Disaggregating Safety: Harm Prevention, Liability, and Control in AI Governance

A safety claim describes a relation between a protected party and a harm. Its force depends on which party it protects and which harm it prevents, and AI developers seldom say. Recent scholarship records the expansion of the word in AI discourse to cover catastrophic risk, systemic injustice, and psychological wellbeing, and finds that many safety benchmarks track general capability. Commercial terms and corporate statements from 2024 through 2026 place the prevention of violence and fraud beside the protection of revenue, intellectual property, reputation, and state interests under the same label. A seven-element test separates these concerns. The test asks each claim to name the threatened party and the anticipated harm, the evidence that the harm is plausible, and the mechanism by which the intervention reduces it. It also asks for the rights or benefits sacrificed, the distribution of costs, and the procedure for appeal and revision. California's frontier AI statute and the federal review of open model weights show the test applied to private claims. The 2026 litigation in two federal courts over the Pentagon's designation of an AI developer as a supply chain risk shows it applied to a government's claim, which failed under one statute and survived under another.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24415916
Primary Topic
Digitalization, Law, and Regulation
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Disaggregating Safety: Harm Prevention, Liability, and Control in AI Governance

Felix Velasco, Josie Jefferson
Zenodo (CERN European Organization for Nuclear Research)
Digitalization, Law, and Regulation
preprint

Disaggregating Safety: Harm Prevention, Liability, and Control in AI Governance

Felix Velasco, Josie Jefferson
preprint en

Abstract

A safety claim describes a relation between a protected party and a harm. Its force depends on which party it protects and which harm it prevents, and AI developers seldom say. Recent scholarship records the expansion of the word in AI discourse to cover catastrophic risk, systemic injustice, and psychological wellbeing, and finds that many safety benchmarks track general capability. Commercial terms and corporate statements from 2024 through 2026 place the prevention of violence and fraud beside the protection of revenue, intellectual property, reputation, and state interests under the same label. A seven-element test separates these concerns. The test asks each claim to name the threatened party and the anticipated harm, the evidence that the harm is plausible, and the mechanism by which the intervention reduces it. It also asks for the rights or benefits sacrificed, the distribution of costs, and the procedure for appeal and revision. California's frontier AI statute and the federal review of open model weights show the test applied to private claims. The 2026 litigation in two federal courts over the Pentagon's designation of an AI developer as a supply chain risk shows it applied to a government's claim, which failed under one statute and survived under another.

Zenodo (CERN European Organization for Nuclear Research)
Digitalization, Law, and Regulation
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon
Disaggregating Safety: Harm Prevention, Liability, and Control in AI Governance — Felix Velasco, Josie Jefferson · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS