The Announcement and What It Actually Says

The Pentagon has announced it is expanding its GenAI.mil platform to give approximately three million military and civilian workers access to ChatGPT and Grok, framing the rollout as a push to meet "warfighter needs." This is not a pilot program or a limited proof of concept. It is an organizational deployment at a scale that has no real precedent in the history of enterprise AI adoption. The security architecture is interesting. The coordination problem underneath it is more interesting.

Most of the coverage I have seen treats this as a story about AI capability, clearance levels, and whether Grok belongs anywhere near a Defense Department server. Those are legitimate concerns. But they distract from a structural question that the announcement leaves entirely unanswered: what happens when three million workers with radically different prior schemas are given simultaneous access to a system that produces outputs whose quality depends almost entirely on how you prompt it?

Identical Access, Unequal Outcomes

This is precisely the variance puzzle at the center of my dissertation research. The Algorithmic Literacy Coordination framework predicts that platform workers with identical access will produce dramatically different outcomes, not because of natural ability differences, but because of differences in the structural schemas they bring to the interaction. Kellogg, Valentine, and Christin (2020) document this pattern in algorithmic work environments: access equality does not produce outcome equality, because the relevant competency is not access but the ability to navigate a system whose logic is not transparent.

The Pentagon deployment assumes that rolling out a tool is the same as deploying a capability. These are not the same thing. A logistics analyst at a regional command who has never articulated a structured query and a defense contractor with two years of prompt engineering experience are not receiving the same tool, even if the interface is identical. They are receiving the same surface with entirely different underlying affordances, because affordances are not properties of the platform alone. They are relational properties that emerge from the interaction between platform structure and user schema (Sundar, 2020).

The Awareness-Capability Gap at Institutional Scale

What makes this deployment worth analyzing carefully is that the Defense Department will almost certainly run some form of AI literacy training alongside it. This is what large institutions do. They announce access and then schedule a training module. The problem, which Gagarin, Naab, and Grub (2024) document in the media literacy context, is that awareness training does not close the capability gap. Knowing that an LLM can hallucinate does not tell you how to construct a prompt that reduces hallucination probability. Knowing that output quality depends on input structure does not give you the structural schema required to act on that knowledge.

This is Hatano and Inagaki's (1986) distinction between routine and adaptive expertise, applied at an institutional level. Routine expertise, the kind most onboarding programs produce, equips workers to use a tool in the conditions the training anticipated. Adaptive expertise equips workers to respond when conditions shift, when the model updates, when a new capability is added, or when a task type falls outside the training examples. The second type is substantially harder to produce and almost never what institutional rollouts actually deliver.

What the "Warfighter Needs" Framing Obscures

The framing of this deployment around warfighter needs is worth examining directly. It positions the rollout as a capability problem that has been solved by providing access to powerful tools. Rahman (2021) describes this as a structural feature of platform governance more broadly: the platform presents itself as an enabler while the actual labor of competence development is externalized to the worker. In the Pentagon context, that externalization has institutional consequences. If a civilian analyst generates a flawed intelligence summary because their prompt schema was inadequate, the failure will be attributed to the individual, not to the organization's decision to deploy a system without addressing the coordination infrastructure required to use it well.

Schor et al. (2020) make a similar argument about platform dependence: the asymmetry is not just economic, it is epistemic. Workers are dependent on systems whose evaluation criteria they do not fully understand. At three million users, that epistemic asymmetry does not disappear because the employer is the federal government. It scales.

The Structural Question Worth Asking

The story here is not whether ChatGPT is secure enough for military use. It probably is, given the architecture described. The story is whether the organization deploying it has any theory of how competence develops in algorithmically-mediated environments, or whether it is simply assuming that access produces capability. Based on how the rollout has been described publicly, I see no evidence of the former. That is the coordination problem. It will not show up in the security audit. It will show up in the variance of outcomes across three million workers, and most of those outcomes will be invisible to the institution that produced them.

References

Gagarin, D., Naab, T. K., & Grub, J. (2024). Algorithmic media use and algorithm literacy. New Media & Society.

Hatano, G., & Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, & K. Hakuta (Eds.), Child development and education in Japan (pp. 262-272). Freeman.

Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, K. S. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

Schor, J. B., Attwood-Charles, W., Cansoy, M., Ladegaard, I., & Wengronowitz, R. (2020). Dependence and precarity in the platform economy. Theory and Society, 49(5-6), 833-861.

Sundar, S. S. (2020). Rise of machine agency: A framework for studying the psychology of human-AI interaction. Journal of Computer-Mediated Communication, 25(1), 74-88.