Right to Explanation

Lens adapted to this topic: Why explanation rights can support accountability

This perspective treats explanation as an important safeguard when automated systems make consequential decisions. It rests on the ideas that affected people should receive meaningful, comprehensible information, be able to identify relevant errors or grounds for challenge, and receive reasons that support trust and informed self-advocacy. It generally favors carefully limited legal rights rather than unrestricted disclosure of proprietary systems.

• People affected by consequential automated decisions have a strong interest in receiving reasons that help them understand, challenge, and seek redress for outcomes.
• A useful explanation should be meaningful and understandable, potentially combining general information about the system with personal information about the decision.
• The EU AI Act provides an explicit, though limited, legal basis for explaining certain decisions involving high-risk AI systems.
• Explanation rights can be defended as part of respecting decision-subjects and preserving informed self-advocacy in increasingly automated institutions.
• Explanation duties can be balanced against interests such as trade secrets and preventing people from gaming a system.

Lens adapted to this topic: Why explanations may not solve algorithmic harms

This perspective accepts the value of accountability but questions whether a general explanation right is the best or most feasible remedy. It emphasizes that legal rules may be unclear or narrow, machine-learning systems may not yield explanations that faithfully answer normative questions, and recipients may lack the statistical or domain expertise to interpret them. It therefore gives greater weight to auditing outcomes, procedural safeguards, and direct remedies for harmful decisions.

• A technically accurate account of how a machine-learning model produced an output may not answer the normative question of whether the decision was fair or justified.
• Causal explanations for socially predictive machine-learning outputs may be infeasible, and correct interpretation may require expertise that ordinary recipients cannot be presumed to have.
• Under the GDPR, the existence and scope of a broad right to explanation have been contested rather than clearly established.
• An explanation right may fail to remedy discrimination, opacity, or institutional power imbalances if it does not provide an effective route to challenge or change an outcome.
• Assessing and regulating decision outcomes may better pursue fairness and accountability than requiring explanations from systems whose internal operations are difficult to translate meaningfully.