AI: Who Gets the Keys?

I would love you to watch Episode 5 of The Human Precipice. But if you'd rather read than watch, I appreciate you taking the time.

AI is going to become the most consequential technology in human history. I think we are already most of the way there. So who is going to control it? Not in some abstract sense. Practically. Whose hands are on the levers? Whose values are in the training data? Whose interests are being served? Whose definition of "safety" decides what gets built and what doesn't? The political class, on both sides of the aisle, obviously have woken up to understand that AI is coming, and with that awakening, they are moving as quickly as they can, to make sure that whatever AI becomes, the keys end up in their pocket.

Before I get into specifics, you should know how I see almost every political question.

I have come to believe, after watching American leadership for most of my adult life, that the system we have does not actually work towards solving problems. It manages them. It uses them. But solving them, in the way you or I would solve a problem in our own life or our own business, that kind of problem solving is mostly off the table.

Healthcare costs. Immigration. The deficit. Education. Infrastructure. The opioid crisis. Housing, Birth Control, and Gun Control. The list is long, and the pattern is consistent. The problem gets named. Hearings happen. Bills are introduced. Sometimes bills pass. Twenty years later the problems are still there. Often worse.

After a while, if you look closely enough, you begin to understand that incompetence is not a plausible explanation. At some point, you need to consider that this inability to problem solve is not an inability to problem solve at all, but rather, it is intentional, and very well executed theatre.

I know this can sound conspiratorial. Images of smoke-filled rooms where the political class meet to plot how to keep things broken. But the mechanism is more banal than that. It is the structure of the incentives.

An unresolved problem fundraises. An unresolved problem mobilizes a base. An unresolved problem keeps cable channel ratings high and campaigns running. A solved problem doesn't. So the system selects, year after year, election after election, for people who are excellent at performative art. They name the problems. They jump up and down. They are convincing. But they solve very little. Because fixing problems would put themselves out of business.

I know this sounds cynical, as I intend it to be, but of course, there are real reformers. There are people who go into politics genuinely trying to change the system. But the system is older and stronger than any wide-eyed idealist. They get chewed up and spit out, or they survive long enough to learn the game and to play by the rules.

It would be silly to argue that all political disagreement is fake. There are real moral and practical disagreements between Americans, on real issues. But the amplification of those disagreements, the way every issue gets weaponized, polarized, and held hostage to fundraising, that is not natural. That is engineered. And it's engineered for the benefit of those that created and maintain the system. An easy to see illustration of this system is exposed in the near perfect balance of political strength. How would you imagine a well engineered system to appear? I will tell you: About 50% of the population leaning one way and 50% the other.

So, this is where I am coming from. Agree or disagree, I wanted you to understand my mindset because it creates a very specific lens through which I see what is happening in AI policymaking today, in the U.S., and in much of the rest of the world.

The Sanders moratorium

In March of 2026, Senator Bernie Sanders and Congresswoman Alexandria Ocasio-Cortez introduced the AI Data Center Moratorium Act. The bill calls for a federal halt on new AI data center construction until a set of guardrails are in place: safety, environmental, economic, and labor related.

The framing, in Sanders' own words, is that we cannot let a handful of billionaire tech oligarchs reshape the future of humanity. More than a hundred local communities and twelve states are moving in similar directions. Denver's mayor declared a moratorium of his own. What would have been considered a fringe idea only a year previously was ramping up to be a full on assault. The video that prompted me to write this essay, and the companion podcast, "AI: Who Gets the Keys?" was one in a series of press events and publicity pieces Sanders has put out over the last several months, first making the rounds in advance of the Bill, and then continuing the push in support of the bill.

I had seen several of these videos but this one really stood out. Sanders sat down with a panel of scientists from the United States and China. The panel was hosted by the Future of Life Institute. I watched the full hour. For the first thirty minutes, the audience heard only one frame: AI as existential threat, suicide race, Oppenheimer moment, slaughterbots, deceptive alignment, intelligence explosions, AI starting nuclear wars before humans realize what is happening, humanity reduced to "gorillas," their wording, not mine.

