ideasIdea
A regulatory sandbox for continuously learning cancer AI
Let AI tools that improve as they learn be used under close supervision in a few hospitals, with pre-agreed rules for what changes are allowed and how they are checked.
Regulation assumes a frozen model; models that update on new data are effectively unapprovable, so deployed models age. The FDA's predetermined change control plan is a step. The proposal is a formal sandbox: a small number of sites, a pre-specified envelope of allowed updates, mandatory shadow evaluation of each update on sequestered data before activation, full audit logs, and a regulator seat at the table, generating the evidence needed to write general rules for adaptive AI.
Hypothesis
Adaptive models in the sandbox will maintain or improve performance over two years while frozen comparators degrade, without safety events attributable to updates, providing the evidence for a general adaptive-AI pathway.
Rationale
Sandboxes in fintech produced workable regulation for novel products faster than rulemaking; the same approach suits AI where the risks are poorly understood in advance.
What would test it
Run a two-year sandbox with three adaptive models at five sites; compare performance trajectories with frozen versions; publish the regulatory learnings.
Maturity
early clinical
Who has to act
regulator
Cost to try
Medium ($1M to $50M)
Years to first evidence
3
Bottlenecks it attacks
- AI that is built but not validated or deployed · Thousands of cancer AI models are published; a handful are in clinical use, and fewer have shown they help patients.
- Regulatory divergence between regions · Regulatory divergence means a drug approved in one country can take years to reach another, or never arrive.