01
Michigan Medicine / Epic Sepsis Model v1
Editorial leadReal-time proprietary sepsis alerts
A deployed alert can add workload while missing most cases.
← Collections
Negative results
The failed and mixed results are part of the evidence.
A useful evidence base includes systems that missed cases, created harm, failed to improve the target metric, or worked only inside a bounded capability frontier. These records make the limits visible instead of selecting only favorable outcomes.
Inclusion criteria
Includes cases explicitly identified as negative, failed, or biased, plus cases whose published outcome reports no significant improvement, lower correctness, missed cases, harm, or a material adverse result.
10 cases · bundle 1.0.0 · reviewed 2026-08-23
02
Netherlands Tax and Customs Administration / Benefits
Editorial leadSelf-learning risk-classification model for benefit applications
Fraud-risk triage can cause institutional harm without fair features, recovery rules, or challenge rights.
03
Large US health system / Optum algorithm
Editorial leadCommercial cost-proxy risk ranking
A predictable proxy can still allocate scarce care unjustly.
Current · updated
1 independent, 1 peer reviewed; publication outcomes are verified.
Canonical case →04
Copenhagen Emergency Medical Services
Editorial leadReal-time ML listening and dispatcher cardiac-arrest alerts
A live emergency-call alert may not improve dispatcher recognition despite higher standalone model sensitivity.
05
Boston Consulting Group
Editorial leadGPT-4 integrated into individual consulting task flows
That access to a general model improves all apparently similar knowledge tasks.
06
Swiss Plasma Center at EPFL and Google DeepMind
Deep-reinforcement-learning magnetic control on the TCV tokamak
Operators set plasma objectives while a simulator-trained policy coordinates all coils on physical hardware.
07
John Deere (Deere & Company)
In-season post-emergence herbicide application (See & Spray targeted spraying)
That a sprayer cannot tell a crop plant from a weed while moving through a field, so the only reliable way to control weeds is to broadcast herbicide across the entire field regardless of where weeds actually are.
Current · updated
1 independent, 1 primary; publication outcomes are verified and reported.
Canonical case →08
TD Bank Group
Mortgage and HELOC pre-adjudication and application memo preparation
That preparing a mortgage or HELOC file for a credit decision requires a human to read the document package end to end - classifying documents, extracting figures, calculating income, checking consents and writing the summary - because borrower documents are too variable for a rules engine to handle reliably.
Current · updated
1 independent, 1 primary; publication outcomes are verified and reported.
Canonical case →09
UCLA Health
Randomized deployment of DAX and Nabla ambient AI scribes
Ambient tools can draft notes in practice, but vendor-specific efficiency and error results require comparison.
10
Be My Eyes
Conversational multimodal visual assistance through Be My AI
A blind or low-vision user can get immediate, private scene descriptions before choosing human assistance.
Watch status: verify the cited source and deployment condition before reusing this case.
Watch · updated
Watch status: verify the cited source and deployment condition before reusing this case.
1 peer reviewed; publication outcomes are verified.
Canonical case →