Every panelist came from the existential-risk camp of AI thinking. That is one school of thought. But it is not the only one. There are serious researchers who think the existential framing is overstated. There are economists who think AI is on track to deliver the largest expansion of human capability since electricity. There are people who built these systems and know them from the inside who would tell you the doomer narrative misunderstands what these models actually are. There are safety researchers who would tell you that halting infrastructure is not how you address the actual technical risks they spend their days trying to solve. None of them were on the panel. I am not saying the people on the panel are wrong to be concerned. Their concerns warrant real discussion. But I am saying the panel was built to produce one conclusion. When every serious counterweight is missing, the audience is not hearing a debate. It was not a discussion. The audience was presented with a curated argument, delivered as expert consensus. And by the way, I am not sure if Bernie thought to invite the Chinese panelists on board to show that they were allies of the U.S., working with American political leadership to slow the rollout of Skynet and the Terminator, but I can tell you that neither Bernie nor anyone on the panel acknowledged the Chinese panelists for what they are. Xue Lan and Zeng Yi are not just academics. Xue chairs China's national AI governance expert committee under the Ministry of Science and Technology. Zeng leads the Beijing Institute of AI Safety and Governance, an institution embedded in the Chinese government's AI policy apparatus. These are not independent voices. They are state-aligned figures from a country that is racing the United States in AI development and shows no sign of pausing anything, while the honorable Senator Sanders parades them before us as if they are the real voices of reason. I cannot see into Senator Sanders' heart and cannot read his mind, but how many explanations for this can there be when you rule out ignorance and incompetence, and I would not call Bernie Sanders either of those things? And yet he invites two high-ranking Chinese officials with known connection to the Red Army, who are actively working to give China the tools needed to dominate globally. Bernie knows these guys have no intention of stopping AI development so what motive can be attributed to Bernie inviting them?

What the bill actually does

Ok, enough about the video. Let's get back to the bill and what it actually does. Simply put, it halts new construction of AI data centers. The data centers that already exist keep running. The trained models that already exist keep operating. The companies that already have massive compute infrastructure keep their head start. Perhaps, permanently.

The startups without data centers already built are frozen out. The smaller competitors trying to break in are frozen out. The open-source community trying to build alternatives that are not controlled by big labs gets frozen out. Foreign companies, like the Chinese labs as an example, not bound by American legislation, keep building. China races on. We pause. The race does not stop. We just stop running in it.

So the bill proposed in the name of stopping the oligarchs, even with casual scrutiny, clearly locks the existing oligarchs in. At the same time, it hands a new set of keys to the federal regulatory class, the agency heads, the licensing officials, the bureaucrats who get to decide when the safeguards are sufficient and when they are not.

The bill also bundles together AI safety, environmental concerns, electricity prices, and job loss, four very different problems with four very different solutions, all of which apparently get solved by the same mechanism. Stopping data centers.

When one solution claims to solve every problem, you should pay attention. That is almost never how real problem-solving works. That is how political marketing works.

The energy bottleneck nobody on the panel mentioned

Here is something the panel did not discuss, and the bill does not address, and the press coverage of the bill mostly did not raise. Sanders and the moratorium movement point to AI's energy consumption as one of the strongest reasons to halt construction. Data centers use enormous amounts of electricity. Hyperscale facilities can draw as much power as a small city. Demand is growing rapidly. Some grids are stressed. Prices in some regions are rising. All of that is true.

What is also true, and what nobody on the moratorium side wants to talk about, is that the United States already has more than two and a half terawatts of new power generation capacity sitting in regulatory queues, waiting to be built. According to research from Lawrence Berkeley National Laboratory, as of the end of 2023, approximately 2,600 gigawatts of proposed generation and storage projects were waiting for grid interconnection. To put that in perspective: the entire installed capacity of the current U.S. power grid is roughly 1,280 gigawatts. So there is more than twice the current U.S. grid sitting in line, waiting for permission to be built.

And it is not coal. More than 95 percent of that queued capacity is zero-carbon: roughly 1,080 gigawatts of solar, 1,030 gigawatts of battery storage, and 366 gigawatts of wind.

So why is it stuck?

It is stuck because the regulatory process for connecting new generation to the grid has nearly tripled in length. Projects built between 2000 and 2007 took less than two years from interconnection request to commercial operation. Projects built between 2018 and 2024 take more than four years, with a median of five years for projects built in 2023. Some specific projects, including some involving the largest tech companies in the world, have been quoted timelines of eight to twelve years.

These delays are not driven by physics. They are driven by a regulatory architecture, federal, state, and utility-level, that has accumulated layer upon layer of process over the last forty years. Each layer was added for a reason. Each layer made sense in isolation. Together they have produced a system in which it is faster to build a new factory than to build the power source that runs it.

So pause for a moment and notice the structure of the situation.

Step one: Government creates a regulatory environment so slow and so expensive that clean energy capacity equal to twice the entire current U.S. grid is stuck in queues, unable to be built.

Step two: That bottleneck causes electricity prices to rise and grids to strain, especially in regions where new demand is concentrated.

Step three: Government points at the rising prices and the strained grids and says: "You see? AI data centers are the problem. We must halt them."

Step four: The proposed solution is more government, this time at the federal level, with a moratorium that adds a new licensing layer on top of all the existing regulatory layers.

