The normative gap problem: accountability when the standard that should govern the decision is itself contested

Accountability requires a norm. To evaluate whether an AI agent acted appropriately, a reviewer must have access to a standard — some specification of what the agent should have done in the circumstances it faced. The norm might be explicit: a protocol, a regulation, a documented threshold. It might be implicit: a professional standard, a community practice, a cultural expectation. In either case, the evaluation asks whether what the agent did was consistent with what it should have done, and that question has no answer without a referent.

The normative gap is the condition where the standard that should govern an AI agent's decision is not settled. Not unknown — it may be actively contested, with legitimate authorities holding different and incompatible views about the appropriate norm. Or it may be genuinely absent — a domain where practice is newer than the norms that should govern it, and where no settled expectation has yet formed. In both cases, accountability cannot function as designed: the mechanism for evaluation requires a baseline that does not yet exist or is not agreed upon.

At the post-quantum security crossing

The appropriate pace of migration to post-quantum cryptographic standards is genuinely contested. Technical arguments support a range of positions: some cryptographers argue for aggressive migration before quantum capabilities emerge; others argue that premature adoption of less-mature standards introduces its own vulnerabilities; still others contend that the migration window itself creates a period of cryptographic uncertainty that is a risk in its own right. An AI agent managing cryptographic infrastructure must make decisions in this contested space — setting migration timelines, selecting algorithms, scheduling key ceremonies — and the standard against which its decisions will be evaluated is not settled.

When accountability is applied to those decisions retroactively, reviewers may apply norms that were not established at the time of the decision, or that reflect a particular position in a technical dispute that remained unresolved when the agent acted. An agent that followed a conservative migration schedule may be evaluated against a later consensus that faster migration was warranted. The evaluation is formally possible — the decision log exists, the outcome is observable — but the normative baseline is supplied retrospectively, and the choice of baseline reflects choices the accountability architecture was never designed to make explicitly.

At the hardware crossing

Norms for AI-assisted physical rehabilitation are established faster in some dimensions than others. Endpoint protocols, safety thresholds, and fall-risk criteria have accumulating clinical evidence behind them. Norms for how much decision-making authority an AI agent should exercise within a therapy session — how often to override the initial protocol, how aggressively to adapt to real-time biometric signals — are more contested. Clinical traditions differ: what one physical therapy community treats as appropriate responsiveness, another treats as deviation from evidence-based protocol.

An AI agent making real-time adaptation decisions operates inside this contested space. Its decisions can be logged; whether they were appropriate cannot be evaluated without first resolving which clinical tradition establishes the referent norm. When institutions or regulatory bodies apply accountability processes, they may apply norms from their own tradition without recognizing that the norm itself was under contest at the time the agent acted. The agent's decisions may be evaluated as violations of a standard that had no settled status when the decisions were made.

At the physical-world care crossing

In care settings, norms for appropriate AI involvement in care decisions vary across cultures, professional traditions, and family expectations in ways that have not been resolved. What counts as appropriate reliance on an AI agent for dietary recommendation in one care community is treated as inappropriate delegation of clinical judgment in another. What is considered respectful of a resident's autonomy in one cultural context is treated as inadequate advocacy in another. These differences are not resolvable by technical means — they reflect genuine normative disagreement about what care should accomplish and who should bear accountability for the decisions within it.

An AI agent operating across multiple care settings will inevitably have its decisions evaluated against different normative baselines depending on which institution, regulator, or cultural framework conducts the review. The accountability gap is not in the agent's decision record or in the evaluator's diligence — it is in the absence of a settled norm that all parties recognize as the appropriate referent.

Making the normative assumption explicit

The structural response to the normative gap is not to pretend the norm is settled when it is not. It is to make the normative assumption explicit and place it under authorization. Every AI agent operating in a contested domain embeds a normative assumption: it operationalizes some version of the contested norm in its decision parameters. That assumption should be visible, documented, and subject to the same authorization processes as the agent's actions.

This does not resolve the underlying normative disagreement. It relocates accountability from the agent's runtime decisions to the normative assumption embedded in its design — which is where accountability belongs when the norm is contested. The evaluative question shifts from 'was this action consistent with the norm?' to 'was this normative assumption appropriately authorized, disclosed, and monitored for drift?' That question can be answered from available evidence. The original question cannot, because the norm it depends on was never settled. At the crossings where practice outpaces consensus, making the embedded assumption visible is the only accountability move that does not pretend to close a loop that remains open.