Irreversible Harm in AI Safety: Harm Class

Harm Class (Reversible or Irreversible)

Definition

A harm class records whether a potential AI-caused harm could be undone after it happens. Reversible harms can be detected and reverted in time; irreversible or irremediable harms cannot, so safety frameworks give them priority and rely on prevention rather than response.

Also called: reversibility, irremediable harm, irreversible harm, Velocity and Irreversibility

NIKOLAI's proposed operational definition

NIKOLAI editorial proposal (unsourced): harm class is a controlled value attached to a risk pathway, threat scenario or outcome. Its values are reversible (the harm can be detected and reverted within a stated response window), difficult to reverse (remediation needs extraordinary effort, resources or time) and irreversible (no feasible remediation), with a separate flag for instantaneous onset and the response window stated where the source gives one.

The operational definition is CASRAI's own proposal, not a definition any named organisation has agreed to. See what NIKOLAI is and is not, or browse the AI safety glossary.

Examples

Examples from published documents

Anthropic treats self-exfiltration and persistent rogue internal deployment as "the main pathways that we consider potentially irreversible, because they undermine our ability to detect (and thus mitigate) future misaligned actions" (Anthropic Risk Report, August 2026, section 2.13)
OpenAI made AI self-improvement its own Tracked Category "because it presents a distinct plausible, net new, and potentially irremediable risk" (OpenAI Preparedness Framework v2, Table 2 note)

Counter-examples

Not to be confused with

Severity is a separate axis: a harm can be modest but irreversible, or very large but recoverable. Harm magnitude in deaths or dollars is recorded under the NIKOLAI element severity-threshold.

Related terms

Related AI safety terms

Terms defined under this element that are not NIKOLAI elements themselves, each as the cited document uses it. Every term is also listed in the AI safety glossary.

Reversible harm: Reversible harm, in Anthropic's August 2026 Risk Report, is harm that "can be reverted even if responded to within a few weeks". Anthropic relies on offline monitoring to catch such harms before a later critical moment.; Source: Anthropic Risk Report, August 2026
Irremediable harm: Irremediable harm is harm that cannot be repaired once it occurs. OpenAI's Preparedness Framework v2 and Meta's framework both use "Instantaneous or irremediable" as a test for which risks to cover: harms that are "inevitable due to a lack of feasible measures to remediate".; Source: OpenAI Preparedness Framework v2

Source of record

Where this definition comes from

Anthropic Risk Report, August 2026, §2.11.1, §2.13
“"What we call “reversible” harm can be reverted even if responded to within a few weeks" (§2.11.1). "Pathways 5 and 6 are the main pathways that we consider potentially irreversible, because they undermine our ability to detect (and thus mitigate) future misaligned actions." (§2.13)”
https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdf
EU GPAI Code of Practice, Safety and Security Chapter, Appendix 1.2.2, p.35
“"(5) Difficult or impossible to reverse: Once materialised, the risk creates persistent changes that require extraordinary effort, resources, or time to remediate, or are permanently irreversible."”
https://ec.europa.eu/newsroom/dae/redirection/document/118119

Crosswalk

How named organisations use this concept

Every row below is a shadow mapping. A shadow row is CASRAI's own reading of a published document. No lab, evaluator or regulator named on a shadow row has declared, endorsed, or been consulted on it. That changes only when an organisation files its own Mapping Declaration.

Organisation	Their term, as published	Match & verification	Source
AnthropicShadow mapping Anthropic Risk Report, August 2026	“"What we call “reversible” harm can be reverted even if responded to within a few weeks" (§2.11.1); "most of our pathways consider reversible harms"; "Pathways 5 and 6 are the main pathways that we consider potentially irreversible" (§2.13).”	exactEQ confidence: high	Anthropic Risk Report, August 2026
AnthropicShadow mapping Anthropic Responsible Scaling Policy v3.4	“Misalignment row: "sabotage leading to irreversibly and substantially higher odds of a later global catastrophe"; "Capability assessments demonstrating that AI models lack the ability to carry out irreversible (which would generally mean unnoticed) sabotage." (section 1 table, pp.7-8)”	closeCL confidence: high	Anthropic Responsible Scaling Policy v3.4
OpenAIShadow mapping OpenAI Preparedness Framework v2	“Tracked Category criterion 5: "Instantaneous or irremediable: The outcome is such that once realized, its severe harms are immediately felt, or are inevitable due to a lack of feasible measures to remediate." (§2.1)” Used as an inclusion criterion that bundles irreversibility with instantaneous onset, not as a per-pathway label.	closeCL confidence: high	OpenAI Preparedness Framework v2
MetaShadow mapping Meta Advanced AI Scaling Framework v2	“Inclusion criterion "Instantaneous or irremediable": "The outcome is such that once realized, its catastrophic impacts are immediately felt, or inevitable due to a lack of feasible measures to remediate." (§3.2, p.14); criterion "Catastrophic": "The outcome would have large scale, devastating, and potentially irreversible harmful effects." (§3.2, p.13)”	closeCL confidence: high	Meta Advanced AI Scaling Framework v2
EUShadow mapping EU GPAI Code of Practice, Safety and Security Chapter	“Contributing characteristic "(5) Difficult or impossible to reverse: Once materialised, the risk creates persistent changes that require extraordinary effort, resources, or time to remediate, or are permanently irreversible." (Appendix 1.2.2)”	exactEQ confidence: high	EU GPAI Code of Practice, Safety and Security Chapter
Frontier Model ForumShadow mapping FMF Risk Taxonomy and Thresholds	“"4. Velocity and Irreversibility: Impacts could be instantaneous or potentially irreversible once triggered. These risks typically lack adequate warning signs before manifesting fully" (s1.3, p.5).”	closeCL confidence: high	FMF Risk Taxonomy and Thresholds

What do these codes mean?

exact: The source term is equivalent to this element
close: The source term is close but not equivalent to this element
broad: The source term is broader than this element
narrow: The source term is narrower than this element
none: No mapping claim — used for false-friend and declared-but-undefined rows
EQ: Equivalent
CL: Close
BR: Source is broader than the element
NR: Source is narrower than the element
FF: False friend — same or similar label, different meaning
DU: Declared but undefined by the source
UV: Unverified

Divergence

Where sources materially disagree

Anthropic's reversibility is operational and tied to detection: a harm is "reversible" if it can be reverted when responded to within a few weeks, and RSP v3.4 glosses irreversible sabotage as sabotage that "would generally mean unnoticed". The EU chapter and the FMF treat reversibility as a property of the materialised harm. OpenAI and Meta fold it, together with instantaneous onset, into one criterion for deciding which risks a framework covers.

Gap

No reversibility label was found in Google DeepMind's FSF v3.1. No source gives a shared response window: the one stated window CASRAI found is Anthropic's "within a few weeks".