If this pattern feels familiar to you, that is because you have seen it before. Government creates the problem. Government uses the problem to justify expanding government. Then the original problem gets used as the justification for the next expansion. And so on.

This is not me being conspiratorial. This is the actual structure of energy policy in this country. The PJM grid, which serves a huge portion of the eastern United States, just held a capacity auction at which prices rose roughly tenfold from one year to the next, partly because the queued clean energy that would have lowered prices could not get built in time. Independent analysis concluded that if just ten percent of the queued renewables had been built, billions of dollars in cost increases could have been avoided.

Everyone in the energy industry knows this. It is not a libertarian fringe view. It is in federally funded research, in mainstream utility journalism, in industry analysis, in the academic literature. The same political class that is now pointing at data centers as the cause of high electricity prices is the political class that built and maintained the regulatory framework that prevented the clean energy that would have lowered those prices from being constructed.

And then they wrote a bill to halt data centers. To save the climate. By stopping the technology that, more than any other technology in our lifetime, might actually let us solve climate change at the root.

What AI could actually do for the climate

In Episode 3 I talked about the fact that AI is currently controlling the plasma inside experimental fusion reactors. It is adjusting the magnets ten thousand times a second to hold superheated matter stable, doing things human engineers had not figured out how to do. DeepMind is now partnered with Commonwealth Fusion Systems, the company trying to build the first commercially viable fusion power plant. If fusion works, and I think it will, possibly within the next decade, AI will have helped get us there. Fusion means clean, essentially unlimited, baseload energy. The end of the energy scarcity problem that has driven half the wars and most of the environmental damage of the last hundred years.

AI is also accelerating materials science: new battery chemistries, more efficient solar cells, better carbon capture compounds. AI is optimizing power grids in real time. AI is modeling climate systems with a precision we have never had before. AI is screening compounds for water purification at scales that were unthinkable five years ago. AlphaFold mapped two hundred million protein structures in a few years, more than every laboratory on Earth had managed in the previous fifty, and that single dataset is now accelerating drug discovery, agriculture, and biofuel research worldwide.

If you genuinely believe the planet is in crisis, the question is not whether AI uses too much power right now. The question is whether AI is the technology that can help us build the post-scarcity, post-fossil-fuel infrastructure that solves the problem at its root.

That conversation cannot happen on a panel where everyone has already decided that AI is the threat. Because if you grant that AI might be the path to solving the climate crisis, you cannot simultaneously argue that AI must be halted to save the climate.

The conversation about jobs

There is also the labor argument. Sanders' bill claims to protect workers by halting data center construction.

In Episode 2 I talked about my own anxiety, working with these tools every day, about training my own replacement. That fear is real. I take it seriously. So do most of the people I know who use these tools well. The disruption to white-collar work over the next decade is going to be substantial, and the people most exposed are not all going to be fine.

But notice what halting American data center construction does and does not do for that problem. If AI is going to disrupt labor, and it is, halting the construction of physical infrastructure on American soil does not stop the disruption. It just means the disruption arrives from somewhere else. China keeps building. India keeps building. Europe keeps building. The AI that takes the American accountant's job in 2030 will exist regardless of whether the data center it runs on sits in Texas or Shanghai. The American accountant is not protected. The American economy just loses the data center.

The harder conversation, the one nobody on that panel wanted to have, is about the economic transition itself.

If we are heading toward a post-labor economy, where machines do most of what used to be paid work, our entire economic operating system needs to be rethought. Healthcare tied to employment does not work in a post-labor world. Worth measured by productivity, which I talked about in Episode 2, does not work either. Tax policy designed for an economy of human wages, when most value is created by capital and machines, does not work. The social contract we built between roughly 1945 and 1980 was designed for a labor-based economy that may not exist by 2045. That is the conversation that actually serves working families. It is also the conversation that requires politicians to do something hard. Imagine a future. Build something new. Negotiate trade-offs. Disappoint constituencies on both sides. Acknowledge that the answers we have grown up with may not work for the world our children are going to inherit.

Halting data centers is much easier than rethinking the social contract. So that is what is getting proposed.

The other side of the aisle

If you have read this far you might be thinking I have written a critique of the left. I have not. I have written a critique of one specific policy that comes from the left. The deeper critique, the one I started this essay with, is structural, and it applies to both parties.

So let me close that loop with the Republican example.

While Senator Sanders has been working to halt AI infrastructure from the left, the Trump administration has been moving in a different direction from the right. The Department of War, working with an AI company called Anthropic, demanded that the company remove its restrictions on two specific uses of its AI: mass domestic surveillance, and fully autonomous lethal weapons. Anthropic refused. The company said it would not allow its technology to be used to surveil American citizens without warrants, and would not allow fully autonomous AI to make kill decisions without a human in the loop.

