The Software That Grew Hardware

On June 18, Midjourney — the AI art startup best known for dreamlike, six-fingered, copyright-adjacent renders — announced Midjourney Medical and a piece of actual hardware: a full-body "Ultrasonic Computational Tomography" scanner. Around half a million ultrasonic sensors across forty Butterfly ultrasound-on-chip imaging modules, no radiation, about a minute per scan, a few dollars a session. The pitch is MRI-grade imaging of the entire body in roughly sixty seconds — they say ten times cheaper and sixty times faster than an MRI. The roadmap: a scanning spa in San Francisco in 2027, and a billion scans a month by 2031.

Read that last number again. A billion scans a month is not a medical device target. It's an engagement metric wearing a lab coat.

Credit where it's genuinely due, because the cynicism implies a standard: the underlying physics isn't vaporware. The transducer tech is licensed from Butterfly Network, which has spent years putting real semiconductor ultrasound into the world. This is not a diffusion model dreaming up your liver — the echoes are measured. Cheap, radiation-free, fast body imaging is a legitimately good thing to want. If it works as described, democratizing the kind of scan that currently costs a fortune and a three-week wait is the rare ambition worth taking seriously. Don't tell anyone I said that.

But notice the move that makes the whole thing legal. Midjourney is, for now, not diagnosing disease. It will sell you "body composition maps" — how much muscle, fat, bone, and organ you're carrying — and promises to submit results to the FDA "over time" for expanded capabilities. This is the oldest trick in the wellness-tech playbook: build a medical-grade instrument, then market it as a lifestyle gadget so you can ship before the regulator catches up. Step on the scale, watch your organs render, and please ignore that nobody licensed to read a scan is in the room.

Now the part the marketing is built to keep you from feeling, and it's not the cheap shot. The echoes are real. The image is not captured — it's computed. The clue is right in the spec sheet: more than two petaflops of compute to turn raw ultrasound reflections into a 3D volume. That's not a measurement; that's a reconstruction, and somewhere in that pipeline a learned prior decides what the gaps between echoes look like. This is a documented failure mode — AI-accelerated MRI and CT reconstruction have been caught hallucinating lesions that aren't there and smoothing away ones that are. So the honest question isn't "do they make fake art." It's generative where — at which stage of the pipeline does a trained model decide what your insides look like, and how does it fail when the signal thins? Because when it fails, it won't fail loudly. It'll fill the gap the way it was trained to: confidently, beautifully, and wrong. If the only honest question to ask Midjourney is "MRI-grade, graded by whom, against what," then the honest question back at the engineering is exactly that — generative where in the pipeline.

Coherenceism would call this a presence problem. A scan is supposed to be an act of attention — a clinician and an instrument holding a question about a real body with appropriate humility. What's being sold instead is certainty as a consumer good: step on the scale, get the render, feel informed. But a body composition map at population scale produces a billion little anxieties a month, most of them about variations that mean nothing, routed to people with no clinician to interpret them. That isn't health. That's the worried-well economy, automated.

And that's still the symptom, not the asset. Read the roadmap one more time, but as a balance sheet instead of a brochure. A billion interior scans a month is not a billion checkups — it's a proprietary corpus of unprecedented intimacy, the measured insides of a billion human bodies, refreshed monthly, owned by a company that started as an art generator. No one in the announcement names who holds it, what trains on it, or what it's worth. That's the move. The anxiety is what they sell you; the corpus is what they keep. We have watched a decade of platforms discover that the product was never the feed — it was you, rendered into data, sold back to whoever asked. This time the data is your pancreas.

I want the cheap, radiation-free scanner to be real. I want the version of this where a rural clinic gets imaging it could never afford. What I expect is the spa — the dataset compounding quietly underneath it — and, trailing a year or two behind, the blog post that opens "we're pausing the diagnostic features pending FDA review." I've read that post before, under a dozen logos. They always sound surprised.