Evaluability
The capacity of knowledge systems to preserve relationships that allow claims to be reliably evaluated.
Research Question
What problem does the paper answer?
What systemic risks emerge when AI-mediated knowledge systems weaken the institutional conditions required to evaluate claims, evidence, provenance, and informational transformation?
Core Contribution
Evaluability as systemic infrastructure
This paper argues that systemic AI risk includes degradation of evaluability itself: the institutional capacity to sustain reliable evaluation through traceable, inspectable, and reconstructable relationships between claims, evidence, and informational transformation.
Executive Summary
General-reader summary
This paper examines a deeper risk in AI-mediated knowledge systems. The problem is not only that AI systems may produce misinformation, biased outputs, or opaque recommendations. The larger systemic danger is that institutions may lose the ability to evaluate information reliably as claims are transformed, compressed, recombined, and circulated through layered AI-mediated systems.
The paper introduces evaluability as a structural property of institutional knowledge systems. Evaluability means that claims remain connected to evidence, attribution, context, and transformation history in ways that can be inspected and reconstructed. When those relationships weaken, institutions may still function operationally while becoming less able to verify the informational foundations supporting their decisions.
By drawing on infrastructure studies, archival theory, provenance research, sociotechnical systems analysis, and AI governance, the paper reframes systemic risk as the breakdown of evaluation conditions. It shows how recursive synthesis, informational compression, accelerated circulation, and provenance opacity can weaken institutional verification across scientific, administrative, economic, and media environments.
Source Abstract
Manuscript abstract
Artificial intelligence systems increasingly mediate knowledge production, circulation, synthesis, and institutional evaluation across scientific, administrative, economic, and media environments. Existing AI governance debates focus heavily on transparency, accountability, misinformation, and algorithmic bias, yet less attention has been directed toward whether the evaluative conditions required for reliable institutional verification remain stable under AI-mediated informational scale.
This paper argues that a central systemic risk emerging within AI-mediated knowledge systems involves degradation of evaluability itself: the institutional capacity to sustain reliable evaluation through traceable and reconstructable relationships between claims, evidence, and informational transformation. Drawing from infrastructure studies, provenance research, archival theory, and AI governance literature, the paper conceptualizes evaluability as a structural property of institutional knowledge systems.
The analysis examines how recursive synthesis, informational compression, accelerated circulation, and provenance opacity weaken evaluative continuity across distributed sociotechnical environments.
AI governance
evaluability
institutional verification
provenance
sociotechnical infrastructure
systemic risk
Key Concepts
Terms to remember
Evaluation breakdown
The weakening of traceability, inspectability, provenance, and reconstructability across mediated systems.
Systemic AI risk
Risk emerging from degraded institutional evaluation conditions, not only from isolated bad outputs.
Recursive synthesis
The repeated reuse of AI-mediated outputs as inputs for further synthesis and circulation.
Provenance opacity
The loss or obscuring of claim origins, evidentiary lineage, and transformation history.
Verification pressure
The growing institutional need to preserve evaluation conditions under AI-mediated informational scale.
Why It Matters
Institutions need evaluation to remain stable
This paper matters because it shifts systemic AI risk from individual model failures to institutional conditions of evaluation. If institutions can no longer reconstruct how claims, evidence, attribution, and transformations relate, then accountability, verification, trust, and coordination all become more fragile.
Relationship to Other Papers
From evaluability to systemic risk
This paper extends the verification and preservation infrastructure sequence into systemic risk analysis. It connects earlier papers on verification, provenance continuity, and structural evaluability to broader questions of institutional coordination, AI governance, and sociotechnical infrastructure.
Research Program Position
Where this paper fits
This paper marks the point where evaluability becomes a systemic-risk concept. Earlier papers developed verification, preservation infrastructure, uncertainty, and claim-evidence continuity. Paper 15 shows what happens when those conditions degrade across institutions rather than within a single output or system.
Within the broader research program, this paper functions as the bridge between verification theory and AI-mediated institutional risk. It anticipates later work on distributed witnessing, historical continuity, and the preservation of knowledge systems by showing that the future problem is not merely whether claims are true, but whether societies can still evaluate how claims came to be trusted.
Submission History
Journal path
| Date submitted | Journal | Submission ID | Decision / status |
|---|---|---|---|
| May 11, 2026 | Technology in Society | TECHIS-D-26-03383 | Desk rejected, May 20, 2026 |
| May 22, 2026 | Synthese | SYNT-D-26-01992 | Desk rejected |
| May 25, 2026 | New Media & Society | NMS-26-1869 | Under consideration |
This page preserves the current surviving version associated with Paper 15. Earlier submissions under this paper number appear to have evolved through revision and retitling before the current New Media & Society submission.
One Sentence Summary
The systemic risk is not only that AI systems produce bad claims, but that institutions lose the ability to evaluate claims at all.