The Specific Event

Anthropic CEO Dario Amodei recently signaled openness to transferring AI technology to a government coalition, a development that has prompted significant discussion about corporate roles in AI governance. This is not a minor policy adjustment. A leading frontier AI laboratory is voluntarily proposing to hand core technological assets to a state apparatus. The framing in coverage has centered on "market dynamics" and "corporate roles," but that framing misses the more interesting structural problem embedded in the offer itself.

What the Offer Actually Reveals

When Amodei signals openness to government technology transfer, the implicit assumption is that competence transfers along with the asset. The government receives the technology, and by extension, the capability to govern it, deploy it responsibly, or at minimum, understand what it has received. This assumption is almost certainly wrong, and the error is a familiar one in organizational theory.

Kellogg, Valentine, and Christin (2020) documented a consistent pattern in algorithmically-mediated work environments: access to a system does not produce competence within that system. Workers with identical platform access show dramatically different outcomes because the structural logic of the platform is not self-evident from the interface. The same principle applies here. A government coalition that receives frontier AI technology does not automatically receive the interpretive schema required to govern it. What transfers is the artifact, not the literacy required to use the artifact well.

The Awareness-Capability Gap at the Institutional Level

Most AI governance discourse operates at the awareness level. Regulators demonstrate that they know AI systems exist, that they produce outputs, that bias is a concern, that safety matters. This awareness is real but limited. Awareness of a system's existence and awareness of its structural logic are categorically different forms of knowledge, and conflating them produces governance frameworks that are technically comprehensive but operationally inert.

Hatano and Inagaki (1986) drew a foundational distinction between routine expertise and adaptive expertise. Routine expertise is procedural: it works when conditions match the conditions under which the procedure was learned. Adaptive expertise involves understanding the principles generating the procedure, which enables response to novel conditions. A government agency trained to apply existing AI regulations - a checklist, a compliance audit, an impact assessment template - has routine expertise. A government agency that understands why large language models produce certain failure modes under certain distributional shifts has something closer to adaptive expertise. The technology transfer Amodei describes would deliver an artifact to institutions that, structurally, have been built to develop the first kind and not the second.

The Transfer Puzzle Applied to Institutional Contexts

My dissertation research on the Algorithmic Literacy Coordination framework is concerned with a specific puzzle: why do platform workers with identical access show divergent outcomes? The ALC framework proposes that competencies develop endogenously through participation in algorithmically-mediated environments, not through prior training or baseline ability alone. The implication is that coordination around algorithmic systems requires a form of schema induction - teaching structural features, not just operational procedures.

Gentner's (1983) structure-mapping theory is instructive here. Transfer between domains occurs when a learner maps relational structure from a familiar domain onto an unfamiliar one. The difficulty for government actors receiving frontier AI technology is that there is no familiar domain from which to draw analogous relational structure. Nuclear technology transfer had decades of institutional precedent. Pharmaceutical regulation had established causal models. Frontier large language models have neither, which means the structural schema required for governance has to be constructed rather than transferred, and construction takes time, iteration, and direct engagement with failure cases that government institutions are poorly positioned to generate.

What This Means for the Governance Conversation

The problem with Amodei's offer is not the intention behind it. The problem is that the governance conversation has not yet developed an account of what institutional competence in AI governance actually requires. Rahman (2021) described how algorithmic systems create what he called invisible cages, structural constraints that shape behavior without being legible to the workers inside them. Governance institutions face the same legibility problem, but with higher stakes and less tolerance for iterative learning.

If the ALC framework's counterintuitive prediction holds - that schema induction produces better transfer than procedural training - then the design of any government AI competence-building program matters enormously. Training government officials to follow compliance checklists is procedural. Building institutional capacity to reason about distributional shift, emergent capability thresholds, and model behavior under adversarial conditions is structural. The former is achievable quickly and will produce reports that satisfy oversight committees. The latter is what governance actually requires, and it is not something that arrives with the technology transfer.

Amodei's offer is worth taking seriously. The framework for accepting it is not yet in place.