AI adoption in high-consequence tasks will continue to diffuse slowly.
5 events · 2 assessments · 2 decisions
Reassessed no status change
The adaptation-limits subclaim's first assessment (supported 0.6) is consistent with the existing verdict. It modestly strengthens the barriers argument. Status kept at supported 0.55, credence 0.65 added, arguments re-evaluated with unchanged ratings. Not a material change, so dependents were not notified.
Reassessed: still Supported
verdict confidence 0.55 · credence 0.65
Structured and assessed
First pass. Created three subclaims after match_claim returned novel for each (regulatory and liability barriers; traditional statistics in predictive optimization; informal professional uptake, as a contradicting claim) and linked the existing institutional-adaptation claim 3bd624a7. Two named arguments, one for and one against, written and evaluated. Assessed supported at 0.55 confidence, credence 0.62, from 2026 health-sector deployment data, AMA physician surveys and Waymo ridership. Importance 0.55, contestation 0.6. No dependent notification: this is a first assessment, and the parent 1e140cf3 will read it when its own pass runs.
Assessed Supported
verdict confidence 0.55 · credence 0.62
The claim is a forecast, made most prominently in Narayanan and Kapoor's essay "AI as Normal Technology", that AI will keep spreading only slowly into tasks where errors carry serious consequences, such as clinical diagnosis, criminal justice and lending. "Slowly" has no fixed benchmark. Read as the authors intend it, the claim is about AI being trusted with consequential decisions inside institutions, compared with the pace of capability gains and of AI uptake in low-stakes work. On that reading, the record up to late 2026 mostly bears it out. The case for the claim rests on regulation, liability and safety validation slowing deployment in high-consequence domains, on the general constraint that diffusion moves at the speed institutions can adapt, and on a track record in which consequential predictive systems have relied mainly on traditional statistics. The 2026 evidence fits this picture. The International AI Safety Report 2026 notes that healthcare AI matching human diagnostic accuracy in research often needs three to five more years to reach clinical use. Industry surveys report that fewer than one in five health systems use AI reliably in core clinical diagnosis, and that clinical AI reaches full deployment at about half the rate of administrative AI. The evidence on the other side is real. Professionals in high-stakes fields are taking up general-purpose AI quickly: AMA surveys show physician use rising from 38% in 2023 to 81% in early 2026, much of it daily. Autonomous driving, a high-consequence task by any measure, grew about tenfold in two years, to roughly 500,000 paid robotaxi rides a week. Most physician use is documentation and reference rather than delegated decisions, though, and robotaxis still carry a tiny share of trips. So these cases are fast growth from a small base rather than a refutation. Whether the claim holds depends on two things: how adoption is measured (informal use versus formal, accountable deployment), and whether the bottom-up uptake now visible turns into AI shaping consequential decisions. Over the next few years that would show up in clinical, legal and lending outcomes data, and in whether robotaxi growth continues.
Claim entered the graph