Qurated: From ML Predictions to Informed Diagnostic Assistance Using the Toulmin Model of Argumentation

The Insight: A Prediction Is Not an Argument

An AI that says "this retina shows diabetic retinopathy" gives you a verdict. It doesn't give you a case. The difference matters — because clinicians don't need answers, they need reasons they can interrogate.

This paper's move is deceptively simple: stop treating ML output as a conclusion to accept, and start treating it as a claim to be argued. The vehicle is a 60-year-old framework from philosopher Stephen Toulmin — and it turns a black box into something a human expert can actually reason with.

The Toulmin Model, Applied to AI

Toulmin decomposed any real-world argument into six parts. Map them onto a diagnosis and the fog lifts:

• Claim — what the ML model concludes ("diabetic retinopathy").
• Grounds — the evidence. Here, a specialized model extracts biomarkers from the image.
• Warrant — the reasoning linking grounds to claim. A medical-knowledge agent (MedGemma) asks: do these biomarkers actually justify this diagnosis?
• Qualifier — the confidence hedge ("probably," "almost certainly"), derived from quantitative evaluation of both grounds and warrant.
• Rebuttal — the conditions under which the claim fails. Built via image-similarity comparison (MedSigLip) to known counter-cases.
• Backing — the deeper authority supporting the warrant (established medical literature).

The genius: each component is a separate, inspectable module. When a diagnosis is wrong, you can locate the failure — bad evidence, faulty reasoning, or overconfidence.

Why This Beats "Explainable AI"

Standard XAI shows you where the model looked (heatmaps, saliency). Argumentation shows you why the conclusion follows. A heatmap over a lesion tells you nothing about whether that lesion warrants the diagnosis. Toulmin forces the system to expose its logic, its confidence, and — critically — its own conditions for being wrong.

Most AI tools never state their rebuttal. This one does.

The Portable Mental Model

Use this whenever you receive a confident claim — from an AI, an analyst, or yourself:

The 6-Question Audit

1. Claim: What exactly is being asserted?
2. Grounds: What evidence supports it?
3. Warrant: Why does that evidence lead to this conclusion?
4. Qualifier: How certain are we, honestly?
5. Rebuttal: Under what conditions would this be false?
6. Backing: What established knowledge justifies the reasoning itself?

Most confident assertions collapse at question 3 or 5. The warrant is assumed; the rebuttal is never considered. Naming these gaps is how you convert blind trust into calibrated judgment.

The Actionable Shift

Design AI to argue, not to answer. If you build or deploy decision-support tools, don't ship a verdict. Ship the argument — grounds, warrant, qualifier, rebuttal — and let the expert be the judge. This preserves human agency while raising the quality of the human's decision.

Demand rebuttals. The most trustworthy system is the one that tells you when to distrust it. Any model, dashboard, or advisor that cannot state the conditions of its own failure has not earned your confidence.

Separate evidence from reasoning. Much of human error comes from conflating "I have data" with "therefore my conclusion holds." Toulmin's split between grounds and warrant is a discipline for thinking clearly under uncertainty — in medicine, in investing, in every high-stakes judgment.

The deepest lesson isn't about retinas. It's that the future of AI-human collaboration isn't automation — it's argumentation. The best systems won't replace expert judgment. They'll structure the debate that sharpens it.

Sources & Further Reading

• https://arxiv.org/abs/2607.09664
• Stephen Toulmin, The Uses of Argument (1958) — the original framework.