Transparency is one of the most widely endorsed principles in AI ethics. The EU AI Act, the OECD AI Principles, the UNESCO Recommendation, ISO 42001, all include transparency as a core requirement. The specific technical manifestation most frequently demanded is explainability: the ability of an AI system to provide a human-understandable account of how it reached a decision.

Postmodern theory does not reject transparency. But it subjects the concept to scrutiny that reveals its limits, and the ideological functions it can serve.

What explainability actually delivers. Current explainability techniques, LIME, SHAP, attention visualisation, provide post-hoc approximations of how a model produced a given output. They are not windows into the model’s actual decision process. A complex neural network does not reason in terms that can be straightforwardly summarised in human language. The explanation is a reconstruction, a simplified narrative that may bear only a loose relationship to the actual computational process.

This matters because the legitimating function of explainability depends on the explanation being accurate. If the explanation tells a person that their loan was denied because of their income level and credit history, but the actual model also weighted heavily a factor that correlates with their race, which the explanation does not reveal, then the explanation is not merely incomplete. It is actively misleading.

Transparency as performance. Postmodern analysis attends to the performative dimension of transparency demands, the ways in which compliance with transparency requirements can produce the appearance of accountability without its substance. A company that publishes a model card for its AI system, documents its training data, and provides an API for algorithmic auditing has complied with many transparency requirements. If the model card omits material information, the training data documentation is incomplete, and the audit API reveals less than it conceals, the compliance is performative rather than substantive.

This is not merely cynicism. It reflects a general principle: the form and content of transparency requirements are shaped by the interests of those who set them. Transparency requirements that are technically demanding but substantively limited serve the interests of companies with the resources to comply at scale while managing what is revealed.

The right to explanation and its limits. The EU’s General Data Protection Regulation includes a right to explanation for automated decisions. Postmodern analysis notes that even a fully accurate explanation of how an algorithmic decision was made does not settle whether the decision was just. Knowing that your loan was denied because your income is below a threshold, your credit history is shorter than a benchmark, and your postal code falls in a high-risk category tells you how the decision was made. It does not tell you whether the threshold is fair, whether the benchmark was set appropriately, whether the risk category is valid, or whether the decision should stand.

The question of justice cannot be resolved by the question of explanation. Transparency is necessary but insufficient for accountability. What is additionally required is the power to contest, to challenge not just whether the model accurately describes you, but whether the standards it applies are legitimate.

Reflection question: Think about a transparency mechanism you have encountered in an AI context, an explanation, a model card, an audit. What did it reveal? What did it conceal? Who was it designed for?