Perspectives
Short pieces on the foundational questions of agentic AI: integrity, authority, epistemic risk, identity, memory, and the conditions under which machine autonomy remains governed.
- 01What Is AI Integrity?AI integrity is the degree to which a system remains faithful to its intended objectives, constraints, permissions, and operating principles as conditions change.
- 02Adversarial AI: Why Governance and Safety Must ConvergeAgentic systems can be redirected through language, context, memory, and delegated authority — which is why AI governance, security, and safety begin to converge.
- 03Epistemic Risk in AIEpistemic risk is broader than hallucination: it concerns the integrity of the relationship between information, evidence, belief, and autonomous action.
- 04AI Governance Across the ArchitectureGoverning agentic AI requires a coherent chain from organizational intent to machine behaviour across models, memory, agents, tools, and runtime.
- 05What Does AI Control Mean?Meaningful AI control is a property of the architecture: permissions, observability, intervention, and the ability to reclaim delegated authority.
- 06Agentic Autonomy and Human AgencyMachine autonomy and human agency are not opposites. The question is which forms of autonomy remain compatible with governed delegation.
- 07From Alignment to AuthorizationAn AI system can be well aligned and still hold too much authority. Authorization translates institutional intent into concrete, revocable boundaries.
- 08Capability Is Not MandateCapability describes technical possibility; mandate defines legitimate scope. Expanding agent capability should not silently expand decision rights.
- 09The Epistemic Attack SurfaceAgentic AI introduces an epistemic attack surface: the information, memory, sources, and communications that shape what a system believes about its environment.
- 10Action Provenance: Knowing Why an AI ActedData provenance asks where information came from. Action provenance asks what caused an autonomous system to act — and what gave it the authority.
- 11Reversibility as a Principle of Safe AutonomyThe more reversible an action is, the more autonomy an organization may reasonably permit. Safe architecture makes failure recoverable, not merely unlikely.
- 12Governance at Machine SpeedWhen machines act faster than people can supervise them, governance must be embedded in the conditions under which they are allowed to act.
- 13Control DebtControl debt accumulates when organizations expand AI autonomy without proportionately expanding visibility, permissioning, containment, and recovery.
- 14When Safe Components Create an Unsafe SystemCompositional risk emerges from the relationships between models, memory, tools, and agents — not from a defect inside any one component.
- 15AI Identity: Who Is Acting?Autonomous systems need authenticated identities, identity-specific permissions, and a preserved record of which principal delegated their authority.
- 16Delegation Without AbdicationDelegation remains delegation when the delegating party can still determine, understand, constrain, and reclaim what has been handed over.
- 17Memory Is GovernanceOnce memory influences future decisions it becomes part of the governance architecture, requiring provenance, permissions, validation, and a lifecycle.
- 18Human Agency as a System PropertyMeaningful human agency depends on what the architecture allows a person to know, authorize, challenge, interrupt, change, and reclaim.