Safe AI, Everywhere Else

There is an uncomfortable contradiction at the center of the AI revolution, and it becomes more visible with every passing month.

The technology industry invests enormous energy in talking about safety. Research teams publish alignment papers. Companies establish ethics boards. Conferences convene panels on responsible development. And much of this work is genuine, reflecting serious people grappling with serious problems under real constraints.

But at the same time, governments, militaries, and corporations are racing to integrate AI into systems where the consequences of failure are measured not in user complaints or quarterly earnings, but in human lives and geopolitical stability.

In June 2026, the White House issued a national-security memorandum directing the U.S. national-security enterprise to accelerate AI adoption and establish partnerships with private industry to make advanced AI systems broadly available for national-security applications. The memorandum also directed an update to existing policy governing autonomy in weapons systems.

This is not a scandal. It is not a secret program. It is the logical, predictable consequence of a technology that has become too powerful for any government to ignore and too strategically significant for any military to forgo.

AI is becoming infrastructure. Not in the metaphorical sense of "it's everywhere," but in the literal, structural sense: it is being woven into the systems that societies depend on to function. The same underlying technology that writes a marketing email can diagnose patterns in genomic data, flag a suspicious vehicle on a highway camera, analyze satellite imagery of a foreign military installation, or assist in a targeting process that ends with a weapon striking a building.

The model does not know which world it has entered. The surrounding institution determines that. And institutions, unlike models, are not optimized for alignment. They are optimized for their own objectives.

A corporation seeks competitive advantage. A military seeks strategic superiority. A government seeks security, or more precisely, the appearance of security sufficient to maintain public confidence. An intelligence agency seeks information. A police department seeks more effective enforcement. A consumer seeks convenience.

None of these objectives are evil. But they conflict with one another in ways that no amount of technical alignment research can resolve.

Privacy conflicts with surveillance. Human employment conflicts with automation. Transparency conflicts with classified operations. Deliberation conflicts with the speed that modern systems demand. Safety conflicts with the competitive pressure to ship first. And democratic oversight, the mechanism by which citizens are supposed to maintain meaningful control over the systems that govern their lives, conflicts with the sheer complexity of the technologies it is supposed to oversee.

This is the real AI safety problem. Not simply "Can we make the machine behave?" That is a critical engineering challenge, and it deserves the attention it is receiving. But the harder question, the one with fewer papers and fewer panels and fewer satisfying answers, is this: "Can we build institutions capable of using extraordinarily powerful machines without losing control of the outcomes?"

The history of powerful technologies suggests that this question deserves more humility than it typically receives.