By Joanna Stern
For 12 years, I wrote technology columns here at The Wall Street Journal . So you know I'll do almost anything for a story. I spent a year letting AI into as many parts of my life as I could. That experiment became my new book, I Am Not a Robot: My Year Using AI for Almost Everything . What follows is an edited chapter. Warning: It gets more intimate than my usual product review.
I KNOW it's a bit odd to do this right off the bat, but I sense something special between us. A connection. A bond. So I think it's time -- time to talk about my breasts.
To quote one of the best episodes of Seinfeld, "They're real and they're spectacular." Or in my case, it's more like they're real and they're... complicated. They're why I started getting mammograms early -- around when I was 30, compared to the usual 40 -- for reasons I'll explain soon. That complication is what makes them the perfect test case for AI "doctors" trained to read breast imagery.
When Annual Mammogram Day came around, I was four weeks into what I called my AI year, weaving artificial intelligence into every corner of my existence. Not just at work -- writing emails, doing research, testing AI vending machines. I'm talking 24-7 AI livin'. Robots helping around the house, on the roads, on the massage table, at the dinner table. If there was a decision to make or a task to do, I wanted to see what happened when I let AI go first. I tried to make AI my everything. Even when it came to my health decisions.
And it turns out healthcare is where that partnership -- human plus AI -- may be advancing fastest. Not replacing doctors but augmenting them. Catching what they miss. Maybe even saving lives. It's not just mammograms, either. The same AI pattern-recognition that's learning to spot breast tumors is being applied to thyroid screening, lung nodule detection and colonoscopies.
Two factors make my breasts particularly challenging for radiologists. They're structurally dense, meaning they contain more glandular and fibrous tissue than fat. Dense breast tissue appears white on a mammogram, the same color as tumors, making it more difficult to detect abnormalities.
The second complicating factor: My mom is a three-time breast cancer survivor, which puts my risk higher than the average woman's. Based on our family history -- including two first cousins who've been through it -- I have a 39% chance of developing the disease in my lifetime.
I watched my mom go through a lot with cancer: lumpectomy surgeries, a bilateral mastectomy, multiple rounds of chemotherapy and radiation. While undergoing treatment in 2001, she lost her hair and wore wigs and scarves to cover her head for nearly two years. There were times growing up when I wasn't sure she'd be here for my high school graduation, let alone my college graduation, my wedding or the birth of my two sons. This year, she turned 73. Many women are not so fortunate.
If AI could improve my own odds, I wanted to give it a try. And soon I'd learn something even more sobering: If this technology had been around three decades ago, it might have spared my mom a surgery and some treatments.
At Mount Sinai Hospital on Manhattan's Upper East Side, I was led to a small changing room and told to strip from the waist up and put on a stiff pink cotton gown that clearly hadn't met fabric softener since the Reagan administration.
Inside the exam room, I came face-to-face with a refrigerator-size mammogram machine. The technician carefully positioned my right breast between two transparent plastic plates. Carefully is a generous term here -- the process is more like arranging a raw chicken on a baking pan for maximum flatness. Once everything was in place, she moved aside and pressed a button and the top plate smooshed my breast like a marshmallow in a s'more made by my overenthusiastic 8-year-old.
"Hold your breath. Don't move," she said. A few seconds and loud beeps later, the plate retracted, and that part was over. One image down, three more to go.
For the men reading this, I don't know whether there's an exact equivalent, but let's try. Picture laying your penis on a cold plastic tray, then watching another tray descend, pressing down and flattening your equipment like a squashed gummy worm for the longest three seconds of your life.
We repeated the process on the left side before conducting a bilateral ultrasound -- a requirement for my aforementioned dense breasts. After 40 minutes of the technician repeatedly dragging a warm and gooey wand across my chest, I had 95 images -- and enough gel on my skin to moisturize a lizard for a year.
At this point, I usually sit in the waiting room biting my nails, bracing for the results. But this time, I got a rare peek behind the curtain.
Dr. Laurie Margolies, the chief of breast imaging at the Dubin Breast Center at Mount Sinai, has spent nearly four decades reading mammograms, starting back when radiologists squinted at film on light boxes. Now she's an AI enthusiast.
