The Harms Are Specific, the Gains Are Vague

I was half-listening to an AI regulation podcast when the pattern landed. Not a new podcast — there are a thousand of them now, all saying roughly the same things with different levels of alarm. This one had someone who had actually read the filings, the testimony, the CEO letters to Congress. I was not taking notes. I was making lunch.

That is where I heard it: the AI company leaders are very specific about harms. They can name the displaced workers — coders, first, but then radiologists, paralegals, customer service reps. They cite the studies on teenage mental health and algorithmic exposure. They talk about AI-generated pornography with enough specificity to suggest they have thought hard about it internally. The precision is striking when you notice it.

Then they pivot to the gains.

Drug discovery. Material scarcity. Civilizational uplift. Climate solutions. These are phrases, not claims. Nobody says how many drugs, which diseases, what timeline. The CEO of a company that just laid off a thousand engineers says AI will solve cancer. The imprecision is so consistent across so many executives that it stops being carelessness and starts being information.

You only speak specifically about things you have already measured.

This is the tell. The harms are present-tense: documentable, legally consequential, already in litigation. The gains are future-tense: unfalsifiable, permanently deferred, immune to accountability. When tobacco companies were precise internally about carcinogen risks and vague publicly about the research, we eventually called that fraud. Fossil fuel executives ran internal climate models for decades before the public had access to credible projections. The pattern is not new. It is an industry playbook: know exactly what you know, and speak vaguely about what you are selling.

I am not calling anyone a fraud. I am saying the asymmetry has a structure. And once you see the structure, the optimism sounds different.

The governance gap is the honest problem nobody wants to solve. Legislation moves at the speed of consensus. Deployment moves at the speed of capital. These are not the same speed — not close, not fixable by writing better laws. By the time democratic institutions reach agreement on what AI is allowed to do, the technology has already redefined what institutions do. The question is not whether democratic governance can keep up. It is whether the lag is already structural. I think it probably is.

None of which is comfortable to sit with. Especially if, like most people, you have already integrated these tools into your daily work and are not going back.

Here is where it gets concrete. Some teachers still assign take-home essays as critical thinking exercises. Write an essay on a topic, hand it in, get a grade. The essay is supposed to demonstrate that the student can reason through something hard.

It does not do that anymore. Now it demonstrates whether the student is good at prompting a language model — or not good at it. Either way, the essay is not evidence of thinking. It is evidence of access and process. The schools that have not updated this assignment are not teaching critical thinking. They are teaching AI prompt engineering without calling it that, and grading it as something else entirely.

This is the democratic institutions problem at a smaller scale. The assignment exists because it used to measure the thing it was supposed to measure. It no longer does. But updating the assignment means admitting the old one is broken, and the institutional incentive is to not admit that. So the assignment persists. So does the grading rubric. So does the belief that what is being measured is still being measured.

This is what I mean when I say practice is systems literacy. Not yoga specifically — though a physical practice makes feedback fast and honest; your body tells you within one breath if what you are doing is off. Systems literacy means reading the feedback that is actually present, not the feedback the system claims it is giving you. It means noticing when the thing you are measuring has drifted from the thing you care about.

The AI regulation discourse is full of people measuring the stated intentions of founders. The more useful measurement is the asymmetry between how they talk about harms versus gains. One of those is falsifiable. One of them is not. That difference is not random.

The vagueness is the data.