During the pandemic, the only radiologist at the Robert Mangaliso Sobukwe Hospital in Kimberley fell ill and was unable to process chest X-rays. This would normally have been a crisis. But a year earlier the hospital had installed RADIFY, an AI platform that flags critical findings on chest X-rays in real time. Using this AI tech the hospital was able to continue processing patient X-rays.
This is local tech from Envisionit Deep AI, founded in 2019 by Jaishree Naidoo, Terence Naidu, and Andrei Migatchev. The deal here is not just saving lives during the pandemic. Africa has one radiologist per 600,000 people while South Africa has only five imaging units per million people. RADIFY is able to label and prioritise over 2,000 X-rays per minute detecting more than 25 pathologies such as pneumonia, TB, COVID pneumonia and breast cancer.
For breast cancer we also have the Mirai Risk Model, currently deployed at Mount Auburn Hospital in Massachusetts, USA. Here the AI analyses patient history and imaging data and then predicts individual breast cancer risk. It then tailors mammogram screening frequency based on the AI’s assessment. This reduces unnecessary screening for low-risk patients while catching more cancers in high-risk patients.
At a simpler level we have the USA Department of Veterans Affairs, which launched its ambient AI scribe in October last year. The system lets a patient return home earlier than usual, then tracks vital signs and alerts healthcare workers if conditions worsen.
Your Apple Watch can do this to a degree, and it is important because it shifts healthcare to proactive, catching it before it becomes worse, rather than reactive, responding when things are already bad. This saves time, money and lives.
This is all good news from AI and healthcare.
The risks
Then mental health. This is a hot topic. Using a chatbot for mental health may be the only option for many. But it is dangerous and we have seen people using a chatbot for mental health commit suicide, at no point did the chatbot try and stop them.
There are however tools such as Woebot that are specifically designed for mental health. The current crop of commercial AI chatbots need guardrails in place for when they’re being used for mental health, simple things such as providing numbers for help lines and certainly not encouraging self-harm.
The time problem
Another big win for AI in healthcare is around saving time for doctors. In reality doctors, nurses and other healthcare professionals are in limited supply due to limits at educational institutions. The earth’s population is growing, but healthcare worker numbers are not growing as fast, so scarcity is a stark reality.
Here the AI runs during the patient/healthcare worker consultation, recording the conversation. After the appointment the AI is then able to summarise the details, saving the healthcare worker time, and in some examples it can even suggest treatments. We’re talking about maybe an hour saved per healthcare worker per day. Hours a week that can be redirected to seeing more patients.
This then brings up the issue of data privacy. It’s one thing having to talk to a doctor or nurse about our embarrassing health issues, but now an AI is listening. If it’s just an AI, then all’s well and fine. But how are these AIs trained? Where is the data stored and how secure is it? My bank just got hacked and all sorts of my private data is now out on the dark web.
Built for the wrong patients
Another real concern is data bias. Research shows dermatology skin tone bias is still happening in 2026. Darker skin has less training data and hence AI gets worse results when used on darker skin.
We saw a simple example of this during the pandemic when we learnt that the pulse oximeter that clips onto your fingertip and uses light to measure blood oxygen saturation can’t read darker skin accurately. A 2020 paper in the New England Journal of Medicine found that pulse oximeters were about three times more likely to miss low oxygen levels in Black patients compared to white patients.
AI is trying to solve this problem by building datasets with diverse dermatology images. But research in 2021 showed that only 5% of healthcare AI training sets used datasets that included diverse dermatology images.
Further, a pilot study testing AI pneumonia detection in Nigeria found that a model trained on USA paediatric chest X-rays correctly classified only 23 out of 97 pneumonia cases. A much lower detection rate than when used in the USA where the dataset originated.
There is also plenty of hype that is simply not substantiated as AI firms chase investment capital and headlines. One example is from the 2025 Medical Image Computing and Computer-Assisted Intervention conference, the premier annual academic event in the field of medical imaging AI. More than 80% of papers claimed to outperform prior methods, yet only around 10% used any statistical significance testing at all. One team of researchers found that 58% of classification papers had a greater than 30% probability their superiority claim was false.
Then there are just the classic tech failures. IBM Watson for Oncology spent $5 billion billing itself as a revolutionary “superdoctor” for cancer treatment recommendations. In 2022 IBM sold its entire health division for a mere $1 billion.
Lots of hype, but are we healthier?
The drugs are coming
The hype is real, but that doesn’t mean we aren’t seeing tangible results. There were over 170 AI-discovered drug programmes in clinical development in the USA at the start of this year. More than half are in Phase 1, with 15 in Phase 3 and another 15 to 20 expected to reach Phase 3 by year end.
It’s noteworthy that as of last year no AI-discovered drug has achieved FDA approval. But AI can identify patterns humans can’t see and it can work far faster than traditional research. Rentosertib from Insilico Medicine is the first AI-designed drug to reach Phase 2a testing. The disease being targeted is idiopathic pulmonary fibrosis, a rare and currently untreatable condition. In trials, patients given the highest dose saw a mean improvement of 98.4ml in lung function over twelve weeks while the placebo group declined by 62.3ml.
That’s a difference of 161ml in twelve weeks. Traditional drug discovery takes 10 to 15 years; AI is compressing that timeline significantly.
Who’s liable?
There are plenty of examples like the above and we’re still in the early days of AI and healthcare. Leap forward a decade or two and imagine what we’ll have by then.
A lot of the progress over these future decades will require wading through the hype, solving for biases and answering who is liable if things go wrong.
The AI companies are firmly stating they are not liable if their AI messes up. So it is left to the healthcare worker or, as is often the case, the patient. This is not good enough, especially when AI companies are firmly in this for profit and are hyping their achievements to boost valuations.
Yes, AI drugs and procedures will go through the required trials and yes, sometimes the trials get it wrong.
But no liability for the AI companies cannot work. They can’t hide in their black boxes pretending it has nothing to do with them.
One option is a pooled no-fault global fund. AI healthcare companies can contribute from their revenue and a harmed patient can claim without having to endure the expensive legal process. With the AI healthcare industry expected to top $18 billion this year, growing at 25% a year over the next decade, there’s enough money in the system to make this work.
AI is being used in healthcare and healthcare professionals are reporting good results. But it’s important we set the ground rules early, rather than reacting to problems long after the horse has bolted.
AI in the Wild
AI in the Wild is a regular column from Simon Brown.
AI is everywhere and only getting better. Record capital raises and valuations, and competing LLMs are all fun, but meaningless to our every day lives. This column will focus on how is it impacting us in the real world.
5th | 15th | 25th every month
