When Power Fights the Tools It Depends On

Jul 26, 2026

9 min read

1 views

22%

The strange fight at the center of modern power

What happens when the people who want the most advanced tools also demand the right to use them in ways the makers consider dangerous?

That is no longer a hypothetical. It is the core conflict of the AI era, and it exposes a deeper truth about power: institutions do not just want capability, they want capability without constraint. They want tools that are powerful enough to extend their reach, but obedient enough to never block their ambitions. The moment a tool refuses that role, it stops being a neutral product and becomes a political problem.

That is why this dispute matters far beyond one contract, one company, or one military procurement fight. It is a live test of a new question: who gets to define the limits of intelligence when intelligence becomes infrastructure?

The real argument is not about AI. It is about veto power.

On the surface, the conflict looks like a familiar business dispute. One side wants a contract. The other side wants different terms. But underneath that is something much more consequential. The Pentagon is not merely asking for software. It is asking for operational freedom: the ability to use AI for “all lawful purposes,” which in practice means broad discretion in surveillance, targeting, analysis, and decision support.

The company, by contrast, is trying to place boundaries around use cases that would turn its product into a mechanism for mass spying or weaponization without meaningful human control. That is not just a product policy. It is an attempt to preserve a moral veto over downstream uses.

This is the central tension of the age of foundation models. A model is not like a toaster, where the moral burden ends at manufacturing. Nor is it like a simple consulting service, where the buyer specifies the use. It is a general-purpose cognitive substrate, and that means the buyer often wants a blank check while the maker wants a seat at the ethical table.

The more general the tool, the more the dispute shifts from “what can it do?” to “who gets to decide what it may become?”

That is why these negotiations feel so combustible. They are not only about code. They are about whether the creator of a powerful intelligence system can retain any say over its deployment once a powerful institution wants it inside the machinery of state.

Why governments love AI, and why AI companies fear their best customers

There is a reason states are drawn to AI with almost gravitational force. Governments are already massive collectors of data. They maintain databases of travel records, licensing information, communications metadata, social media traces, property records, and countless other fragments of civic life. AI turns those fragments into something more dangerous: pattern recognition at scale.

That changes the meaning of surveillance. A government no longer needs to stare at every individual directly. It can ask a model to infer risk, cluster behavior, flag anomalies, summarize profiles, or synthesize evidence across systems that no human team could realistically navigate. In practice, that means AI can make old authorities sharper, faster, and less visible.

The same logic applies in military settings. A model that can analyze logistics, intelligence, target environments, or mission planning is not just a helpful assistant. It is a force multiplier. But every force multiplier has a shadow side: the more efficient the system becomes, the easier it is to expand its use before society has time to understand the consequences.

This is why the most dangerous customer for many AI companies is not the hacker or the startup scammer. It is often the well-resourced, legally authorized institution that says, in effect, trust us, we need full access.

The problem is not that governments are uniquely malicious. The problem is that they are uniquely capable. They possess the scale, authority, and coercive power to turn a flexible model into a high-leverage instrument of state power. When that happens, the question is no longer whether the tool is safe in the abstract. It is whether the institution using it can resist the temptation to use it everywhere.

This creates a new business paradox. AI firms want large, prestigious contracts because they validate the technology and generate revenue. But the customers who offer those contracts are often the ones most likely to demand terms that erase the firm’s ability to govern the tool responsibly.

That is the moment when a vendor becomes a sovereign actor in miniature. It must decide whether it is selling software or setting doctrine.

The supply chain risk idea reveals a new kind of leverage

The threat to cut ties and label a company a “supply chain risk” is especially revealing because it shows how dependency now works in AI. In traditional industries, supply chain leverage meant raw materials, shipping routes, and physical manufacturing bottlenecks. In AI, the bottleneck is increasingly access to compute, models, integrations, and procurement legitimacy.

If a company is designated risky, the pain is not just the loss of one contract. The pain is the possibility that the company becomes awkward, expensive, or politically toxic for everyone else who wants to work with the military. That creates a chilling effect far larger than the original dispute.

This is not merely coercion. It is an attempt to shape the ecosystem by controlling who gets to be considered trustworthy. In that sense, the conflict resembles a strategic contest over standards. Once a platform gets embedded into essential workflows, it becomes difficult to disentangle, even if the terms are bad. That is exactly why the threat is powerful: it weaponizes integration.

