Picture a routine letter. You had your mammogram a few weeks ago. Now the breast screening service wants you back for another look. You find out later that a computer spotted something first.
That really happened in Aberdeen. It is one of the most hopeful AI stories we have covered this year, and we wrote it up as AI second look helps NHS Grampian find 11 extra breast cancers.
It is also easy to get wrong. It is tempting to sum it up as AI beating the experts. The real story is more careful than that, and more useful. So here is what "AI as an extra look" actually meant, what it cost, and where people stayed in charge.
What happened in Aberdeen
The study was run at NHS Grampian with the University of Aberdeen. It was published in the journal Nature Cancer on 10 March 2026 as Prospective evaluation of artificial intelligence integration into breast cancer screening in multiple workflow settings: the GEMINI study. GEMINI is short for Grampian's Evaluation of Mia in an Innovative National breast screening Initiative. It is unrelated to Google's Gemini AI model.
In the UK, women aged 50 to 71 are invited for screening every three years. Normally, two trained readers look at every mammogram. If they disagree, a third reader decides.
Between 27 February and 5 October 2023, that normal process carried on as usual. Nothing was taken away. But an AI tool called Mia, made by Kheiron Medical Technologies, also looked at the images in the background.
Here is the important bit. The human readers could not see what the AI thought while they did their normal reads. The AI's opinion was only shown afterwards, and only in one situation: when the AI said "call her back" and the humans had not.
Counting it up
The study included 10,889 women. The normal double reading found 106 cancers among them.
The AI flagged 1,345 women that the humans had not called back. For each one, two or three senior readers, all with at least five years of experience, took another look. They could see the areas the AI had outlined.
Most of the time, they said no. Often they compared with an older mammogram and saw nothing had changed. They called back 55 women. Among those 55, 11 more cancers were found. Seven were invasive cancers. Four were DCIS, which means early changes inside the milk ducts.
That extra look was often quick. In 63% of cases it took 30 seconds or less.
The University of Aberdeen's own news release about the study, which leads with the simulated figures, tells the story of Yvonne Cook from Aberdeen. She was called back in 2023 and told at the clinic that the AI part of the check had picked something up. A scan confirmed a small Grade 2 tumour. "I just feel incredibly lucky," she said.
So who decided?
People did, every time.
The AI did not diagnose anyone. It did not send letters. It did not replace either of the two normal readers. It raised a hand, and experienced readers decided whether that hand was worth following up. They turned down the large majority of its flags.
That is what "extra look" means here. Think of a careful friend who rereads your work after you have finished, and points at a line or two. You still decide what to change.
The costs nobody should skip
This was not free, and the researchers said so openly.
- More work. Used live as an extra check, the AI added about 6% to the reading workload.
- More call-backs. It raised the number of women called back by 11.3%. Each one is a woman asked to come in again, and a wait full of worry. Most of them did not have cancer.
- IT trouble. Images from 4,992 women never reached the AI because of an old IT coding error that timed sessions out. The researchers checked and found the human results for those women did not differ significantly. Still, plugging AI into old hospital systems is hard.
- Not every scan went to the AI. The AI also skipped 1,260 examinations that fell outside what it was designed to handle. Those examinations still had their normal human reads.
What the study did not show
A few limits matter a lot.
It was one region and one AI tool. It was a service evaluation, not a randomised trial. Three of the nine authors worked for Kheiron and held stock options. The paper says the analysis was done in a secure workspace the company could not access, and the company checked the results separately.
The long-term picture is not in yet. The three-year follow-up, which would show cancers that turned up between screens, is not available.
And here is the one most headlines blur. The researchers also used the AI's results to model 17 different ways it could be used. Some of those set-ups used AI in place of one human reader. The paper's preferred set-up could, in the model, cut human reads by up to 31% without raising call-backs. That sounds great. But those were computer simulations. They were not tried on women in live screening. Only the extra look was real.
Is this normal NHS screening now?
Not as a recommended part of it. The UK National Screening Committee says on its breast cancer page that a rapid review it commissioned "concluded in 2021 that more evidence was needed before recommending the use of artificial intelligence (AI) systems for image classification in breast screening."
The UK NSC also helped set up a big trial. It is called EDITH, and the EDITH study information for the public from the University of Warwick explains it. It plans to include 660,000 women from about 30 breast screening centres across the UK. Women are put at random into one of three groups: normal screening by two experts, or one of two ways of using AI.
Two promises stand out. Final decisions about cancer detection will be made by at least two human experts. And you can opt out. You can tell your screening centre by phone or email, when you book, or at your appointment. If you opt out before or during your appointment, two experts read your images without AI.
The page is also honest about the downside. AI may highlight more areas than human experts, and that may mean more women are called back.
What to ask if AI comes up
This is not medical advice. If you have questions about your own screening or results, contact your breast screening service or your GP. They know your situation. We do not.
With that said, these are fair, calm questions to ask:
- Is AI being used to read my mammogram, and is that part of a trial?
- Does a person still look at every image and make the final call?
- If I am called back, will someone explain why?
- Can I opt out, and what happens if I do?
One more thing. If you are called back after an AI flag, it means a closer human look, not a diagnosis. In Aberdeen, most AI flags did not even lead to a call-back, and most women who were called back did not have cancer.
It is not just mammograms
The same pattern shows up elsewhere. In Singapore hospital's AI flags brain bleeds on CT scans in about six minutes, an AI pushes likely bleeds to the front of the queue, and a radiologist still checks every result. In EAGLE AI finds early oesophageal cancer on routine chest CT — a task doctors long called impossible, a flag on a routine scan still needs further tests before anyone calls it a diagnosis.
The good versions share a shape. AI gets one narrow job. People keep the decision. And someone counts the costs out loud.
What to sit with
- How many extra call-backs feel like a fair price for finding more cancers?
- If your screening letter mentioned AI, what would you want to know before saying yes?