Von menschlichen Experten begutachtet: Die KI scheiterte bei wichtigen Schritten bei der Generierung einer Scoping-Überprüfung

A group of international researchers found that AI models currently fail to generate valid scoping reviews. While known issues such as hallucinations and difficulties with information hierarchization remain challenges (which may be solvable via future agentic/continuous learning frameworks), the study identifies a more critical structural „blindspot“, the inability to verify data behind subscription paywalls during inference.

The model relied heavily on Open Access content or abstracts, resulting in a biased dataset. This raises important questions regarding the „zero-click“ internet: as users shift toward AI-generated summaries, paywalled research could become less visible to the discovery process.

IMHO this dynamic may eventually prompt authors and publishers to adapt their practices to ensure AI discoverability, or potentially incentivize publishers to develop proprietary AI interfaces that specifically prioritize their own archived content.