Danah Henriksen and I recently wrote a paper, currently in press, titled “The Mirror and the Black Box: AI Metaphors and What They Mean for Learning.” It’s about how the metaphors we choose for AI shape what we can and can’t think about it. The paper traces a continuum of metaphors and concludes by asking what it means for how we think about learning. In short, it is a paper about how framing shapes understanding. So, I started wondering: what happens if you reframe the paper itself?

My first attempt was a direct imitation of something I had seen my friend Leon Furze do for a beautiful page he published recently called The AI Iceberg. It was a single scrolling page where the metaphor of an iceberg organized both the content and the experience. You didn’t just read about layers of complexity beneath the surface. You scrolled down through them. The form was the argument.

I looked at it and thought how would that work for the paper that Danah and I had written. So I began, as all creative people do, by imitating Leon. I built a seven-act vertical scroll, driven by animations, with interactive flip cards on a color spectrum to represent the different metaphors. It had a historical timeline that one could scroll, and a dramatic arrow “flip.” It was cinematic. It worked, but it didn’t feel right.

It was too long. By the time you reached the end, you’d forgotten the texture of the beginning. You couldn’t see the whole argument at once. The format was linear, but the argument wasn’t. A scholarly paper isn’t really a story that unfolds from beginning to end; it’s a central claim surrounded by the voices that complicate, deepen, and challenge it. This design could not do that. The design was a container for the argument, not an expression of it.

What happened next wasn’t a design decision so much as a recognition. I’d been circling this idea for a long time without knowing it. Years ago, when I was a junior faculty member at Michigan State, I rescued a box of papers from a pile headed for the trash. In that box was a 1979 paper by Lee Shulman, and a quote from it stayed with me. (As it so happens, Shulman later became a hugely influential figure in my intellectual life. Matt Koehler and I extended his work on Pedagogical Content Knowledge to develop TPACK). But in that paper, it was Shulman’s way of writing, his way of layering argument with commentary and historical context, that moved me, and shaped my sense of what good scholarly prose could be.

As I have written elsewhere, he described his description of the analysis as being akin to the Talmudic view of four levels of Biblical exegesis:

p’shat: explication of the plain meaning of the text; d’rash: interpretation of plain meanings; remez: broader inferences based on discerning nuances or “hints” from the text; and sode: barely bridled speculation soaring effortlessly from the text and tethered loosely, if at all, to its sources (normally the level of exegesis assigned to the Kabbala, or mystical literature). The reader should be forewarned that there will appear little interpretation of the first kind, much interpretation of the second and third kinds, and occasional forays into the dangerous altitudes of the fourth.

That layered quality of the Talmud was what resonated. The Mishnah sits at the center of the page: the core text, the law, the argument. Surrounding it is the Gemara, generations of rabbinical commentary, each voice responding to the text and to each other. It is, arguably, the original hypertext. Centuries before Ted Nelson coined the word, the Talmud had this architecture of core text surrounded by annotation, with marginalia pointing outward to yet more commentary. I’d been interested in hypertext since the early days of the web. I’d been shaped by Shulman’s layered scholarship.

And here I was, trying to find a form for a paper whose argument is that the frames we put around ideas determine what those ideas can mean.

The Talmud model solved everything. I compressed the entire argument of the paper into a single paragraph of about 180 words. Seven sentences, each flowing into the next, readable in thirty seconds. That paragraph sits at the center of the page. Flanking it on both sides are commentary panels. Click on any phrase in the core text and the commentary expands: the taxonomy of metaphors, the historical timeline, the analysis of anthropomorphism, the references and connections. The paragraph works simultaneously as prose (skim the whole argument), as navigation (click any phrase to go deep), and as a keynote spine (point to a phrase, expand it, teach from it).

But having the concept and having the thing are two different experiences. The instantiation took a while. What does a Talmud-inspired layout actually look like on screen? What typeface carries the weight of a “core text” without looking pretentious? How wide should the commentary panels be? How do you handle the transition when a reader clicks from one annotation to another? These aren’t cosmetic questions. Each one is a decision about what the reading experience means. The choice of EB Garamond for the body text, Playfair Display for the titles, a warm cream background instead of white, these emerged through iteration, through building and looking and adjusting. The aesthetics of the interaction, as it were, took time to find.

And of course, I built this entire thing in conversation with AI. Not by writing code, but by describing what I wanted, reacting to what appeared, pushing back, trying again. I wrote about a similar process recently when I redesigned three old web projects from the early days of the internet. In that post I described working with Claude as a partnership where “I really had no clear idea of what the final result would look like. The design emerged out of this dialogue.” The same thing happened here, but the stakes felt different, because this time the content wasn’t a satirical glossary or a set of palindromic poems. It was a scholarly argument. And the tool I was using to reimagine that argument, the AI, was also the subject of the argument itself.

There’s a recursion to that which I find genuinely interesting. The paper says: the metaphors we use for AI shape how we understand AI. And here I am, using AI to build a new metaphor for what a scholarly article can be. The tool that the paper theorizes about is the tool that helped me reimagine the paper’s own form. That recursion raises questions I find more interesting than the technology itself. What does an article mean? What might an academic argument look like if we stopped treating “article as document” as the only option? “Article as document” implies: linear, self-contained, authoritative, finished. What if it could also be: layered, conversational, expandable, alive?

The Talmud figured this out a long time ago. A text is never just a text. It’s a text in conversation with every reader and commentator who came before, and every one who will come after. The margins aren’t secondary to the center. They’re where the thinking happens.

I don’t know if anyone besides me will find this particular experiment compelling. But I know this: I’ve always been able to imagine how scholarly work might look and feel in a different medium. What I couldn’t do was build it to the standard in my head. I could hand-code HTML, sure, but I couldn’t get the drop cap right, or the three-panel layout, or the way commentary should slide into place when you click a phrase. The design in my mind was always better than what my technical skills could produce. That gap is what’s closing. And that, it seems to me, is what is genuinely new about this moment.

From hand-coded HTML in the 1990s to conversations with AI in 2026, the tools change but the principle doesn’t. You make something. The something talks back. And if you’re paying attention, you end up somewhere you didn’t expect.

How cool is that.

The site is at punyamishra.com/ai-metaphors. Take a look and let me know what you think.