A Wall of AI at ASCO - but is it working?

TL;DR

At ASCO this week, AI dominated the oncology agenda, with Tempus AI alone bringing 37 abstracts. But one Penn abstract pointed in a different direction — at the patient — and found that the AI information cancer patients actually encounter online is low quality, hard to read, and silent on risks. Oncologists are now reporting patients walking in having already been briefed by an AI tool on their phone, with no validation of what it said. Most AI-in-oncology strategies have a blind spot here: enormous capital goes into the models, far less into the layer that explains those models to the person on the receiving end. The model can be clinically right and still fail at the bedside.

American Society of Clinical Oncology (ASCO) this week was a wall of AI aimed at the oncologist. Tempus AI brought 37 abstracts. Mayo Clinic brought AI reading the tumour microenvironment. One Penn abstract pointed the other way, at the patient, and found the AI information patients actually see is low quality, hard to read, and silent on the risks.

The team screened the first 320 webpages and videos a cancer patient finds when they search Google or YouTube for AI and their disease. That is the material shaping how patients understand a diagnosis, and it is failing them.

The finding that should land with anyone building in this space came from the clinic, not the search results. Oncologists now report patients walking in to ask about something an AI tool already told them. The technology is in the exam room. It arrived through the patient's phone, not through the health system, and nobody validated what it said on the way in.

This is where most AI-in-oncology strategies have a blind spot. Enormous capital is going into models that touch the patient, diagnostics, decision support, multimodal prediction. Far less attention is going into the layer that explains those models to the person on the receiving end. That layer gets treated as communications, a thing you bolt on after the science works.

The cost shows up directly in uptake. A patient who does not understand why an AI-guided recommendation differs from what they read online is a patient who hesitates, second-guesses, or opts out.

The model can be clinically right and still fail at the bedside.

Key takeaways

• ASCO this week was dominated by AI aimed at the oncologist — Tempus AI brought 37 abstracts, Mayo Clinic brought AI reading the tumour microenvironment.
• One Penn abstract pointed the other way: the AI information cancer patients actually find online is low quality, hard to read, and silent on risks.
• The first 320 webpages and videos a cancer patient finds when searching Google or YouTube for AI and their disease are the material shaping how patients understand a diagnosis — and it is failing them.
• Oncologists now report patients walking in to ask about something an AI tool already told them; the technology arrived through the patient's phone, not through the health system, and nobody validated what it said on the way in.
• Most AI-in-oncology strategies have a blind spot: enormous capital goes into models that touch the patient, far less into the layer that explains those models to the person on the receiving end.
• That explanatory layer gets treated as communications — a thing you bolt on after the science works.
• A patient who does not understand why an AI-guided recommendation differs from what they read online is a patient who hesitates, second-guesses, or opts out.