Evidence gap map: AI governance frameworks in the scholarly literature

In plain terms

"AI governance framework" is one of the most-used phrases in the field — but what kinds of frameworks does the scholarship actually contain, and on what evidence do they stand? The engine retrieved a corpus on the question, coded every abstract into a framework matrix, and counted. The picture: a literature rich in conceptual models and principle sets, thin on statute, and almost empty of empirical testing — 21 of 30 matrix cells contain no papers at all.

Finding (screened coverage map): Across a 5x6 framework matrix, 21 of 30 cells have no papers in this corpus of 16 included records; the populated cells concentrate in aspect 'conceptual_governance_model' (6 papers) and evidence type 'conceptual_normative' (8 papers).

Background

Policy Window's research engine selected the review kind BEFORE this article was drafted: a review-kind selector matched the question ("What does the scholarly literature report on AI governance frameworks?") and the measured corpus features (29 records, 24 with abstracts, no comparable quantitative effects, abstracts only) to an evidence gap map (selection f41ed8259d2bd65e). The topic is the engine's own study queue's top unserviced entry. The article's status is a screened candidate routed to human review, not an adjudicated truth (Sacred Rule 9).

Method

Every record was coded exactly once against a declared frame — five framework aspects (conceptual models, ethics-principles sets, statutory frameworks, national/institutional frameworks, sector application frameworks) by six evidence types — with a verbatim rationale span per inclusion and a stated reason per exclusion (13 records excluded: off-topic or no abstract). The coverage table is arithmetic over those codings. Review report befadbb8963c3084; corpus b6f941a82b8594df.

What the map shows

The populated region: conceptual governance models (layered models, roadmaps, definitional work, hybrid designs), national and institutional frameworks (the NIST AI Risk Management Framework, Singapore's evolving framework, public-administration designs), one statutory entry (the EU AI Act analyzed doctrinally), and sector application frameworks in health, education, and marketing. The named empty cells include nearly every aspect crossed with empirical evidence: one included paper reports a quantitative survey, none reports qualitative empirical study of a framework in use, and no framework aspect carries more than a single empirical paper in this corpus. Frameworks are proposed far more often than they are tested.

Limitations (disclosed by the engine)

• scoping retrieval with fixed queries, not a systematic search (coverage is query-bounded)
• coding from abstracts only — full texts were not consulted; cells count papers, not extracted effect estimates
• single-annotator coding (in-session:claude-agent); no second coder, no kappa
• corpus bounded to 29 retrieved records; counts are corpus-relative, not field-level claims

Verify

Every count above re-derives offline from the committed corpus and codings — no model, no network, no trust in the institute required: PYTHONPATH=src python scripts/verify_review_framework_ai.py # exit 0 = re-derives offline