Posted on by

Establishing accountability

In an increasingly AI-driven world, the governance of these powerful systems has become a pressing concern. At the heart of responsible AI lies accountability, the principle that those who design, develop, and deploy AI should be answerable for its actions and impact. Especially within the European Union, this concept is viewed as a cornerstone for ensuring fairness, ethical alignment, and harm mitigation in AI systems. Yet, despite its significance, accountability in AI is often vaguely defined, hindering effective policy-making and public discourse. This vagueness masks the trade-offs inherent in different approaches to AI governance.

Novelli et al examined accountability in AI in 2024. They began by defining accountability as a “relation of answerability,” emphasising the obligation to justify one’s conduct to a recognised authority. This relation relies on three fundamental conditions: authority recognition (acknowledging the legitimacy of the forum holding the agent accountable), interrogation (the forum’s right to scrutinise the agent’s conduct), and limitation of power (the forum’s ability to constrain the agent’s actions based on its evaluation).

Building on this foundation, the authors present an “architecture” of accountability, identifying seven key features that shape the accountability relationship:

  1. Context: the “what for,” specifying the AI’s purpose and field of use.
  2. Range: the “about what,” defining the scope of tasks and responsibilities subject to accountability, across design, development, or deployment phases.
  3. Agent: the “who,” identifying the individual, corporate, or collective entity accountable for AI actions.
  4. Forum: the “to whom,” specifying the entity to which accountability is owed, which may be the public, data subjects, or regulators.
  5. Standards: the “according to what,” encompassing legal rules, ethical principles, and technological requirements against which AI behaviour is assessed. (
  6. 6) Process: the “how,” outlining the rules, metrics, and procedures used to assess adherence to standards, potentially involving internal supervision, external audits, or human-machine interaction.
  7. Implications: the “what follows,” specifying the potential consequences of the accountability assessment, ranging from sanctions to revisions.

These features are then organised around four overarching goals that accountability seeks to serve within a governance framework:

  1. Compliance: ensuring AI systems adhere to ethical, legal, and technical standards, often implemented through preliminary checks.
  2. Report: facilitating explanation and justification of AI behaviour, enabling a dialogue between the AI system, its operators, and affected parties. Here, the authors explain that simple transparency may not be functional, and better alternatives may be in interpretable AI that describes internals in a user-understandable way or explainable AI that reports only what’s needed to interact.
  3. Oversight: actively seeking information, gathering evidence, and evaluating AI performance, performed by different internal or external bodies.
  4. Enforcement: linking the monitoring and evaluation of AI to consequences, from fines and sanctions to system revisions. The authors argue that while these goals can be complementary, policymakers often prioritise certain goals, implicitly shaping the direction of AI governance.

The article emphasises a sociotechnical approach to accountability in AI. Acknowledging that AI systems are not isolated technologies but are embedded within complex social and organizational contexts, the authors argue that accountability must consider the interplay between technological capabilities, human roles, and existing values. This approach helps unpack three key features of accountability: the range (influences of values and incentives), agents (interactions among technology, humans, and the environment), and standards (normative and epistemic reasons for conduct). The article’s shift to a sociotechnical approach emphasises the relevance of an external authority, and rejects accountability as a purely internal reflection, as it would be if it were purely moral and didn’t carry authority recognition, interrogation, and limitation of power.

The article concludes by acknowledging the complexities of addressing accountability in AI, underscoring the importance of ethical, legal, and political deliberation in striking a balance between different accountability policies, and their impact on individuals.

Further reading

Novelli, C., Taddeo, M. & Floridi, L. Accountability in artificial intelligence: what it is and how it works. AI & Soc 39, 1871–1882 (2024).
https://doi.org/10.1007/s00146-023-01635-y.

s00146-023-01635-y

How will you ensure that the use of AI is accountable in your organisation?
Record your thoughts in your learning diary before you move on….

< Previous Page
Next Page >