The administration's response was to designate the company a "supply chain risk," a label normally reserved for adversary nations like China and Russia, never before applied to an American company. The President directed every federal agency to stop using their products. Court documents later revealed the supply chain risk designation was issued specifically because the company had been, in the government's own words, "hostile in manner through the press," in other words, because they refused to comply quietly. That is textbook First Amendment retaliation, against an American company, by the government, for the offense of public disagreement.

I should disclose something here: I use all the major frontier models, Gemini, ChatGPT, and Claude, to help build this channel. You can weigh what I am saying with that in mind. But the structural point does not depend on me. The court filings are public. The pattern is clear.

Sanders, from the left, wants to halt AI in the name of protecting workers from oligarchs. The policy itself locks in incumbents and hands new keys to regulators.

The administration, from the right, wants to force AI companies to remove safeguards in the name of national security. The demands include unrestricted domestic surveillance and autonomous lethal weapons.

Different teams. Different methods. Same instinct. Get the keys. Grab control.

Why the keys matter

And now I want to come to the part that I think is the most important, and was the most challenging to write about.

Whoever controls AI is going to control something more than an industry or a market. They are going to control, maybe not in a literal sense, but in the sense that practically matters, they are going to control reality.

If the AI that helps you research a topic has been trained to emphasize some things and de-emphasize others, you are not researching the topic. You are researching the topic the way someone wanted you to see it. If the AI that helps you understand a piece of legislation is built to nudge you toward one conclusion, you are not understanding the legislation. You are absorbing a position. If the AI that mediates your access to history, science, medicine, and policy has been quietly shaped, and you cannot see the training data, you cannot audit the alignment, you cannot see what was filtered out, then you are not thinking for yourself anymore. You are thinking inside the lines somebody else drew.

This is not a hypothetical. Every AI system is shaped by choices. What data to train on. What to weight. What to suppress. What "alignment" means. What "safety" looks like in practice. Those choices encode values. Sometimes they are made deliberately; sometimes they are made through unexamined assumption; sometimes they are made under pressure from regulators, advertisers, or governments. But they are always made.

There is no neutral AI. There is only AI shaped by someone.

So the question of who shapes it is enormous.

The people who built newspapers had a kind of power. The people who built broadcast television had more. The people who built social media had more still. The people who build AI, and the people who get to decide what AI is allowed to say and not say, will have the most epistemic power any institution has held in human history. By a margin that is hard to overstate.

This is the prize the political class is fighting over right now. Not the data centers. Not the environmental costs. Not the jobs. Those are the surface arguments. The thing underneath is the keys to the epistemic infrastructure of the next era of human civilization.

And here is the part where my view gets sharp: the honest answer to that concern is not to hand the keys to fewer people. It is to make sure they are distributed widely. Multiple labs. Multiple approaches. Open source where possible. Transparency where possible. Real competition. Citizen access. The exact opposite of what either a federal moratorium or a coerced government monopoly produces.

Concentration is the danger. Not abundance. Not the technology itself. Concentration.

Why this matters more than the usual political fight

I want to end where I started. With the lens I bring to American politics, and with the question of why AI is different.

I said earlier that I think the political class on both sides has discovered that unresolved problems are more useful to them than solved ones. I believe that. I have watched it for about forty years and I believe most of the architecture of contemporary American politics, the tribal coverage, the engineered outrage, the carefully maintained inability to actually fix anything, serves the people inside the system better than it serves the people outside it.

AI is the first technology in our lifetimes that genuinely threatens that arrangement.

Not because AI is going to fix politics. It is not. But because AI, done right and distributed widely, would let ordinary people do things the political-media class has long held a near-monopoly on. Synthesize information. See through marketing. Read a bill and understand it. Compare what a politician said today with what the same politician said five years ago. Cut through curated panels and access primary sources. Reach independent conclusions without needing a guild of professional explainers as middlemen.

That is the threat. Not to billionaires. To the people whose power depends on you not knowing what they know. If you want to understand why both ends of the political spectrum are reaching for the AI keys right now, left through moratoriums, right through coerced compliance with state security demands, you do not have to assume any individual politician is acting in bad faith. You only have to notice the pattern.

Both ends of the system see the same thing. Both ends are responding the same way. Both ends are using the language of protecting you while structurally taking the technology that could most empower you and bringing it under the control of the institutions you have the least leverage over.

Watch what they do. Not what they say.

That has always been the right rule for evaluating political actors. With AI, it is no longer just the right rule. It might be the only rule that matters.