Back to AI Research

AI Research

LAAF: A Layered Accountability Architecture Framewo... | AI Research

Key Takeaways

  • LAAF: A Layered Accountability Architecture Framework for LLM Applications provides a systematic approach to assigning responsibility when Large Language Mod...
  • Large Language Models (LLMs) operate in hospitals, courtrooms, banks, and public service desks, where fluent, confident outputs are treated as authoritative even when ungrounded or incorrect.
  • When such an output contributes to harm, who is answerable, and through what mechanisms can responsibility be traced, explained, and acted upon?
  • The corpus is read through a four-layer classification device spanning provenance, application logic, human oversight, and governance and redress, cross-cut by traceability, role clarity, and continuous monitoring.
  • The authors argue that because LLMs are probabilistic systems that can produce ungrounded or incorrect outputs, accountability cannot be treated as a simple technical fix.
Paper AbstractExpand

Large Language Models (LLMs) operate in hospitals, courtrooms, banks, and public service desks, where fluent, confident outputs are treated as authoritative even when ungrounded or incorrect. When such an output contributes to harm, who is answerable, and through what mechanisms can responsibility be traced, explained, and acted upon? Following PRISMA guidance, five databases were searched from January 2022 to March 2026 against four review questions; of 4,512 records identified, 122 primary studies were included, together with 12 regulatory and standards documents analysed as primary sources. The review consolidates a sociotechnical account of accountability as an actor-forum relation resolved into five dimensions, and synthesises mechanisms across four families: technical controls, human oversight, organisational governance, and documentation and traceability, each with a maturity assessment. The corpus is read through a four-layer classification device spanning provenance, application logic, human oversight, and governance and redress, cross-cut by traceability, role clarity, and continuous monitoring. Both are mapped onto the EU AI Act, whose high-risk obligations have applied since 2 August 2026, the NIST AI RMF with its Generative AI Profile, ISO/IEC 42001, and sectoral guidance in healthcare, consumer finance, education, and the public sector. Four persistent gaps emerge: under-specification of human oversight, absence of shared accountability metrics, disciplinary disconnection, and limited empirical evaluation, alongside five structural tensions that no surveyed instrument resolves. The review closes by consolidating the classification device into an integrated accountability architecture, LAAF, with cybersecurity aligned to the OWASP LLM Top 10 (2025); it is a synthesis of the surveyed evidence rather than a validated artefact.

LAAF: A Layered Accountability Architecture Framework for LLM Applications provides a systematic approach to assigning responsibility when Large Language Models (LLMs) cause harm in high-stakes environments like healthcare, finance, and law. The authors argue that because LLMs are probabilistic systems that can produce ungrounded or incorrect outputs, accountability cannot be treated as a simple technical fix. Instead, they propose a framework that maps technical, human, and organizational mechanisms onto current regulatory requirements to ensure that when a system fails, there is a clear, traceable path to identify who is answerable and how to provide redress.

The Accountability Gap

The authors identify a structural problem: LLMs are often deployed without a named accountable actor, a traceable record of how outputs are generated, or a defined process for correction. While the technical literature focuses on measuring hallucinations and bias, the governance literature often lacks concrete metrics for oversight. This creates a "disciplinary disconnection" where technical capabilities outpace the mechanisms required for responsible use. The paper notes that accountability is frequently conflated with transparency or explainability, which the authors define as insufficient on their own—transparency without accountability is merely publication without consequence.

The LAAF Framework

To address these gaps, the authors developed the Layered Accountability Architecture Framework (LAAF). This framework organizes accountability into four layers:

  • Provenance: Tracking the origin and history of the model and data.

  • Application Logic: Managing the retrieval pipelines, prompts, and tools used to generate outputs.

  • Human Oversight: Defining who reviews system outputs, what evidence they use, and their authority to overrule the model.

  • Governance and Redress: Establishing the organizational policies and procedures for handling failures.
    These layers are cross-cut by three requirements: traceability, role clarity, and continuous monitoring. The framework also incorporates cybersecurity practices aligned with the OWASP LLM Top 10 (2025).

Regulatory Mapping and Findings

The authors conducted a systematic review of 122 primary studies and 12 regulatory documents, including the EU AI Act, the NIST AI RMF, and ISO/IEC 42001. Their analysis revealed four persistent gaps: the under-specification of human oversight, the absence of shared accountability metrics, a disconnection between technical and governance disciplines, and limited empirical evaluation of proposed mechanisms.
Franklin analysis: The paper emphasizes that LAAF is a synthesis of existing evidence rather than a validated, finished artifact. The authors explicitly state that their work is intended to provide a structured way to organize accountability across the LLM lifecycle, noting that because LLM behavior is dynamic—changing with model updates or new data—accountability must be treated as a continuous, lifecycle-wide property rather than a one-time inspection.

Why It Matters

This research is significant because it moves the conversation from abstract principles of "responsible AI" to a practical, layered model for deployment. By mapping these mechanisms to binding regulatory instruments like the EU AI Act, the authors provide a roadmap for organizations to align their LLM applications with legal and ethical standards. The framework highlights that because hallucinations are a systemic property of generative models, they must be governed through institutional processes rather than simply waiting for technical improvements to eliminate them.

Comments (0)

No comments yet

Be the first to share your thoughts!