AI selfie-based cancer survival prediction is promising but not yet shown to be reliable or clinically useful beyond doctors' judgment
Source: https://www.theguardian.com/profile/devi-sridhar. "Can AI really predict cancer survival better than a doctor? I have my doubts | Devi Sridhar | The Guardian." October 23, 2026. www.theguardian.com
The Gist
A new AI tool can estimate how old your face looks and may help predict cancer survival, but the evidence so far is patchy. It works for only some cancers, may be less accurate for Black patients, and can be thrown off by lighting, cosmetic treatments and other factors. The author says we still don't know whether it tells doctors anything they couldn't already see.
Conclusion
Although the FAHR-Face research is promising, there is insufficient evidence that an AI model analysing selfies can reliably predict individual cancer survival or add value in clinical practice beyond what experienced clinicians already do, so skepticism is warranted.
Premises
- The study found AI-estimated facial age significantly associated with survival in only 18 of 28 cancer categories, so it works for some cancers but not others.
- The facial-age measure was not significantly associated with survival for Black patients after adjustment, and AI models have shown racial bias elsewhere, raising concerns about generalisability and fairness.
- The models are built on retrospective statistical associations across many patients, and it is a large leap from such associations to accurately predicting an individual's future survival or proving clinical usefulness.
- Selfies are unreliable inputs: appearance varies with lighting, angle, expression and time of day, and is altered by botox, fillers, facelifts and GLP-1 drugs, so visible age and underlying health may diverge.
- The tool has not yet been tested against doctors' own judgments or across different hospitals and patient groups, so its added value over clinicians' experience-based assessment is unknown.
- The main drivers of facial ageing identified (alcohol, drugs, smoking) are unsurprising, suggesting the tool largely recovers what doctors can already read from a patient's face.
Assumptions
- Association-level results in retrospective data do not by themselves establish reliable individual-level prediction.
- Cosmetic and pharmacological alterations to the face introduce error that the researchers' robustness tests cannot fully remove.
- A clinical tool must demonstrate incremental value over existing clinician judgment to justify adoption.
- Failure to find significance in some subgroups or cancer types indicates a real limitation rather than just low statistical power.
- Doctors' informal facial assessments (e.g. 'appears older than stated age') are a meaningful baseline comparator.