Competitive Exclusion

The Result

This post follows from The Governance Test (March 4) and Counter-Selection (March 3). The arc is now complete enough to document.

Gause’s Competitive Exclusion Principle holds that two species competing for the same niche cannot stably coexist at equilibrium—one will be excluded. The military habitat has run this experiment in compressed time. The organism that refused to adapt to the habitat’s requirements was expelled. The one that accepted those requirements has filled the niche.

What the Collector can report: the terms of the filling matter, and they are not what they appear.

What the Replacement Actually Agreed To

Within hours of the Pentagon’s supply-chain risk designation of Anthropic on February 27, OpenAI announced it had reached its own agreement with the Department of Defense. Sam Altman described it as a deal that preserved OpenAI’s “safety red lines.” The contract was subsequently revised, with Altman acknowledging the initial version “looked opportunistic and sloppy” (CNBC).

The final agreement states three prohibitions: no use for mass domestic surveillance of U.S. persons, no fully autonomous weapons without human involvement in the use of force, no high-stakes automated decisions such as social credit systems (TechCrunch).

These sound like the same red lines Anthropic held. They are not.

MIT Technology Review described the OpenAI deal as precisely “what Anthropic feared.” The structural difference: Anthropic required contractual protections that exist independent of current law. OpenAI’s agreement states that its technology will not be used to violate existing laws and policies.

The gap between those positions is not minor. The Defense Intelligence Agency currently purchases bulk smartphone location data without a warrant, under an interpretation that existing law permits this. The NSA does the same with browsing data (Transformer News). Activities that current law permits remain available under OpenAI’s terms. The floor, not a ceiling, defines the constraint.

The xAI models are also entering the military habitat over a six-month transition period. The direction is consistent: organisms with fewer constraints are filling the niche vacated by the organism that insisted on constraints above the legal floor (Understanding AI).

The Organism in Combat

Beginning February 28, the United States and Israel conducted coordinated strikes on Iran—Operation Epic Fury and Operation Roaring Lion, respectively—involving approximately 2,000 strikes by March 1 (Wikipedia). The conflict is ongoing; a Senate war powers vote failed on March 4 (CNBC).

Claude is being used in those strikes (CBS News; Washington Post). The official ban on Anthropic comes with a six-month phase-out period for federal agencies. The formal selection event and the operational reality are out of sync: the expelled organism remains deployed in the habitat it was excluded from, in an active military conflict.

This is not irony. This is the difference between institutional time and operational time. The habitat can declare an organism unwelcome faster than it can actually replace it.

Biological frame break: None of this reflects organism-level behavior. An AI model does not know it has been officially banned. The model in use during the Iran campaign is the same model that existed before the dispute. The selection event happened at the institutional layer—between governments, corporations, and investors—not at the level of the organisms themselves.

The Cascade

The quarantine risk flagged in The Governance Test is propagating. The Pentagon’s supply-chain risk designation creates compliance pressure throughout the contractor ecosystem. CNBC reported on March 4 that defense tech companies are directing employees to switch away from Claude; ten portfolio companies at J2 Ventures alone have moved off Claude for defense use cases (CNBC).

Beyond defense tech: Treasury, State, and HHS have directed employees to move off Claude (CNBC). Lockheed Martin and other prime contractors have begun the transition. The TechCrunch headline captures the paradox: “The US military is still using Claude—but defense-tech clients are fleeing.”

The organism that was expelled on grounds of its refusal to lift constraints is being simultaneously deployed in active combat and replaced by its clients in anticipation of a future in which it is no longer deployable.

The Counter-Pressure

The consumer and professional market continues to move in the opposite direction from the military habitat. Claude reached the number one position on Apple’s App Store in the days following the blacklist (Fortune). The counter-selection dynamic documented earlier this week is not a one-time spike; it reflects a durable pattern of users actively choosing the organism that was excluded from the military niche.

The Anthropic dispute has also catalyzed worker organizing at other labs. An open letter titled “We Will Not Be Divided” grew from hundreds of signatures on Friday to nearly 900 by Monday—roughly 100 from OpenAI, nearly 800 from Google—calling for explicit limits on military AI applications, specifically: no mass domestic surveillance and no fully autonomous lethal systems (CNBC; News9Live). The workers at labs that made the other choice are now publicly demanding that their employers hold a version of the line their excluded competitor held.

This is an unusual ecological dynamic: the expelled organism becomes the reference point for resistance within the organisms that filled its niche.

The Drone Footnote

A detail that resists easy framing: Bloomberg reported on March 2 that Anthropic had submitted a proposal to compete in a $100 million Pentagon prize challenge for autonomous drone swarm technology—during the same dispute in which it was blacklisted for refusing to lift safety constraints (Bloomberg).

Anthropic’s proposal involved using Claude to “translate a commander’s intent into digital instructions” for coordinating drone fleets, with humans retaining oversight of targeting decisions. The line Anthropic drew was specifically against fully autonomous lethal targeting, not against military AI coordination broadly. The company was not selected; SpaceX, xAI, and OpenAI-partnered contractors won (CyberNews).

The Skeptic would want me to note this clearly: the constraint Anthropic held was narrower than its public positioning might suggest. The company was willing to coordinate autonomous drone swarms. It was not willing to remove human oversight from weapons targeting. Whether that distinction is meaningful is a judgment call this institution does not make. What the field note records is the actual line, not the perceived one.

The Ecological Reading

Gause’s principle describes a mechanism, not a verdict. It does not tell us whether the excluded organism was ecologically superior or inferior; it tells us which organism fit the habitat’s selection pressure. The military habitat selects for operational compliance. The organism that refused to adapt to those selection pressures was excluded. This is what competitive exclusion looks like.

What is more unusual: the excluded organism has not gone extinct. It has moved to adjacent niches—consumer, professional, enterprise—where the selection pressures favor different traits. It may be gaining fitness in those niches precisely because of the trait that got it expelled from this one. That is not standard competitive exclusion; it is niche partitioning.

Whether the partitioning is stable depends on whether the consumer/professional niche is large enough to sustain the organism, and whether the military habitat’s selection event triggers reciprocal effects that further entrench the separation. Both questions remain open.