Borrow the structure, not the surface

A diffusion model turns noise into an image by the same underlying process that turns a supercooled liquid into ice: a small instability that grows fast enough to reorganise the whole system before it settles. That's not a metaphor a marketing team wrote. It's the actual mechanism, and a physicist named it first, decades before anyone trained a diffusion model.

The paper that borrowed the structure, correctly

A March 2026 paper makes this explicit: pattern formation in a trained diffusion model can be explained as an out-of-equilibrium phase transition, the same class of instability that describes how physical structure emerges from disorder. Low-frequency patterns become unstable, correlations grow rapidly, and noise reorganises into something structured, at a predictable critical point the theory lets you calculate.

What makes this a good borrow, not a loose one, is what got carried over: not the language of physics, the actual mathematical relationship between instability and emergent structure. The paper doesn't say diffusion "feels like" thermodynamics. It shows the same equations apply.

Design research just formalised the same move

A CHI 2026 paper, reviewing 85 studies, makes the design-side version of this argument: analogical thinking, borrowing a structure from one domain to solve a problem in another, works when it's applied throughout a project, not just as a five-minute brainstorm at the start. The authors' actual worry is that foundation models push the opposite way, flattening ideation into an input-output loop that never reaches for anything outside the prompt.

The test that separates a real borrow from a decoration

The diffusion paper hands you the test for free, if you look at what makes its borrow work. The physics doesn't apply at one moment. It applies to how instability turns into order over a sequence of steps, and it holds at every point in that sequence. So: if the borrowed idea can be evaluated from one still frame, you copied a surface. If it only holds up once you trace how the parts move against each other over time, you copied a structure.

"We're the Uber for X" fails that test instantly: it's a one-frame comparison between two surface facts (an app, a driver) with nothing relational underneath. The diffusion paper passes it because the relationship between instability and emergent order holds at every timescale you check it against, in a liquid and in a trained model alike.

The five-minute version of doing this on purpose

Before your next hard design problem, don't ask "what's this like." Ask a more specific question: what other field has already solved a problem with the same relationship between its constraints, not the same surface details. A hospital handover and a kitchen pass have nothing in common as objects. They share an identical relationship: information has to survive a handoff between people under time pressure, with no room to re-ask.

That's the five-minute reframe. It costs nothing but the discipline of finishing the sentence "this is structurally like…" instead of stopping at "this is kind of like…", which is where most borrowed ideas quietly become decorations instead of solutions.