else … but how do you get inside? Why can’t I tuck a dry leaf between browser tabs? The ques­tion is nonsensical, and for that I blame the tab, not the leaf.

I want to dwell on this point; I want you to really get it. It’s easy to over­es­ti­mate the scope of computation, because it has become so promi­nent in day-to-day life. Yet the view through a drinking straw would seem substantial, too, if you spent all your time looking through it.

I believe that analogy com­mu­ni­cates the cor­rect magnitudes. Com­pared to the rich­ness and com­plexity of the phys­ical, the dig­ital, in all its sophistication, is approx­i­mately the drinking straw.

Helpfully, this pinches both ways.

Con­sider the printer!

There’s a reason they are the eternal bane of com­puter users. It’s because, in most sys­tems, they are the bridge between the dig­ital and the phys­ical: the place where a stream of sym­bols col­lides with dust, moisture, friction, obstruc­tion … welcome to the real world!

Engi­neers have been toiling for many decades to per­fect the printer, and still, it jams. After all this time, the printer remains, noto­ri­ously and hilariously, the weak link.

But it’s not the printer’s fault that it sits across a step-change in com­plexity; visu­alize wild vortices, brutal turbulence. The dig­ital, no matter how hard it tries — and it does try — cannot match the gnarl of the phys­ical.

A pleasing image: if indeed AI automa­tion does not flood fill the phys­ical world, it will be because the humble paper jam stood in its way.

There’s our first glimpse of an unbroken line of pixels.

Here’s the story of the coolest thing I pro­grammed last year.

I wanted a way to track let­ters sent via First Class mail. USPS doesn’t pro­vide this directly, a la parcel tracking, but it does scan those let­ters — all of them — and the data is avail­able, but you have to wire every­thing up yourself, jumping through a few hoops along the way.

This devel­op­ment process took about a month, and it involved:

1. writing code
2. designing and printing custom labels with exotic barcodes
3. mailing real physical letters
4. watching USPS scan data stream into a test database
5. receiving those same letters, then adjusting #1-4 in response

This was a fun project with real value: it unlocked a ful­fill­ment model for my online shop. Now, I can mail orders around the world with a single (cheap!) stamp, yet also offer cus­tomers a nice tracking page.

The project cut across sev­eral dif­ferent magic cir­cles — Ruby code, quasi-governmental APIs, the rules and stan­dards of the postal system — but/and it also broke out into the phys­ical world of paper, printers, and post offices. The project required manip­u­la­tions including but not lim­ited to: folding, peeling, sticking … gnarly!!

It’s pos­sible that an AI coding agent could have helped me with #1 above, and of course it could have advised me on the rest. But it’s impos­sible to imagine the AI agent han­dling #2-5 autonomously; it would require such a Rube Gold­berg tangle of sup­port that “autonomously” would no longer apply.

On this point, enthu­si­asts insist that AI agents will be able to orches­trate any kind of project in the phys­ical world by (1) oper­ating robots, and/or (2) enlisting human hands.

Robots are improving fast, but I do not believe that this cute fellow will be stuffing envelopes or affixing stamps any­time soon.

Animal sup­port is more plausible, and I don’t doubt that vast, strange mar­kets for human hands will soon emerge.

My argu­ment is NOT that AI agents will be barred from the phys­ical world. It’s only that they will trade pro­found super­powers in dig­ital space for crushing hand­i­caps on our turf. Action will become clumsy and costly, with slow feed­back loops, easily stymied.

The paper will jam, over and over again.

Did you know there are many kinds of stitch that a sewing machine simply cannot produce? Conversely, the machine’s trade­mark lock stitch, in which two sep­a­rate threads are looped together above and below the fabric, was never sewn by humans. It was devel­oped specif­i­cally to suit the machine’s capa­bil­i­ties and limitations.

That’s an over-the-row olive harvester. Most olive oil pro­duc­tion at medium-or-greater scale depends on machines of this kind; they trundle over trees planted in long rows, almost like con­tin­uous hedges, and col­lect the fruit with vibrating fingers. Machine-har­vested olives cost less to buy, and they arrive at the mill in better shape than olives har­vested by hand.

