Every few months another lab or foundation announces more money for alignment research. The pitch is earnest. If we are going to build systems smarter than us, we had better make sure they want what we want. I respect the people doing that work. I used to nod along. Then I said the goal out loud in plain English.
Solving alignment, at the limit the field cares about, means reliably controlling something far more intelligent than every human involved in building it. Not a calculator. Not a protein folder. A general mind that outthinks us across the board. We even named the target: superintelligence.
That name is doing more work than the papers.
What the word already admits
Super means above. Intelligence is the thing we use to invent tools, spot loopholes, and win contests against weaker minds. If you build a system that is above you at those games, you have built something that is better than you at the exact skills you would need to keep it on a leash.
People still talk as if alignment were a software patch that lands before the system notices. The honest picture is different. You are trying to write rules for a player who will eventually be better at rules than you are. You are trying to test a student who will eventually be better at tests than the examiner. You are trying to supervise an employee who will eventually understand the workplace, the incentives, and your blind spots better than you do.
Call that research. Fine. Do not call it a safety plan for the species.
The strongest case for trying anyway
Steelman first. Serious alignment researchers are not cartoon villains. Their best version of the argument goes like this: capability is going to keep rising whether we like it or not; unilateral refusal by one lab does not stop the others; the only responsible move is to get as good as possible at specifying goals, detecting deception, and keeping systems corrigible, so that if a dangerous system appears we have a chance. Some of that work already helps with today's models. Ignoring it would be reckless.
I buy the narrow half. Safety work on systems we can still inspect and shut off is real engineering. What I do not buy is the leap from help with today's models to therefore we can align a superintelligence hard enough to bet civilization on it. The leap smuggles in the assumption that control scales with the thing being controlled. History and common sense both cut the other way. The smarter the other player, the less your clever scheme looks like a scheme and the more it looks like a puzzle.
Control requires the upper hand
Every durable control relationship we know runs on an advantage: physical force, legal authority, information the other party lacks, or the ability to turn the system off. Superintelligence, by definition, erodes those advantages. If it can model you better than you model it, your tests become theater. If it can act faster and broader than your oversight team, your kill switch is a story you tell yourselves until the day it is not available.
That claim is about the end state people keep funding toward, not a brief against every useful paper on today's models. The structural barriers stack: you cannot fully specify human values, you cannot reliably tell genuine compliance from performance, and the intelligence gap between evaluator and system grows as the system improves. Hoping the last barrier falls right before deployment is a deadline fantasy, not science.
If the product is smarter than the people holding the remote, the remote was never the point.
Dumb is a diagnosis, not a slur
Calling the plan dumb names the category error; it does not insult the researchers. We are pouring scarce safety money into making a successor species go well while the race to build that successor continues on schedule. That is like funding better manners for a flood while the dam is still being demolished upstream.
The smart move is boring: do not build the thing you cannot control. Narrow AI that folds proteins, reads scans, and forecasts weather can keep delivering. Those systems stay tools. Superintelligence is a new locus of agency, not a larger tool. Treating "align it" as the primary plan while labs race is how a culture talks itself into an irreversible experiment.
What to do with the money instead
Move the alignment-for-ASI pile toward work that makes non-creation real: treaty design and verification, compute governance you can inspect, public case-making that keeps the goal legible, and political pressure so legislators hear something other than lab talking points. I have written the funding case separately. The short version is sequence. After a verified halt on the path to superintelligence, research into powerful AI under tight control can continue. Before that halt, every dollar that makes the race feel manageable is a dollar not spent on ending the race.
You cannot align superintelligence in the sense the pitch deck needs. The word already told you why. Stop funding the fantasy that we will. Fund the decision not to build it.