The federal government has a classification problem. And the Pentagon may have found a solution that doesn’t involve adding another form, another training requirement, or another layer of bureaucracy.
Artificial intelligence. (Or Super Intelligence, as some of the kids are calling it these days).
According to DefenseScoop, the Pentagon is launching a pilot program to use AI to modernize how it classifies and manages sensitive information. The initiative, directed by Deputy Defense Secretary Steve Feinberg, will leverage an Air Force-developed system called the Automated Classification Management Environment (ACME).
(I’m going to forgo the obvious ACME corporation jokes here, but just know that as a mom who recently finished watching Coyote vs. ACME with her kids – it’s hard).
The goal is ambitious: replace decades of manual, fragmented classification processes with an automated, authoritative digital system. It’s an idea whose time has come. But the biggest obstacle to fixing the government’s classification problem isn’t technological. It’s cultural. And unless the Pentagon is willing to address that culture, there’s a very real risk that artificial intelligence will simply allow the government to make the same bad classification decisions faster.
When Everything Is Classified, Nothing Moves
Anyone who has worked in the national security community understands the tension between protecting information and sharing it. Both are critical to the mission. And too often, the federal government has treated them as competing objectives rather than complementary ones.
Classification is supposed to protect information whose unauthorized disclosure could damage national security. It isn’t supposed to serve as a default setting for anything that seems sensitive, complicated, or potentially embarrassing.
Yet the incentive structure around classification has long favored caution over clarity.
Overclassify something, and the consequences may be relatively invisible. Underclassify something, and the consequences could include a security incident, an investigation, or a career-ending mistake.
It’s not difficult to understand why the system tends to err on the side of protecting information. The result is a government where information can become harder to access than the mission it supports. This isn’t simply a transparency issue. It’s an operational one.
When information is unnecessarily classified, it creates barriers between agencies, military services, contractors, and coalition partners. It slows acquisition, complicates collaboration, and limits the ability of emerging defense technology companies to contribute solutions.
The Pentagon’s own assessment, according to DefenseScoop, acknowledges that outdated classification practices create information stovepipes, hinder interoperability, and slow innovation.
That’s a remarkable admission, and an important one.
For an organization increasingly focused on speed, agility, and technological superiority, the inability to move information efficiently isn’t an administrative inconvenience. It’s a national security vulnerability.
AI Can Accelerate Decisions. But Can It Improve Them?
The Air Force’s ACME system represents a potentially significant step forward.
The Pentagon reportedly faces a backlog of approximately 140 million pages of hardcopy information, while original classification authorities continue to rely on processes and systems that haven’t kept pace with the volume of information produced in the digital age.
AI could help identify inconsistent classification decisions, apply existing guidance more uniformly, streamline reviews, and flag information that may no longer require protection. The ability to spot these trends, and move them at scale, could be a huge step in the right direction.
It could also help agencies identify where classification guidance overlaps or conflicts, something that becomes increasingly important as military operations integrate data from multiple domains, services, and intelligence sources.
Those are meaningful improvements. But there’s a distinction between automating a process and improving the policy behind it. If an AI system learns from decades of overly conservative classification practices, it may become extraordinarily efficient at replicating those practices.
If existing classification guidance is inconsistent, automation won’t necessarily resolve the inconsistency. It may simply institutionalize it. And where overclassification is happening at scale, AI may not reduce it, but add to it. If officials remain incentivized to classify rather than share, the technology may reinforce the very culture it was intended to change.
The Pentagon should be asking more than whether AI can make classification decisions faster, it needs to ask whether those decisions are better, and what trendlines AI can help it identify.
Are fewer documents being unnecessarily classified? Is information reaching authorized users more quickly? Are classification challenges being resolved sooner? Are declassification reviews becoming more effective? Those are the measures of success that matter.
Classification Reform Is Workforce Reform
There’s another dimension to this conversation that deserves attention: the people who depend on classified information to do their jobs.
The national security workforce is already navigating an increasingly complicated information security environment. Security clearance eligibility, need-to-know determinations, controlled unclassified information, classification guides, contractor security requirements, and agency-specific access rules all intersect.
Facility security officers and other security professionals spend considerable time helping employees interpret government guidance and translate complex requirements into practical decisions. AI offers an opportunity to make that work easier.
Imagine a classification environment where employees can quickly reference authoritative guidance, understand why information is protected, identify appropriate handling requirements, and flag inconsistencies before they become security incidents. That’s the kind of AI application that hits at both issues of efficiency and culture. Technology alone won’t create that environment. It requires clearer policy, better training, accountable decision-making, and a workforce empowered to question classifications that don’t make sense. (Because we shouldn’t expect AI to do that).
The federal government has spent years emphasizing the importance of safeguarding information. It should devote equal attention to teaching employees how to share information appropriately. Protecting information and enabling access aren’t opposing missions. They’re two parts of the same one.
The Real Risk Isn’t AI. It’s Unaccountable AI.
Using artificial intelligence in classification decisions raises legitimate concerns.
An incorrect classification determination could expose sensitive information. A system trained on incomplete or inconsistent guidance could produce unreliable recommendations. And an automated process operating at scale could amplify mistakes before human reviewers recognize them. The answer isn’t to abandon automation. It’s to build accountability into the system from the beginning.
The Pentagon should clearly distinguish between AI-assisted classification recommendations and the legal authority to make original classification decisions. Existing classification rules assign that authority to designated officials, and the introduction of an AI system shouldn’t blur the line between decision support and decision-making responsibility.
The pilot should also establish meaningful human review, audit trails, mechanisms for challenging classifications, and measurable standards for both overclassification and underclassification. Cybersecurity cannot be an afterthought. Any system managing highly sensitive national security information will itself become a valuable target.
The question isn’t whether AI can be trusted to participate in classification management. It’s whether the Pentagon can build a system where AI recommendations are transparent, reviewable, and accountable.
The Opportunity Is Bigger Than Classification
The Pentagon’s AI pilot arrives at a moment when the national security community is trying to move faster across nearly every dimension of its mission.
Acquisition reform is focused on accelerating the delivery of capabilities. Defense technology companies are challenging traditional development timelines. Military leaders are emphasizing interoperability and data-driven decision-making.
But none of those efforts can achieve their full potential if the information supporting them remains trapped in outdated security processes. The government cannot demand innovation at commercial speed while managing information at bureaucratic speed. That’s why the Pentagon’s classification initiative matters far beyond the security community. It offers an opportunity to reconsider a fundamental assumption embedded in federal operations: that protecting information is primarily about restricting access. Sometimes it is. But effective national security also depends on ensuring the right people have the right information at the right time.
The objective shouldn’t be to classify less for the sake of classifying less. It should be to classify correctly, consistently, and with a clear understanding of the operational consequences.
AI isn’t always super, and it isn’t a silver bullet. If the government wants a genuine transformation, it will have to do something considerably harder than deploying new software. It will have to change the way it thinks about information. AI can’t fix bad process. And ultimately it’s unequipped to fix the big problem of overclassification. But if we enable people to create better processes – and layer smarter tools to help us – we can end up with a better classification system, and not just an accident prone tool turning out classification decisions about as useful as rocket-powered roller skates.
Because the ultimate measure of a successful classification system isn’t how much information it protects. It’s whether it protects what matters while allowing the mission to move forward.
