National infrastructure proposal

UK Healthcare AI Safety Registry

A national, publicly searchable record of risks, near misses and reported harms from healthcare AI deployments. Grounded in the MIT AI Risk Repository taxonomy.

Concept prototype
Sarah Amani · Jun 2026
v0.9
Total reports
0
Near misses
0
Reported harms
0
Alerts issued
0
Trusts reporting
0
Report rate vs deploy
0%
Quick report · in 10 seconds designed for the end of a busy surgery

Two fields required. Full details optional. Reference number returned immediately.

Trend
Monthly reports submitted
By system type
Where the signal is coming from

Recent incidents

Incident database

The full record, anonymised before publication

Each incident classified using the MIT AI Risk Repository causal and domain taxonomies. Reports with linked outcomes shown in green.

12 incidents
Product directory

AI products with linked incident reports

Medicines and Healthcare products Regulatory Agency (MHRA) registration status cross-referenced. Use before procurement or when authoring clinical safety cases under DCB0129 (the NHS clinical risk management standard for digital health vendors).

10 products
MHRA registration note
Status cross-referenced at last update. Inspired by HSJ (Health Service Journal) finding that only 3 of 23 registered AVT (ambient voice technology) vendors were searchable on the MHRA prior to May 2026.
National safety alerts

Issued where patterns indicate systemic risk

Required action with deadline, cross-referenced to originating incidents. Closing the loop is what changes the reporting calculation.

Analytics

Signal, not noise

356 reports · Jan 2025 – May 2026
Severity distribution
By harm rating
Care setting
Where reports originate
MIT AI Risk Repository taxonomy
Incident classification by domain

Mapped against the MIT Domain Taxonomy (7 domains, 24 subdomains). airisk.mit.edu

Contributing factors
Where AI systems fail
Quarterly volume
Trajectory of reporting
Underreporting signal

Where the silence is loudest

Five Yellow Card reports across 15M+ ambient voice tool consultations is a measurement failure, not a safety record.

73%
Near miss or no harm
26%
Reported by clinicians
14
Days avg. to publication
Trust AI safety lead directory

Named accountability, publicly visible

A named trust AI safety lead is the minimum unit of governance that makes learning from this registry possible.

12 trusts
Submit a report

Minimum friction, maximum signal

Two fields required. Full details optional. Designed for the clinician who finishes surgery and notices something off. This registry accepts reports from patients too.

Prototype only — in a live service, submissions would be encrypted and reviewed by the clinical safety panel before publication.
Minimal report
Two fields, ten seconds
+ Add full details (optional but valuable)
By submitting you confirm all patient-identifiable information has been removed and this report is made in good faith to support NHS patient safety learning.
Governance & methodology

Theoretical foundation & operating principles

This registry complements — does not replace — LFPSE, MHRA Yellow Card, and trust-level incident reporting. Classification grounded in the MIT AI Risk Repository.

Theoretical foundation
MIT AI Risk Repository
A living database of 1,700+ AI risks with a Causal Taxonomy (Entity · Intent · Timing) and Domain Taxonomy (7 domains, 24 subdomains). Every incident in this registry is classified against both, giving NHS clinical safety officers a peer-reviewed shared vocabulary for DCB0129 hazard analysis.
→ airisk.mit.edu
Concept design: May–Jun 2026. Improved following: Luke Kellaway (Health Service Journal) on ambient voice technology (AVT) underreporting · Yash Morje on distributed failure pathways · Paul Brady on Failure Mode and Effects Analysis (FMEA) for AI · Oana on Health Services Safety Investigations Body (HSSIB) observation O/2026/087 on Electronic Prescribing and Medicines Administration (EPMA). Informed by: NHS Digital clinical risk standards DCB0129 (vendor) and DCB0160 (deploying organisation) · LFPSE (Learn From Patient Safety Events) · the MHRA Yellow Card scheme · the NHS Accessible Information Standard (DAPB1605) · the AHRQ (US Agency for Healthcare Research and Quality) AI Safety Program · UNESCO AI Ethics Recommendation (2021) · EU AI Act Annex III (referenced as international context only; the UK is not bound by it) · UK National Commission on the Regulation of AI in Healthcare (2025–26) · the MIT AI Risk Repository.