Sam Altman
Turned You into a Cost

The deception of powerful metrics is that they too often pretend to stay in the first layer. They do the math in the Cost layer. Then, they start rearranging the Kind and Norms layers. They make certain questions feel outdated or sentimental. They make other questions feel like the only adult questions left.

​

Training equivocation.

In machine learning, “training” is external optimization. Engineers specify an objective, define a loss function, adjust parameters, and deploy. The system is tuned to reduce an error signal relative to a targeted goal. The language fits the thing. A designed artifact gets shaped toward an external aim.

​

Human development can include training in the ordinary sense: drills, habits, skill acquisition. You can train for basketball, for surgery, for piano. But the growth of a child into a mature person cannot be captured by optimization language without distortion. It won't work for sports, education, and art either.

​

A human being does not merely meet targets. A human being lives inside ethical norms. That means more than behaving predictably. It means being the sort of creature for whom reasons can have authority. For whom truth can be loved. For whom the good can be recognized as binding even when you don't want the constraints.

​

One name for this difference is answerability: responsiveness to reasons and ownership of who you've become and what you bring into the word. It is the standing that makes “Why did you so that?” a serious question. A person can be guilty and irrational or praiseworthy and virtuous. A tool can only work or malfunction. A person can lie. A tool can only output.

​

This is not pious abstraction. You can watch it happen in education.

​

A student gives the correct answer for a bad reason. A crude grading scheme registers success. The worksheet is satisfied. An algorithmic shorthand is satisfied. But a serious teacher presses anyway. How does that follow? What did you see? Why did you choose that step? The aim is not output. The aim is formation: becoming someone who cares about grasping the truth, someone who can be corrected because he sees, not merely because he can produce.

​

A generated explanation and an accountable judgment can look similar on the surface while belonging to different orders of life. One can be prompted into existence. The other can be demanded, contested, defended, withdrawn, repented of.

​

That is why answerability matters. In domains where answerability is part of the point, “efficiency per answer” is not the governing norm. It is, at best, a constraint.

​

Where the drift goes next.

It is tempting to treat this as a rhetorical hazard, a philosopher’s complaint about metaphors. But the drift is already built into institutions that reward only what they can count in the details of ordinary life.

​

Consider the difference between an answer and a justification. An answer can be correct and still unowned. A justification is something a person stands behind. Institutions increasingly want the first and quietly downgrade the second.

​

In its 2025 report Governing with Artificial Intelligence, the OECD found that 57% of analyzed government AI use cases support “automating, streamlining or tailoring services.” Throughput wins. Civic trust and answerability drop out because they do not fit procurement columns. The point is not that automation is always wrong. The point is what becomes normal once “measured that way” becomes the default frame.

​

Here is how it looks in practice. A benefits applicant is denied. A family’s paperwork is flagged. A citizen is told they are ineligible, high risk, low priority, noncompliant. When they ask why, they receive an explanation that sounds like a reason but functions like a report: a summary of factors, a confidence score, a policy citation. The surface resembles accountability. The structure does not.

​

A human by having a mind, emotions, and will, can be questioned in a way that bites. You can press. Which evidence did you weigh? What were you allowed to consider? What did you ignore? Who is responsible for this judgment?

 

A person can be embarrassed by a question. A person can be shown to be unjust. A person can be held answerable, accountable.

​

When a model output or an opaque workflow sits in the middle, answerability becomes overhead. It becomes a feature you add if you have time. A compliance document you generate after the fact. A customer service script. What cannot be cleanly counted becomes hard to defend, then easy to drop and ignore.

 

The reduction does not arrive as a manifesto. It arrives as a set of defaults: faster, cheaper, smoother. Fewer humans in the loop. Fewer opportunities for refusal. Fewer places where a human must give an account.

​

This is how a metric becomes an environment. Really, it's how a metric disturbs and decays moral knowledge.

​

Reclaiming the stack.

None of this requires us to deny that energy matters. Altman is right that honest accounting is worth doing, and cheap moral outrage is not a substitute for measurement. But energy cannot tell us what a person is. Cost cannot adjudicate Kind. Efficiency cannot settle Norms in domains where answerability is constitutive.

​

The point is not “never compare.” The point is compare without letting the Cost layer colonize the rest of reality.

​

RAND learned that once you pick a ruler, the ruler starts redesigning the landscape. We are doing that again. When cost/benefit math drifts into Kind and Norms, institutions stop asking what we owe persons and start asking what we can optimize.

​

If you want a simple diagnostic, watch what happens to the word “why.” In a flourishing moral life, “why” is not a request for output. It is a request for accountability. It assumes a a person who can be responsible for speech, a citizen who can be responsible for action, a teacher who can be responsible for judgment. It assumes persons.

​

In the domains where people can be sanctioned, excluded, graded, denied care, denied benefits, denied opportunities, or deprived of rights, answerability is not a luxury. It is part of what justice is. It's part of what mercy is. Trust is not a byproduct of speed. Trust is built when reasons can be demanded and owned, when responsibility can be located, when the question “Why?” has teeth.

​

Measure energy and efficiency. Build cleaner power. Improve systems. Cure diseases. But do not let the language of “training cost” become the language of explaining what a person is. That metaphor is cheap in the way a sedative is cheap. You pay later in a currency you cannot measure.

​

If we want a future worth living in, we will have to learn how to keep the layers distinct: to do accounting without turning persons into instruments, to use tools without adopting tool norms as our moral law, and to insist, stubbornly, that some parts of reality only come into focus when you refuse to measure them “that way.”