Sociotechnical evaluation & risk framing12

A survey of generative-AI safety evaluations found 85.6% operate at the model-capability layer, only 5.3% at the human-interaction layer and 9.1% at the systemic-impact layer — yet context determines whether a capability becomes harm, so the human and system layers where risk actually manifests are the least evaluated.

Sources: weidinger2023

Appears on: /pan-lab

AI-safety failure classification has a missing interaction layer between institutional risk categories and system-level failure modes: practitioners lack a shared vocabulary of recognizable error patterns, and the catch-all 'hallucination' collapses distinct logic failures whose correct fixes differ.

Sources: beyer2026

Appears on: /pan-lab

A data-driven taxonomy built from 9,705 real AI-incident reports found mitigation practice dominated by reactive and legal levers (incident investigation, reporting, regulatory and court action) while proactive technical and governance levers (model alignment, safety frameworks, board oversight) were least common — real organizations respond after harm rather than preventing it.

Sources: popchanovska2026, slattery2024

Appears on: /pan-lab

Trustworthiness measured at the model or benchmark level does not transfer to the deployed system: standard benchmarks compare models but do not cover the aspects that matter most in a specific application context, so safety and responsibility are properties of the system-in-context — its users, incentives, and institutions — not of the model alone.

Sources: mitra2025

Appears on: /pan-lab

AI failures often originate not in individual models but in the architecture of the decision process - recurring failure topologies including temporal feedback instability (small errors amplified through loops) and relational propagation (errors spreading through network structure) - so safety is a property of the decision architecture, not the model alone.

Sources: cemri2025, perdomo2020

Appears on: /pan-lab

The volume's child-welfare chapter reports that a widely deployed screening score predicts whether a child will be placed out of the home within two years, which is the system's own future response rather than the maltreatment the worker is deciding about. The chapter treats the gap between the modelled target and the decision's actual question as a design property of the deployment, not as a defect in the model's accuracy.

Sources: zhang2026b

Appears on: /pan-lab

The volume's criminal-justice chapter separates three routes by which a risk instrument inherits disparity. The training target is usually arrest, charge or conviction rather than offending itself, which is largely unobserved. The strongest inputs are typically prior system contacts, which carry the disparity of the enforcement that produced them. And the sample is drawn from the population the system already touched. The chapter treats these as distinct routes, so a remedy aimed at one does not address the others.

Sources: ahn2026

Appears on: /domains/security-operations-fraud

The volume's poverty chapter uses the targeting literature's paired vocabulary for the two directions in which a system steering a scarce resource fails: an inclusion error reaches someone the program did not intend to reach, and an exclusion error leaves out someone it did intend to reach. The chapter reports reducing both as the stated aim of machine-learning-assisted eligibility and proxy-means targeting. It is a narrative review and measures neither rate itself; the accuracy results it summarises belong to its sources.

Sources: zeng2026

Appears on: /pan-lab

The disability chapter is the one setting in the volume where the AI is the assistance itself rather than a decision aid steering an institutional decision about a person, which is the case that most tests the operator-network boundary. The resolution the framework already carries holds: model institutional propagation, keep clinical and operations staff in the operator network, and record the assisted person's own outcome externally. No outcome for a served person is computed from any diagram here.

Sources: wang2026a

Appears on: /pan-lab

The older-adults chapter supplies the harder substitution case: where virtual contact substitutes for face-to-face contact, or a monitoring device substitutes for a person looking, the party who stops looking can be an unpaid family caregiver rather than an employee the organization can train, roster or audit. Substitution drawn on a staff link assumes an authority relationship that does not exist in that case, and no lever in the catalogue reaches that person.

Sources: shen2026

Appears on: /pan-lab

The housing chapter names consequences that sit outside an operator-network model entirely rather than being merely unmodelled attributes of people: spatial stigma and the disinvestment or gentrification pressure that follows an area being algorithmically labelled as declining on superficial visual indicators, and a landlord's willingness to rent. These are place- and market-level effects, and nothing in a governance diagram computes them.

Sources: shin2026

Appears on: /pan-lab

The volume's conclusion proposes four distinct evaluation layers for AI in social work: technical performance, distributive impact, procedural fairness and experiential legitimacy. It is a normative proposal, not an empirical finding, and the chapter reports no original measurement of any layer. Read against this Lab: pathway and lever direction sit in the technical-performance layer, the external equity surface is distributive impact and is recorded rather than derived, and procedural fairness and experiential legitimacy have no surface here at all.

Sources: yang2026d

Appears on: /pan-lab