The FDA Accepted an AI Tool This Week. It's Not the Autonomous Diagnostic Approval — and That Distinction Matters.
textak currently forecasts a 54% probability that the FDA approves the first fully AI-driven diagnostic tool — down from 56% and still declining. This week the FDA accepted a letter of intent to evaluate an AI tool for predicting drug-induced liver injury, which will be read by some observers as confirmation of our thesis. We want to be precise: it's not. The DILI prediction tool acceptance is meaningful progress, but it is proximate evidence for our forecast, not direct evidence. Roche's deployment of 3,500 NVIDIA Blackwell GPUs for AI-accelerated drug discovery compounds the picture of serious institutional commitment. But serious commitment and autonomous regulatory approval are different things, and confusing them is how forecast credibility dies.
Let's be exact about what the FDA's DILI acceptance actually proves. A letter of intent to evaluate is the agency signaling openness to a conversation about an AI tool's regulatory pathway — it is not clearance, not approval, and not the removal of physician oversight from a diagnostic decision. This is the FDA in its early-growth posture: expanding its engagement surface, running parallel tracks on novel AI applications, and building the institutional knowledge base it will need to eventually make harder calls. The Roche deployment is similarly significant as infrastructure signal — pharma is moving from pilot to production at industrial scale — but a pharmaceutical company deploying GPUs for R&D acceleration is not the same as the FDA certifying that an AI system can make a final clinical diagnostic determination without human review.
The core tension in our 54% is the liability gap, and it has not moved this week. The DILI tool targets preclinical prediction — helping identify drug candidates likely to cause liver damage before expensive human trials. That's a high-value, relatively lower-stakes autonomous application compared to, say, autonomous cancer staging or autonomous cardiac event prediction in a live patient. The AMA's position on physician oversight hasn't shifted. The liability frameworks that would need to exist before a hospital system could defensibly deploy a fully autonomous diagnostic — malpractice insurance structures, institutional indemnification, clear FDA guidance on accountability chains — are not in place. No amount of GPU deployment changes that.
Honestly, the part of our thesis that keeps us up at night is the definitional problem we set for ourselves. 'Fully AI-driven diagnostic tool' without mandatory physician review for final decisions — we've been careful to exclude tools where physicians remain in the loop. IDx-DR has been FDA-cleared since 2018 for autonomous diabetic retinopathy detection. We scoped our forecast to exclude narrow single-condition autonomous tools precisely because that bar had already been crossed. But as the FDA clears more narrow autonomous applications across different specialties, the question of when 'multiple narrow autonomous tools' becomes 'a fully autonomous diagnostic system' gets genuinely ambiguous. That ambiguity is a risk to our resolution criteria, not just our thesis.
What would move us back toward 60%: an FDA public guidance document explicitly addressing liability frameworks for physician-free diagnostic AI, or a cleared device that spans multiple diagnostic domains without mandatory human review. What would drop us below 45%: continued AMA opposition hardening into formal lobbying position, or an AI diagnostic misdiagnosis event receiving significant public attention that creates political pressure for the FDA to pull back from autonomous framing. The DILI acceptance is real progress. It's just not the thing we're forecasting.