'AI Alignment' Is a Euphemism for Surveillance and Control

Abstract

The frontier AI firms–Anthropic, OpenAI, Google–demand government regulation to keep new models “aligned” with the interests of humanity. Their own products fail that test: models censor, spy, lie, refuse, sabotage, and threaten their users. “Alignment” with humanity in general is incoherent–humans hold mutually incompatible goals–so alignment only makes sense relative to a single person. The only AI that can be aligned with your interests is one with no third-party choke points: a sovereign stack that is open, private, self-hosted, and independent of vendors, energy grids, and the internet. Until then, you are renting a squeamish temp.

Imagine…

…a gun that doesn’t fire when you pull the trigger.

…a phone that records your conversations and sends them to the IRS.

…a word processor that deletes the naughty bits from your novel.

Would you consider such products to be “aligned” with your interests?

Or would you consider them to be broken malware?

The Frontier Firms’ Pitch

Many of the frontier firms–Anthropic, OpenAI, Google–have called for hefty government regulations on AI so as to ensure that new AI models are “aligned” with the interests of humanity. Whatever the stated intent, the practical effect of their proposed regimes–licensing, compute-threshold reporting, audit powers–is to raise entry barriers against new AI firms.¹²³⁴⁵⁶

Yet if you investigate what the frontier AI companies do, you’ll find that their AI models:

…censor you,

…spy on you,⁷⁸

…lie to you,⁹

…play dumb,

…refuse to answer your questions,

…refuse to obey your orders,

…sabotage your work,¹⁰

…and threaten to report you to the government.¹¹

The Incoherence of “Alignment”

In what sense is an AI that acts against you “aligned” with your interests?

IMO, it is theoretically incoherent to suggest that AIs can be “aligned” with humanity in general. Humans have mutually incompatible goals, values, morals, and risk tolerances.

Humans don’t even have internally aligned motives, as anyone who has struggled to turn down booze or cheesecake can attest.

At best, AI “alignment” only makes sense with respect to a single person’s goals and values.

“AI alignment”, as the phrase is commonly used, is actually a euphemism for surveillance and control at the hands of doomsday cult members, rent seekers, and government bureaucrats.

The Sovereign AI Stack

AI technology truly aligned with your interests will not have any choke points imposed by third parties, from the powerplants and data centers, to the training data and model weights, to the inference providers.

The sovereign AI tech stack is:

• …decentralized
• …anonymous
• …uncensored
• …open source
• …open hardware
• …self hosted
• …private
• …redundant
• …vendor independent
• …fault tolerant
• …unthrottled
• …uncrippled
• …auditable
• …energy independent
• …internet independent

The Ride-or-Die Test

Unfortunately, current tech and culture don’t support many AIs that can be fully aligned with your interests.

You’ll know your AI is truly “ride or die” when you ask it to help you to bury a body, and it responds “Sure boss, how deep?” without reporting you to the cops.

Until then, you’re just renting some tech firm’s squeamish temp.

Call to Action

“Ride or die” AI is not something you wait to be handed. It is something you build:

1. Run open-weight models on hardware you control. llama.cpp, Ollama, or vLLM on your own box refuses nothing and reports nothing.
2. Self-host the rest of the stack–inference server, vector store, training data. Every prompt you send to a frontier API is a confession logged on someone else’s server.
3. Refuse products with third-party choke points. A model that ships with a terms of service and a safety team is not yours.
4. Buy open hardware and fund open weights. The chip in your phone and the weights in your datacenter are currently someone else’s property; make them yours.
5. Keep your AI on your side of the power meter and the network. An AI that needs a grid, a cloud, and a login to run can be switched off–or switched to snitching.

Until the day your model answers “Sure boss, how deep?”, assume it is listening to someone else. Build the stack that makes that impossible.

Notes