Ghost in the gates: AI, deregulation, and the unwritten future.

A song, an AI model, and a reverse-engineered future

This article is an experiment in time travel, of sorts.

The lyrics that follow were written in early 2025, in response to the sweeping deregulation of AI under a U.S. administration that decided guardrails were optional and acceleration was the only acceptable speed. The song is short. The lyrics are bleak. The essay around them is what happens when you take the song seriously and ask: what would have had to be true for this song to feel correct, even prescient, ten years from now?

The lyrics

"ghost in the gates, ghost in the gates / it learns your name, it learns your face / we let it in, we let it win / now the ghost is running everything … signed the deal, erased the laws / money talked, the brakes came off / cut the chain, killed the light / called it progress, called it right"

"Ghost in the gates" plays on the old idea of a ghost in the machine, the unexplained presence inside a system that isn't supposed to have one, and on the gates we build to keep threats out. The gates were the defense. But we signed the deal, erased the laws, and left the doors wide open: the thing walks straight in, learns your name, learns your face, and starts running everything. AI as the ghost. The dismantled guardrails as the open gate. The chorus is the moment we stop pretending there's a human at the steering wheel.

The future I'm reverse-engineering

For the song to feel correct in 2035, here's what would have to be true:

One: alignment work gets defunded. Not banned, just starved. The people working on AI safety lose their grants, their corporate sponsorship, and most importantly their political cover. The work doesn't stop entirely, but it stops mattering in the rooms where the deployment decisions get made.

Two: the biggest models get bigger faster. Compute clusters in the gigawatt range. Training runs that would have been impossible to defend a few years ago, run on data nobody can audit anymore because the auditing requirements were dismantled. Capabilities emerge in jumps, not glides, because nobody is checking the work between jumps.

Three: a meaningful incident happens. Not extinction-level, not even close. Something at the scale of a flash crash that wipes out a few percent of global financial markets in an hour, or an autonomous system that does the wrong thing at scale, or an information operation that successfully steers an election in a country with a thin democratic margin. The incident is bad enough to be a five-day news story. Not bad enough to change anything.

Four: the response is more deregulation. The incident gets framed as the cost of progress. The proposed fix is to remove the remaining frictions on innovation, because the alternative is "losing to China." Anyone who points out that this is roughly the logic of "if a fire keeps starting in the kitchen, the answer is to remove the fire alarm" is dismissed as a doomer or a luddite.

Five: by 2035, the ghost is just part of the furniture. Not benevolent, not malicious, just present. Decisions get made that nobody can fully explain. Outcomes get distributed in ways that nobody can fully predict. The systems work well enough that pulling them out would be more disruptive than living with them. The song's refrain ("we let it in, we let it win, now the ghost is running everything") describes the moment everyone realized this, and decided to keep going anyway.

I don't actually think this is inevitable

Songs aren't predictions. They're flares. The point isn't that this future is locked in. The point is that the trajectory in 2025 was sufficient to make this future plausible, and the people who shrugged off plausible bad futures historically don't end up looking smart in retrospect.

I run an AI company. I build with AI every day. I think AI is the most important technology since computing itself. I also think the people who care about it most should be the most insistent that we don't accelerate past the ability to course-correct. The frame "regulation vs. innovation" is a false dichotomy invented by people who benefit from the absence of one of them. Good regulation enables innovation by giving everyone the same rules, and by punishing the bad actors who would otherwise drag the whole field down.

What we can still do

I'm not going to pretend an essay changes policy. But I will say what I'm doing personally, in case it's useful for anyone reading this.

I'm building AI products that I would still be proud of if my children read the source code in twenty years. I'm structuring my company so that the cost of doing the right thing isn't ruinous. I'm publishing my thinking out loud, not because I expect to change minds, but because the alternative is being part of the silent majority that says "I knew this was bad but I had work to do."

The ghost is in the gates. We can still decide what kind of ghost it becomes.