1Part 1 of 6

What AI actually does during an endoscopy

It helps to bring the concept down to earth, because "artificial intelligence" sounds like far more than it is. In the endoscopy room today there are two families of systems:

• CADe (detection). It analyses the video in real time and draws a box around anything that might be a polyp. This is the one in widespread use. It decides nothing: it draws your attention.
• CADx (characterisation). It goes a step further and tries to predict whether a lesion is neoplastic by looking at its surface and vessels, without needing to remove it and send it to the lab.¹

There is also a third, less glamorous but very practical line: systems that monitor the quality of the examination, for example assessing whether the colon is properly clean or whether the whole length has been inspected. ²

Key idea: what sits in endoscopy rooms today is not a machine that diagnoses, but an attention detector. It points at where to look. The person who looks, judges and decides is still the doctor.

2Part 2 of 6

What the evidence says: yes, it detects more

Here the answer is clear and positive. The most complete meta-analysis pooled 28 randomised trials with 23,861 patients and found that AI-assisted colonoscopy:

• Increases the adenoma detection rate by 20% (RR 1.20; 95% CI 1.14-1.27).
• Reduces the adenoma miss rate by 55% (RR 0.45; 95% CI 0.37-0.54).
• Lengthens withdrawal by only 0.15 minutes — around 9 seconds.
• Works just as well in expert hands (RR 1.19), which was not obvious beforehand.
• ³

It holds up in routine practice too. The British COLO-DETECT trial, run across 12 NHS hospitals with 2,032 patients, found more adenomas per procedure (1.56 versus 1.21) and a detection rate of 56.6% versus 48.4%, with no increase in complications. ⁴ For transparency: that trial was funded by Medtronic, the manufacturer of the system being evaluated — worth bearing in mind when reading its results.

One methodological detail is worth knowing: results depend quite a lot on how the study is designed ⁵ and the benefits seen in controlled trials tend to shrink in everyday clinical practice, partly because in a trial the endoscopist knows they are being watched and performs better. ⁶

3Part 3 of 6

The uncomfortable limit: more adenomas is not the same as less cancer

This is the part that rarely makes the headlines, and it is the most important. When you look at what is being found extra, a decisive nuance appears: the increase is driven mainly by diminutive lesions, and there was no significant difference in the detection of advanced adenomas — precisely the ones capable of becoming cancer. Sessile serrated lesion detection did not improve either. ³

And there is a cost: AI use was associated with 39% more resections of lesions that turned out not to be neoplastic. In other words, more things get removed that did not need removing, with the small risk and expense that entails. ³

One result illustrates the limits nicely: the TIMELY trial, international and led from Spain, tested AI in people with Lynch syndrome (high hereditary risk of colon cancer) and found no improvement at all: 0.64 adenomas per colonoscopy in both groups. Its conclusion was that in this population the cornerstone remains a meticulous technique. ⁷ If you want the background, here I explain what risk a polyp really has of becoming cancer.

On the optimistic side, a modelling study estimated that rolling AI out across population screening could prevent, each year in the United States, around 7,194 additional colorectal cancer cases and 2,089 deaths, while saving the system money. That is a simulation, not an observation: useful for guiding decisions, not for taking as fact. ⁸

4Part 4 of 6

The 2027 question: is it blunting our skills?

This is the big open debate, and it began with a study published in 2025 that shook the endoscopy community. At four centres in Poland, researchers compared how endoscopists performed colonoscopies without AI, before and after they had grown used to working with it. The result made everyone uncomfortable: adenoma detection in non-AI colonoscopies fell from 28.4% to 22.4%, an absolute difference of 6 percentage points. ⁹ This is the phenomenon that has been christened deskilling: you lean on the assistant so much that you lose your edge without it.

That said, it would be dishonest to give you only half the story. It is an observational, retrospective study, not a trial, and other work does not confirm it. A prospective 2026 study with 13 endoscopists and 5,013 colonoscopies measured exactly this — a phase without AI, a phase with it, and a phase after removing it — and found no deskilling at all: once AI was withdrawn, performance returned to baseline. What it did find was that AI temporarily helps less experienced endoscopists while they use it. ¹⁰ The debate is very much alive in the specialty journals. ¹¹

Key idea: the real risk is not that the machine replaces the doctor, but that the doctor stops looking as hard because they trust the machine is already looking.

5Part 5 of 6

Beyond the colon

Almost all the solid evidence is in colonoscopy, but three other fronts are advancing:

• Leaving diminutive polyps in place. The Italian PRACTICE trial compared removing every polyp against leaving in situ the diminutive rectosigmoid ones that AI classified as non-neoplastic. The strategy proved non-inferior (adenoma detection 44.7% versus 46.5%) with no complications. It could spare unnecessary polypectomies.¹³
• Barrett's oesophagus. There are systems for detecting early dysplasia, an area where the human eye often fails; the current challenge is making them work outside the setting where they were trained.¹⁴ I wrote about this here: Barrett's oesophagus and cancer risk.
• Stomach. It helps detect and characterise premalignant lesions and early gastric cancer, where selecting the right patient determines whether endoscopic resection will be curative.¹⁵
• Capsule endoscopy. Perhaps the neatest fit: a capsule generates more than 50,000 images per study and reviewing them is exhausting. Tools already in clinical use cut reading time substantially by selecting the relevant images, though false negatives need watching.¹⁶

6Part 6 of 6

What this means for you (and what to ask)

If you are having a colonoscopy, AI is good news: it adds rather than subtracts, and it barely lengthens the procedure. But it is not the most important question you can ask. What most determines the quality of your colonoscopy is still human and organisational.

Myths and facts

The five I hear most when this subject comes up in clinic.

My approach in clinic

Related reading

Quality colonoscopy

The indicators that actually predict a good examination.

Read post →

Will a polyp become cancer?

Why not all polyps carry the same weight.

Read post →

Your report, translated

What each term in your endoscopy report means.

Read post →

Frequently asked questions