All Data Is Good Data
How the Midjourney ultrasound project started a war between medicine and tech.
On Friday, Midjourney, a generative AI company best known for automatically creating DeviantArt images from text prompts, announced it was pivoting to medicine. Their web copy opens with blunt honesty: “Today we’re gonna announce something a little weird and a little crazy, but also spectacular and filled with hope.” What follows is a cinematic trailer straight out of Ridley Scott’s Alien franchise, depicting Midjourney’s dream of a new whole-body ultrasound scanner that is “as powerful as MRI, and as casual as a trip to the spa.”
Their first scanning spa won’t be available until 2027, so I’m not sure why they decided to announce this on Friday. But it ignited a heated fight between tech boosters and medical practitioners, lighting up my mostly dormant Twitter timeline all weekend. I am embarrassed to confess that it was custom-made to engage me and my niche tastes. I couldn’t look away!
Let’s first be clear: everyone was fighting about vaporware. The actual technology is based on a 2026 Nature Biomedical Engineering paper. That paper, by a Caltech lab, is a nice engineering systems integration project combining a lot of technology that’s been around for decades. There is no fancy AI, and the imaging algorithm is written in Matlab and run on a cheap GPU (an RTX 3070). The paper doesn’t demonstrate any particularly cutting-edge imaging, presenting a few reasonable images of abdominal fat. Instead, the paper promises acceleration, claiming the ability to perform a whole-body scan in about a minute.
Midjourney thus promised something not yet demonstrated:
“In an ideal and near-term future, we take this information and watch how it changes over time. We compare it to the general population, we talk to doctors, nutritionists, coaches, trainers, and AI friends. We become more aware of our health and we improve our lifestyles. We make smarter, more proactive, more frequent decisions. And we live longer, healthier lives, better lives.”
Now, this text may seem clichéd and innocuous, but two communities on Twitter read this paragraph very differently.
On the one side were tech boosters, who seethed with vitriol against the medical establishment. They believe that all information is good, and more information can only improve decisions. They embrace some perverse version of Sutton’s Bitter Lesson that more data fed into more computing can solve all problems.
On the other side were medical doctors, most of whom also seemed to really hate the medical establishment! They’re on Twitter, after all. But they had compelling arguments against indiscriminate scanning.
First, they note that ultrasound is a limited imaging modality, and the new full-body system doesn’t eliminate those limitations. Ultrasound can’t see through bone or air, and has other fundamental limitations with contrast. You can’t fix a fundamentally physics-limited scanner with more frequent scans.1
Second, though conventional wisdom says that early detection improves disease outcomes, we have 75 years of evidence that is far from encouraging. The case has been most illuminating in oncology. Some cancers are untreatable no matter when they are detected. With advances in therapies, some cancers are treatable even if you catch them late. The correlation between early detection and cure is, unfortunately, much lower than cancer researchers had hoped in the 1950s. The medical establishment has tempered its expectations over the last 20 years, but the lay population hasn’t caught up.
Third, frequent scanning means you have to run more tests to follow up on what you see in a scan. This is a nightmare. Imaging is never perfect. Let’s say that you build a technology that is 95% accurate at detecting a disease that afflicts one in a thousand people. A crude probability calculation2 then says that for every hundred scans that come back positive, only 2 of them actually have the disease. What about those other 98 people? They are healthy people who will be subjected to the stress of potentially harmful follow-up tests. Biopsies aren’t fun!
The fourth lesson also comes from oncology. We don’t understand the dynamics of many progressive cancers. Even the best imaging technology still can’t tell the difference between harmful and harmless masses. Some growths look bad and appear to grow over time, but they never cause harm. On imaging, these look the same as tumors that will kill you in six months. Why should we be constantly raising alarms about information that doesn’t lead to actionable, beneficial interventions?
These arguments fell on deaf ears. The tech boosters were furious that the medical conservatives were against more information. They yelled that doctors are terrible at mathematics. They yelled that doctors are gatekeepers preventing people from accessing information that will make them better. They yelled that doctors don’t know innovation if it stares them in the face. One particularly vitriolic person wrote: “You guys thought that Lister and Pasteur both needed to be crushed, and for the most part, you haven’t changed.”
Unsurprisingly to me, the tech boosters didn’t present any scientific arguments. I only saw unfounded assertions about Bayes’ Rule and Blackwell’s theorem, and how “false positives aren’t a problem.” The Midjourney scanner is yet another project in the optimaxxing biohacking space, seeking technological means to wrest control of your body.
What the tech side of the argument came down to was this seemingly anodyne statement in the Midjourney blogpost: “You want as much data as you can get about your health as quickly and as cheaply as possible,” and that you should strive to maximize the “megabytes per second per dollar of information” you capture about yourself. Massive time series data would unlock new diagnostics, overcoming the limitations of single scans. All data is good data, and all data leads to improvement. This assertion is not backed by evidence. It is, as Midjourney wrote, hope.
I side with doctors on this one, though for a reason that they didn’t seem willing to advance. We shouldn’t obsessively overmedicalize ourselves. There is an inherent harm in always seeing yourself as a patient, ailing unless proven healthy. More information can make you sicker.
To people who doubt this: Do you think AI will be able to turn iPhone photos into MRI images? If so, hit me up and we can go talk to some VCs.
This calculation is called “positive predictive value” and is an application of Bayes’ rule.
Kind of feels like the same crowd that convinced itself that token-maxing was the right way to use AI is now trying to convince everyone that scan-maxing is healthcare.
The hope is always that data has positive network effects rather than diminishing marginal returns and the log relationship between training data and capability for LLMs sure seems like it should've settled this for even the most enthusiastic boosters of this line of thought, but noooo