AI evidence extraction

Inspectable AI for Science: A Research Object Approach to Generative AI Governance

Ruta Binkyte, Sharif Abuaddba, Chamikara Mahawaga, Ming Ding, Natasha Fernandes, Mario Fritz · ArXiv.org · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

1/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

Methodology & findings

Study design

Conceptual framework proposal with illustrative workflow demonstration.

Primary method

Design science / infrastructure research. The authors propose a conceptual framework and demonstrate feasibility through an illustrative end-to-end transparency pipeline for literature review writing.

Main result

The paper argues that "legitimacy in AI-assisted scientific writing is procedural rather than binary" and proposes that "treating AI systems as Research Objects whose configuration, interaction history, and outputs can be documented, inspected, and evaluated alongside data and software artifacts" provides a structured approach to governance. The framework shifts accountability "from attribution disputes toward methodological transparency" by operationalizing AI usage as structured provenance.

Research paradigm

Design science / infrastructure research

Author conclusions

The authors conclude that "Generative AI is rapidly becoming embedded in the everyday practice of scientific writing" and that existing approaches "struggle to capture the infrastructural role that AI now plays in research workflows." They argue that the AI-RO framework "shifts governance from attribution disputes toward methodological transparency" and stress that "legitimacy in AI-assisted scholarship depends less on whether AI is used than on whether its influence is inspectable." They further emphasize that "future research infrastructures, particularly in S&P domains, should incorporate mechanisms for secure AI-assisted workflows' provenance capture and controlled disclosure as first-class requirements."

Risk of bias

No empirical study conducted; this is a conceptual proposal. Potential bias: the framework is presented as normative without user validation or empirical testing of whether the proposed mechanisms actually improve governance outcomes.

Limitations

The authors acknowledge that "the framework and demonstration intentionally focus on a single stage of the research lifecycle (i.e., drafting a RELATED WORK section) and do not constitute a comprehensive governance solution." They further note that "the proposed workflow assumes standalone structured interaction with AI, whereas real-world AI use is often exploratory, iterative, and difficult to scope" and that "a rigorous validation of the framework requires human study to evaluate the usefulness of the artifacts for scientific review and to assess the trustworthiness of the scientific work." The framework is characterized as "an initial architectural proposal that delineates a design direction rather than a fully operational standard."

Open questions raised

The paper identifies the following future research directions: (1) Standardization of AI usage documentation; (2) Developing privacy-aware provenance infrastructure and anonymization protocols for AI-RO; (3) Supporting cultural norms that reward transparency and usable audit interfaces; (4) Encouraging development and adoption of open and responsibly trained models; (5) Developing protocols and tools guiding AI use in research; (6) Building tools and interfaces to support provenance in unstructured, iterative, and fragmented interactions with AI; (7) Developing AI-RO governance norms tailored to security and privacy research.
The authors identify several future research directions: (1) standardization of AI usage documentation; (2) developing privacy-aware provenance infrastructure and anonymization protocols; (3) supporting cultural norms that reward transparency and usable audit interfaces; (4) encouraging development of open and responsibly trained models; (5) developing protocols and tools guiding AI use in research; (6) building tools for provenance in unstructured, iterative interactions; and (7) developing AI-RO governance norms tailored to security and privacy research.
Standardization of AI usage documentation
Developing privacy-aware provenance infrastructure and anonymization protocols for AI-RO
Supporting cultural norms that reward transparency and usable audit interfaces
Encouraging development and adoption of open and responsibly trained models

Data: https://github.com/RutaBinkyte/AI-RO; https://osf.io/zbrvuCode: https://github.com/RutaBinkyte/AI-RO; https://osf.io/zbrvuExtracted from: pdfAgreement 70%

Explore related topics

Responsible AI Use AI Governance Position Paper