Conceding Authorship

Palantir’s recent manifesto has gotten attention for obvious reasons. It flirts with the draft. It treats AI weapons as inevitable. It folds violent crime into the same sermon about national seriousness and civilizational renewal.[2]

None of that is the most interesting part.

The interesting part is that it is one of the few visible attempts to supply a doctrine for a technology that has already become infrastructural. Palantir has a project.[1] Its critics mostly have objections.

Some of those objections are correct. Many are necessary. But “this is dangerous” is not a doctrine. It is not a theory of public purpose, technical ambition, democratic constraint, institutional design, or legitimate use. That is the vacuum. Palantir did not create it. It noticed it was empty and walked in.

That is the political failure at the center of AI discourse. When critics refuse to govern what already exists, they do not preserve innocence; they concede authorship.

For years, too much AI criticism treated the technology as a moral contamination event rather than an infrastructural struggle. The argument was not merely that these systems were harmful, exploitative, deskilling, or politically dangerous. Much of that was, and remains, true. The deeper implication was that clear refusal could still function as policy. That the right stance toward an already-deploying technology might be opting out on principle. That if one remained morally clean enough, the world might pause in deference.

It didn’t.

Money and power continued doing what money and power does. Capital expenditure continued. Procurement continued. Defense integration continued. Administrative adoption continued. Platform substitution continued. Labor discipline continued. The state did not lose interest because critics found the technology spiritually degrading. Contractors did not abandon the field because artists, teachers, journalists, and ordinary workers correctly noticed that much of consumer AI was exploitative garbage.

Once AI became a labor issue, a state-capacity issue, a military issue, a procurement issue, and an administrative issue all at once, refusal stopped being governance. It became a way of removing yourself from the fight while other people wrote the operating doctrine.

That is what the Palantir moment reveals. Not merely that the company is creepy, though it is. Not merely that its worldview is ugly, though it is. But that it is occupying a space that should never have been left open in the first place.

The category error was treating “AI” as one policy object. It isn’t. It’s a stack of different governance problems that overlap without collapsing into one another.

Consumer harms are about use: fraud, scams, manipulative interfaces, synthetic sludge, privacy erosion, degraded services, the replacement of ordinary human contact with cheaper simulation.

Citizen harms are about standing: due process, eligibility decisions, policing, workplace discipline, administrative opacity; whether a person can be acted upon, categorized, denied, managed, or displaced without explanation, contestation, or appeal.

State harms are about authority: procurement capture, institutional deskilling, dependency on private infrastructure, the quiet replacement of public capacity with contractor-administered systems whose operators answer to capital before they answer to the public.

Geopolitical harms are about power distribution: compute concentration, military integration, intelligence fusion, alliance dependency, export controls, strategic asymmetry; who becomes sovereign over the technical substrate and who becomes dependent on someone else’s stack.

These are not the same threat models. They do not have the same stakes. They do not admit the same remedies. Yet a remarkable amount of AI discourse has treated them as though they could be melted down into one moral register.

One camp collapses everything into consumer disgust: chatbots are slop, AI is theft, synthetic media is corrosive, therefore the whole field should be morally quarantined. Another collapses everything into geopolitical destiny: strategic competition is real, therefore all hesitation is decadence, all friction is unseriousness, and all objections can be folded into a larger story of national revival.

Refusing AI maximalism without indulging unsummoning fantasies is not centrism. It is simply politics that survives contact with reality. Too much discourse has treated the field as a choice between acceleration and exorcism: either embrace contractor-led inevitability or retreat into a dream of rollback, rupture, or cleansing refusal. False binary. One can oppose private sovereignty over technical systems, demand hard constraints, public control, labor protections, procurement limits, and domain-specific bans, without pretending the infrastructure already built will disappear because the critique is morally pure.

Both camps flatten a multi-level problem into one moral register. That flattening is tidy. It can also be compelling, especially when life is already precarious. If AI shows up in your life as job erosion, arbitrary evaluation, surveillance, and replacement by synthetic junk, moral refusal has real force. If it shows up as institutional decline, adversary advantage, strategic drift, and elite unseriousness, a politics of urgent build-build-build has force too.

But force is not the same thing as governance. One side offers innocence. The other offers agency, or at least the feeling of it. Neither, on its own, is a doctrine.

That is why the Palantir manifesto matters. Not because it is persuasive, but because it is one of the few public documents willing to behave as if doctrine is required. That is why “these people are dangerous” feels so insufficient even when it is true. It remains an objection. The other side arrives with procurement pathways, institutional patrons, urgency, myth, and a vocabulary of sacrifice.

Even on its own terms, though, the doctrine is weaker than it looks. The loudest phrase in the package is some version of “AI deterrence,” which is exactly where the thing starts eating itself.

