Depth Governance

In the AI Story

The governance problem is structural: the quality dimension that matters most (whether the output serves genuine purpose, whether the analysis captures relevant causality, whether the code will scale under realistic conditions) is precisely the dimension the surface conceals. An AI-generated statistical analysis may exhibit perfect formatting, appropriate citations, and sophisticated vocabulary while resting on a fundamentally flawed causal model. A Claude-written function may compile cleanly, pass surface tests, and fail catastrophically under edge conditions no one thought to specify. An essay may demonstrate apparent understanding of a philosophical position while misrepresenting its actual content—The Orange Pill's Deleuze fabrication being the paradigmatic case. In each instance, surface inspection provides false confidence, and only depth evaluation—reading the analysis against the data, tracing the code's logic under load, checking the philosophical reference against the original source—reveals the gap between appearance and reality.

Implementing depth governance requires four institutional commitments. First, evaluation expertise: the organization must retain people with sufficient domain knowledge to assess judgment quality, not merely output formatting. This is expensive—domain experts are scarce and costly, and their time spent evaluating is time not spent producing. Second, protected evaluation time: depth governance cannot be performed in the gaps between production tasks. It requires sustained, focused attention that the continuous partial attention regime of AI-augmented work systematically eliminates. Organizations must build temporal structures—scheduled review periods, mandatory pauses, evaluation sprints—that protect the time required. Third, process transparency: evaluators must have access not merely to outputs but to the process generating them—the prompts used, the iterations performed, the alternatives considered and rejected. This requires documentation overhead that slows production. Fourth, explanation culture: the organization must normalize the demand for explanation—'Why did you accept this output? What did you verify?'—transforming it from an exceptional audit into a routine expectation.

The economic logic is Williamsonian: depth governance is a credible commitment to quality because it is expensive. An organization that announces commitment to judgment quality while continuing to measure and reward output volume, that praises careful evaluation while promoting based on shipping speed, is engaged in cheap talk—costless utterances contradicted by resource allocation. The workers correctly infer that the actual priority is speed, and they optimize accordingly: accepting AI output that looks right, because the organization's incentive structure rewards the appearance of productivity over the reality of judgment. Depth governance aligns incentives by making evaluation observably costly: the firm sacrifices short-term output for long-term capability, and the sacrifice is what makes the commitment credible. Workers observing this trade-off receive a reliable signal that the organization genuinely values judgment—and they respond by investing in developing it.

Origin

The term is new to this volume, synthesizing Williamson's governance framework with the specific challenge AI's smooth outputs create. The underlying problem—monitoring quality when surface inspection is uninformative—is classical in economics of information (Akerlof's lemons problem, Spence's signaling) and in organizational theory (performance evaluation under information asymmetry). But AI has made the problem acute by systematically producing outputs whose surface quality is decorrelated from their depth quality. The Deleuze error, the COMPAS case flaws, the legal brief fabrications, the integration leaks—each represents a category of failure that surface governance cannot catch and that demands institutional mechanisms designed specifically for depth evaluation. The concept extends Williamson's monitoring framework into the novel informational environment AI has created.

Key Ideas

Surface governance fails under AI. Traditional quality mechanisms checking syntax, formatting, and specification compliance cannot detect errors concealed beneath smooth presentations.

Depth governance evaluates process. Not whether the output looks right, but whether the judgment producing it was sound—a shift from product to process evaluation.

Expertise is the prerequisite. Depth evaluation requires domain knowledge sufficient to assess judgment quality, making it expensive and scarce.

Cost signals commitment. Depth governance is credible precisely because it sacrifices productivity—the expense distinguishes genuine commitment from cheap talk.

Explanation must be normalized. The demand 'Why did you accept this?' transforms from audit to routine, from accusation to standard practice, enabling governance without destroying morale.