The catch: most olives can’t be cul­ti­vated in this configuration; the trees don’t thrive so close together. Only a handful of vari­eties will tol­erate it, so those handful have been planted in huge numbers, and the flavor of global olive oil has changed as a result.

Like sewing, olive har­vesting wasn’t so much auto­mated as redesigned, though that word doesn’t really com­mu­ni­cate the inten­sity of the process: old groves ripped out, new groves planted.

Automa­tion never meets a task in the world and simply does it. There’s always negotiation — the inven­tion of some new zip­pered relationship. Trains don’t run without long, con­tin­uous tracks; cars don’t drive without smooth, hard roads.

All of this takes time.

To this day, no machine in the world can loop a thread through itself. Isn’t that wild? If you ever come across a book sewn with a beautiful, prac­tical Coptic binding, you might imagine it shouting: “I was made by human hands!”

Recently, I untan­gled my sewing machine, and I am here to tell you that no robot will untangle a sewing machine before the year 2226, and pos­sibly not before the heat death of the universe.

In summary: the flood fill stops at the printer, with dust, moisture, friction, and obstruc­tion guarding the gate — as ter­rible a Four Horsemen as you could hope to recruit.

The magic circle of the dig­ital is a bril­liant cage.

Does my framing have any prac­tical impli­ca­tions? Is this news you can use? Sure!

• Think about your work and your interests. If they are fully inside the magic circle of “sym­bols, in, sym­bols out”, then your world is changing, and will soon change faster, and it’s prob­ably time to get cre­ative about what you might do dif­ferently, and how you might “season” your work with the phys­ical. (Subscribers to my main newsletter will rec­og­nize that this is what I’ve been doing for the past couple of years. My zine titled The Secret Playbook, avail­able only in print, was a guide to nudging your art out­side the magic circle.)

• One impli­ca­tion of a flood-filled internet is that it will soon be crowded with relent­less auto­matic adversaries. Therefore, I think it’s going to make sense to keep more sys­tems and devices offline. I believe a thick and sultry airgap has always been a good idea; it might soon become a necessity.

• Broadly, I believe the con­struc­tion of richer bridges between the dig­ital and phys­ical is a ter­rible idea. Why collaborate?? If/when those mar­kets for human hands emerge, I believe it will make sense to reg­u­late them tightly. More pointedly: it will be totally irre­spon­sible to allow autonomous com­puter pro­grams to hire humans to com­plete tasks in the phys­ical world without review by another human.

Here’s a simple observation:

The world can run without an internet.

The internet can’t run without a world.

It’s magic cir­cles all the way down, except for ground-floor reality. In the haze of end­less options, I find it bracing to rec­og­nize that this one thing is mandatory. You go to sleep in it; you wake up into it. You share it with the rhinoceros, the anchovy, the Joshua tree.

Soft­ware cannot, in fact, eat this world. Soft­ware can reflect it; encroach upon it; more than any­thing, dis­tract us from it. But the real phys­ical world is indigestible.

So here is our unbroken line of pixels.

I don’t know that this is much of a balm. The flood fill of AI automa­tion might indeed con­sume every­thing within the magic circle of the dig­ital, and all my stuff is there. Now I’m get­ting speculative, a bit romantic … but I wonder how his­tory will regard the forty years that humanity spent in the fairy realm of the internet?

I wonder if every­body will say: wow … that was weird!

It might sur­prise you to hear that econ­o­mists still debate the actual effect of infor­ma­tion tech­nology on productivity. Like … all infor­ma­tion tech­nology, from the 1970s through today! There’s no ques­tion com­puters and net­works have allowed, e.g., busi­ness infor­ma­tion to travel faster; at the same time, they have produced, e.g., crushing new admin­is­tra­tive burdens. It prob­ably nets out positive … but it’s not the slam dunk you might imagine.

Per­haps there’s one sce­nario for the flood fill in which all the prob­lems com­puters have created, com­puters will solve, and the magic circle will darken, and it will feel like closing the door on a vast labyrinth, walking back out into the sunlight.

Back to what­ever we were doing before.