An AI image company built a medical scanner, and the market moved before the evidence did
On June 17, 2026, Midjourney — a company known for generating images from text — announced a full-body medical scanner it calls “Ultrasonic CT.” Within days, the stock of its hardware partner, Butterfly Network, rose 33%. No part of the device had FDA clearance. And founder David Holz said plainly, in Bloomberg’s reporting, “We’re not even using any AI in this yet.”
Around 12 people had been scanned at the time of the announcement. The stated goal is 50,000 scanners worldwide by 2031, with capacity for 1 billion scans a month. That distance — between what has been demonstrated and what has been claimed — is the part worth studying.
What actually happened
Midjourney launched a new division, Midjourney Medical, and unveiled a whole-body ultrasound device that produces a 3D body map in about 60 seconds, using sound waves and water instead of radiation or magnets. The imaging is not Midjourney’s own work. It is built on Butterfly Network’s ultrasound-on-chip technology under a licensing agreement worth up to $74 million over 5 years, signed in November 2025, with the prototype using 40 Butterfly imaging modules per system, according to Butterfly’s filing.
The signal for anyone buying AI in a regulated sector is this. A consumer AI brand extended into an evidence-governed domain, framed its product against an established gold standard (MRI), gave it a name that implies more than the device does, and deferred the regulatory pathway. The market reacted to the brand and the demo, not to validated results.
Why this is a problem, not just a product launch
Heads of Compliance, DPOs, COOs, and CFOs in regulated firms are pitched AI systems almost daily. The Midjourney case compresses a familiar pattern into one announcement, which makes it a clean teaching example.
The name “Ultrasonic CT” suggests computed tomography. It is not. There is no X-ray and no ionising radiation — it is ultrasound, as the company’s own materials confirm and as independent explainers point out. The performance framing — image quality “comparable to” or “in ways superior to” MRI — came from a launch event, not from peer-reviewed testing. Radiologists pushed back quickly. One associate professor at NYU Langone Health noted the device does not currently outperform modern ultrasound, CT, or MRI on the evidence shown, even while calling the direction promising, per Radiology Business.
None of this makes the project illegitimate. It makes the gap visible. Capability claims, brand strength, and market enthusiasm are not the same as validated performance. In a domain where being wrong carries clinical, financial, or legal consequences, the buyer’s job is to separate what is demonstrated from what is asserted.
Here is the same launch read two ways:
| Item | Midjourney’s framing | Independent status, June 2026 |
|---|---|---|
| Image quality | Comparable to, or in ways superior to, MRI | Not independently validated; specialists say it does not yet outperform modern ultrasound, CT, or MRI |
| Scan speed | About 60 seconds, roughly 100x faster than MRI | Demonstrated on a prototype; around 12 people scanned at announcement |
| Regulatory status | FDA approvals to be added incrementally | No FDA clearance; launching as body-composition “wellness,” not diagnosis |
| Role of AI | AI reconstructs a 3D body map | Founder: “We’re not even using any AI in this yet”; imaging uses Butterfly’s ultrasound-on-chip |
| Scale | 50,000 scanners, 1 billion scans a month by 2031 | Stated goal; first location planned for end of 2027 |
How to evaluate an AI system before it touches a regulated workflow
The discipline below applies whether the system is a medical scanner or a customer-facing agent in finance, insurance, or pharma. Each point is a concrete check a buyer can run.
Separate demonstrated capability from claimed capability
Ask what has actually been tested, on how large a sample, and with what results. The honest version of the Midjourney number is “around 12 people scanned,” not “a billion scans a month.” For any AI system, require a defined benchmark set and the measured outcome against it before you accept a capability statement. A demo answers “can it do this once.” A benchmark answers “how often is it right.”
Map the regulatory perimeter first, not after
Midjourney begins with body-composition maps, which require no clearance, and defers diagnostic FDA approval. That is a legitimate sequencing choice, but a “wellness-first” framing can obscure where the regulated line sits. For DACH firms, the equivalent questions are concrete: which functions process personal data under revDSG, where does data reside, and which decisions fall under sector rules or the EU AI Act for your EU operations. Know which parts of the system are inside the regulated perimeter before, not after, you deploy.
Verify where the AI actually operates
The branding implied AI was doing the imaging. The founder said it was not. This is worth checking on every system you evaluate: where do models actually run, what do they decide, and what is deterministic or rule-based. A plain map of “model here, retrieval here, fixed logic here” tells you where errors can originate and where they cannot.
Require source-grounded answers, not model confidence
An AI system in a regulated workflow should retrieve from a verified knowledge base, cite the source of each answer, and decline to assert beyond it. This is the difference between a system that reasons from your approved data and one that produces fluent, unverifiable text. Lab51 builds custom AI agents from a client’s own data, with grounded retrieval that pulls answers from a controlled knowledge base and a blacklist of claims the agent must never make. The same architecture that stops an agent inventing a competitor advantage is what stops it inventing a fact in a compliance-sensitive answer.
Demand an accuracy report against a “must-get-right” set
Before scaling, agree on a set of 20 to 50 questions the system must answer correctly, with the answers approved by your team. Test against expected requests and against deliberately adversarial and random inputs. The deliverable is an accuracy report with a pass threshold you set, not a slide that says the system performs well. If a vendor cannot produce one, that absence is itself information.
Read the naming and the market signal critically
“Ultrasonic CT” describes ultrasound with a CT label. When terminology overstates function, treat it as a prompt to look closer. And separate market reaction from product validation: a 33% stock move is a signal about expectations, not a clearance event. Both can be true at once — a project can be genuinely interesting and still unproven.
Why this matters now
AI brand extensions into regulated domains are accelerating, and consumer-grade confidence is increasingly applied to decisions that carry real exposure. The volume of AI vendor pitches reaching Swiss and DACH firms is rising at the same time the regulatory environment is tightening, with revDSG obligations and EU AI Act duties phasing in for firms with EU operations. The cost of deploying an unvalidated system into a regulated process is no longer theoretical. The practical response is to make the evaluation discipline above a standing requirement, applied to every AI system before it reaches a regulated workflow, rather than a one-time check during procurement.
The question to keep asking
The Midjourney launch is a useful mirror because it shows how far a confident demo and a strong brand can travel before the evidence arrives. For any regulated buyer, the test is the same one the radiologists applied: where is the validation, who produced it, and does it hold up when read in isolation from the marketing. A system that can answer that cleanly is one you can defend internally. A system that cannot is a risk you are accepting on someone else’s word.
