Monitoring, oversight & deployment governance6

Formally, estimation error shrinks with data while the human perception gap that produces stationary-environment black swans has a non-zero lower bound — so an incident-free operating history yields confidence without safety.

Sources: lee2025

Appears on: /pan-lab

A frontier risk-management framework in practice ties deployment authority to measured capability-vs-safety zones — green (routine plus monitoring), yellow (controlled with strengthened mitigations), and red (suspend).

Sources: shanghaiartificialintelligen2025, greenblatt2024, meinke2024

Appears on: /pan-lab

An authoritative review of deployed-AI monitoring finds staleness, performance drift, the right cadence of re-evaluation, and who acts on detected anomalies to be unresolved open challenges — and that systems can behave differently when they believe they are monitored — so post-deployment oversight is an unsettled, gameable control rather than a fixed guarantee.

Sources: rao2026

Appears on: /pan-lab

Two chapters of the volume describe the same gap from opposite ends. The governance chapter names an ethical capacity gap: many social workers have not been trained in data science or AI oversight, which it argues leaves them ill-prepared to question or interpret the algorithmic outputs they are nonetheless answerable for. The literacy chapter's professional-development framework assigns audiences by tier, placing sanctioned-tool lists, ethics review boards and vendor bias-mitigation terms with agency leadership while the skill to audit a decision and advocate for a misclassified client is taught at the tier below; it also reports professional-body guidance placing the duty to train on the employer rather than on the individual practitioner. Both are arguments about where authority sits relative to capacity. Neither reports a measured rate of either.

Sources: yang2026a, huang2026b, britishassociationofsocialwo2025b

Appears on: /pan-lab

Automated content moderation fails in two directions at once, and which direction it favors is a governance choice rather than a technical default. The volume's LGBTQIA+ chapter documents both halves landing on the same population: identity terms such as 'trans', 'queer' and 'nonbinary' have been flagged as inappropriate content while overt hate speech aimed at that population evades detection. The platform natural experiment already in this registry shows the same choice made explicitly rather than by default: with human review capacity withdrawn, the deployer said it would over-enforce rather than let harmful content stay up. The chapter is a peer-reviewed secondary synthesis and supplies the direction and the vocabulary, never a magnitude; the removal and reinstatement figures in this registry come from the platform's own transparency reporting under a separate claim and are not restated here. No outcome for the people whose content is moderated is computed anywhere in this Lab.

Sources: downey2026, youtubegoogle2020

Appears on: /domains/cases/youtube-covid-enforcement, /pan-lab

Two chapters converge on the procedural half of a second look: embed a procedure requiring a practitioner to document the reason for agreeing or disagreeing with an algorithmic suggestion, so supervision can examine how the person and the tool worked together instead of counting how often they agreed. The housing chapter supplies the failure it is written against — staffing reduced on the assumption that the tool is efficient turns review into rubber-stamping, and an organization that penalises deviation makes its own oversight performative. Both state it as a design prescription and neither offers an effect size.

Sources: yang2026d, shin2026

Appears on: /practice/structured-dissent