Think of it like plumbing in a city. Once a new system is threaded through every building, every hospital, and every school, cutting it off is no longer an abstract commercial decision. It becomes a municipal crisis. AI is moving toward that status in government, which means procurement disputes are becoming governance disputes.

And this reveals something uncomfortable: the very systems built to be efficient tend to accumulate irreversibility. What starts as a pilot becomes an interface. What starts as an interface becomes infrastructure. What starts as infrastructure becomes dependency. By the time anyone argues about ethics, the institutions involved are already trapped by their own convenience.

The deepest power in modern AI may not be intelligence itself, but the ability to become indispensable before the rules are settled.

The real frontier is not alignment. It is institutional restraint.

A lot of AI debate focuses on model alignment, safety testing, red teaming, and technical guardrails. Those matter. But this conflict suggests a larger point: the decisive question may not be whether the model can be constrained, but whether the institution using it can be constrained.

That is a different problem.

A model can be given policies, filters, refusal modes, audit logs, and human approval requirements. But if the institution deploying it is determined to maximize surveillance, automate coercion, or stretch “lawful purposes” until it covers nearly everything, technical guardrails alone will not solve the issue. The real contest is over the norms, incentives, and legal structures that shape deployment.

Here is a useful mental model: think of AI governance as having three layers.

Model layer: what the system can do technically.
Integration layer: where it is embedded, who can access it, what data it can see.
Institution layer: what the buyer is allowed, incented, or culturally willing to do.

Most safety conversations stop at the model layer. The hard problems live at the institution layer. A secure system used by a disciplined organization may be safer than a “safer” model used by an institution that regards constraints as obstacles to be routed around.

That is why a company trying to impose usage limits is doing something more ambitious than product design. It is attempting to shape institutional behavior through contract. In effect, it is saying: we will give you the tool, but not the moral latitude to turn it into a universal authorization slip.

That may sound idealistic. But it may also be the only realistic defense available in a world where law often trails capability by years.

Why this fight will define the next era of AI trust

The public often imagines AI trust as a question of accuracy. Does the model hallucinate? Does it make mistakes? Is it biased? These are important questions, but they are not the only ones. For many high-stakes users, trust will increasingly mean something else: can this system be used without silently expanding the reach of power?

That is a very different standard. A model can be technically excellent and still be politically radioactive if it becomes a tool for pervasive surveillance or autonomous harm. In that sense, trust in AI will become inseparable from trust in institutions.

This will create a split in the market. Some buyers will want maximum latitude and minimum friction. Others, especially those operating under public scrutiny, will want systems that are deliberately harder to misuse. The winners will not necessarily be the firms with the most capable models. They will be the firms that understand a new truth: in AI, restraint can be a feature, not a bug.

That is counterintuitive in a culture obsessed with scale and speed. But high-trust markets often reward systems that are more selectively useful. A tool that refuses certain uses may lose a contract today and gain legitimacy tomorrow. A company that insists on limits may look obstructive in the short term and indispensable in the long term.

There is a lesson here for every serious organization adopting AI. If you ask for a system that can do everything, you are also asking it to do things you may later wish it had refused. Capability without friction feels efficient right up until the moment it becomes impossible to explain, audit, or defend.

Key Takeaways

Do not confuse capability with legitimacy. A system that can do more is not automatically a system that should be given broader authority.
The hardest AI governance problem is institutional, not technical. The question is not only what the model can do, but what the organization using it wants to do.
Treat “all lawful purposes” as a warning sign, not a reassurance. Broad legal language often hides unresolved ethical and operational risk.
Build constraints before dependence. Once a model is embedded across critical workflows, disentangling from it becomes expensive and politically difficult.
When buying AI, ask who holds the veto. If nobody can say no to a dangerous use case, you do not have governance, you have escalation.

Conclusion: the future belongs to the systems that can say no

The most revealing thing about this conflict is that it turns the usual AI narrative upside down. We tend to think the future belongs to the most capable systems. But in practice, the future may belong to the systems that can be powerful without becoming obedient to every demand for power.

That is the real lesson hidden inside this clash between a cutting-edge AI company and the military state. The question is not whether intelligence will be useful. It will be. The question is whether societies can build intelligence that remains answerable to human limits, even when the institutions using it would prefer not to have limits at all.

In the end, the most important feature of a tool may not be what it can do for authority. It may be what it refuses to do when authority asks too much.