After my exams, we sat together in front of three glowing monitors and examined my four mammogram images using ScreenPoint's Transpara AI software, which has been trained using millions of images of malignant and benign masses to recognize subtle patterns that might escape even the most experienced human eye.
Margolies knew the system well. She's also on the company's medical advisory board, a relationship worth noting -- though she was quick to say her opinions came from decades of reading mammograms, not a marketing script.
A large "L" appeared next to the images.
"AI does not think there's anything on your mammogram," she said. The "L" stood for "low" chance of cancer, which Margolies said was less than 1 in 2,500. She explained that if the AI tool had found something suspicious, there would have been an "E" for "elevated," with a triangle or circle highlighting the problematic spot.
Margolies zoomed and panned across my breast tissue like a detective examining crime scene photos.
"Nothing I see here is of concern," she said, agreeing with the AI's assessment.
But the extreme density of my breasts made mammograms just the starting point. "Your cancer -- if you had one -- unfortunately could have been hidden by the dense tissue. Even the AI might not have picked it up," Margolies cautioned.
Later, when I spoke with ScreenPoint CEO Pieter Kroese, he claimed that the company's studies show Transpara AI to be about 20% more accurate in reading highly dense breasts than radiologists alone are. "That tells you AI sees things that a human eye can barely see or cannot see at all, " he said.
ScreenPoint's studies on dense breast findings were conducted internally, but an independent study, led by UCLA and published in 2025 in the Journal of the National Cancer Institute, found that Transpara could flag subtle signs of breast cancers that develop between routine screenings. These cases, in theory, could respond to early treatment and reduce certain cancer risk by up to 30%.
Margolies turned to my ultrasound images, pulling them up in the Koios DS Breast tool. One showed a "B," for "benign," when the doctor's cursor isolated it with her mouse. But another, on the right breast closer to the nipple, returned an orange "S," for "suspicious."
Margolies wasn't worried. This mass, she said, was there eight months ago and hadn't changed shape or size. Stable lumps are good lumps.
She seemed to trust the AI every time it read "benign." But when the AI indicated something was suspicious, she was more skeptical. She explained that when the AI puts up the big orange "S," it is right only 30% to 40% of the time. That's by design.
"If you are the FDA, that's the way you'd want the program to be, more sensitive than specific. You wouldn't want to miss breast cancer. You wouldn't want a radiologist to not biopsy something even though the computer said it was benign, if it looked suspicious."
Ultimately, the Koios AI marked three of the small masses in my right breast with the big orange suspicious rating. Only one, located at 9 o'clock and close to the nipple, slightly concerned Margolies, even though it hadn't changed in size or shape.
"We should double-check this one in six months on an ultrasound because of your extreme breast density and your family history," she said.
The human doctor with 40 years of experience was right -- the small oval in my right breast didn't look "suspicious." She was also right to order follow-up testing, given the AI's suggestions and my history. Thankfully, that mass turned out to be nothing.
Sometimes, though, humans aren't right, and they miss things. That was certainly the case back in 1993 with my mom. When she first discovered a lump in her left breast, she had a lumpectomy to remove it, then went through radiation. Six months later, following a routine mammography of her right breast, she was diagnosed with DCIS -- ductal carcinoma in situ, a precancerous condition that could develop into invasive breast cancer -- which led to her decision to get a double mastectomy.
"I soon learned that those DCIS calcifications had already appeared on the mammogram taken six months earlier," my mom told me as we sat together reading over this chapter. "If the radiologist had carefully reviewed both breast mammograms at the first occurrence, I would have skipped the lumpectomy and radiation and gone straight for bilateral mastectomies."
Instead, the radiologist missed it as a result of being so focused on the lump in the left breast. "It seems to me this new AI technology would have paid equal attention to both images," my mom said.
In 2016, Geoffrey Hinton, often called one of the godfathers of AI, made a bold prediction: "People should stop training radiologists now. It's just completely obvious within five years deep learning is going to do better than radiologists."
By 2023, he had walked that back and adjusted the timeline by another 10 to 15 years. "I think I was off by about a factor of three, but I'm still convinced I was completely right in the long term," he said on a podcast.
(MORE TO FOLLOW) Dow Jones Newswires
May 04, 2026 05:30 ET (09:30 GMT)
Copyright (c) 2026 Dow Jones & Company, Inc.
Comments