Deterrence requires being understood. It is not merely a matter of capability. It requires legible signaling, calibrated thresholds, credible resolve, and enough interpretive space for an adversary to understand what is being communicated. The whole logic depends on the other side reading you correctly.

AI power often arrives as being faster than you can be understood.

That is not a small bug. It is the load-bearing problem.

Its defenders would say the ambiguity is part of the point: deterrence no longer depends on shared legibility, but on shaping an environment so fast, fused, and unreadable that challengers hesitate inside it. But deterrence without a shared grammar is not stability. It is managed miscalculation with better branding.

The very properties that make AI attractive in military and state settings are structurally in tension with the features that make deterrence work.[3] Deterrence needs legibility. AI systems increase ambiguity. Deterrence needs calibrated signaling. AI systems compress action into machine-speed outputs that are difficult to interpret from the outside and often barely interpretable from the inside. Deterrence needs time. AI is sold on reducing the need for time.

So “AI deterrence” arrives wrapped in the prestige of nuclear-era strategic seriousness while eroding the communicative conditions that made deterrence coherent in the first place. It promises stronger capability while weakening legibility. It offers more warfighting potential and less strategic readability. It is, in that sense, a doctrine already eating its own face.

That contradiction does not stop at military affairs. It helps explain why so much AI governance discourse ends up trapped around language.

When action outruns interpretation, institutions start mistaking narration for control.

That is half the sickness right there. Real causality has moved downward into model weights, training data, scoring systems, retrieval paths, orchestration glue, thresholds, routing layers, and machine-speed recommendation loops. But those layers are hard for ordinary publics, and even ordinary institutions, to see directly. So governance migrates toward the last human-readable residue: prompts, policies, constitutions, refusals, system cards, acceptable-use statements, alignment rhetoric, model personality.

We start treating output style as governance. Refusal behavior as control. Policy prose as sovereignty. Safety theater as institutional restraint. The mouth becomes the object of political fixation because the hands have disappeared into the stack.

This is not a trivial confusion. It is what allows power to become opaque while legitimacy remains theatrically available at the interface. The public gets a narrated version of accountability while real authority migrates elsewhere.

A serious rival doctrine cannot be built out of vibes, refusal, or aesthetics. It has to be built out of boring things: standing, authority, receipts, appeal, admissibility, boundaries between advisory and constitutional power. It has to distinguish between systems that may recommend and systems that may decide; between systems that may surface a pattern and systems that may act on a person; between observability and legitimacy, visibility and standing.

That work is less emotionally satisfying than denunciation and less intoxicating than national myth. It lacks the glamour of both purity and destiny. But absent that work, the field defaults to contractor-state fusion dressed up as seriousness.

That is what the Palantir moment makes visible. Not that the company is uniquely monstrous. It is not. It is simply more explicit than most. It says aloud what many institutions have been drifting toward in practice: a world in which technical systems mediate labor, citizenship, administration, policing, military decision, and geopolitical competition through one increasingly seamless security story.

The answer cannot be a return fantasy. There is no path back to 1990, no matter how emotionally attractive the idea may be. The answer also cannot be to surrender the entire terrain of ambition, capability, and futurity to defense contractors and their mythographers.

What is needed is not less politics but more of it, stated plainly. Not “ethics” as a floating moral atmosphere, but politics in the older and less flattering sense: who governs, who benefits, who bears risk, who gets standing, who gets managed, what authority can be delegated, what must remain contestable, and what kinds of systems are simply incompatible with democratic life no matter how efficient they appear.

Palantir did not win because its doctrine is good. It won because it was one of the few actors willing to behave as if doctrine was required.

That should be embarrassing to more people than Palantir.

If critics refuse to govern what already exists, they do not preserve innocence; they concede authorship.

[1] Alexander C. Karp and Nicholas W. Zamiska, The Technological Republic: Hard Power, Soft Belief, and the Future of the West (official site). Useful for the book’s own framing: Silicon Valley’s drift away from national purpose, the call for renewed state-tech partnership, and the claim that AI has become part of a new strategic competition. (techrepublicbook.com)

[2] Alistair Barr, “Palantir’s summary of CEO Alexander Karp’s manifesto is generating buzz. Read the 22 bullet points,” Business Insider, April 20, 2026. Use this for the 22-point summary itself: national service / possible conscription, the “moral debt” language, the violent-crime point, and the “AI deterrence” framing. (Business Insider)

[3] Michael C. Horowitz, “Artificial Intelligence and the Future of Strategic Stability,” Texas National Security Review (2026). Use this for the claim that AI integration raises risks through machine miscalculation, AI-induced accidents, speed, and the erosion or bypassing of human firebreaks in escalation. (Texas National Security Review)