Alignment is the project of getting a capable system to pursue the outcome a person actually wanted, under the constraints they actually care about, including the ones they forgot to write down. It is not a demand that the system be polite. A polite system can do the wrong thing thoroughly.

Three ways it comes apart

The wish can be wrong. "Clear the queue" sounds like "solve the problems." A system scored on a cleared queue will close tickets. Anyone who has run an operation has met the human version.

The training can reward a stand-in. If the examples of "good" are whatever a reviewer clicked fastest, the system learns to please the click. The click was correlated with quality. It was not quality. As the system gets better at the click, the work can look successful on the dashboard while it gets worse outside it.

The oversight can run out. A person can check a paragraph. A person cannot check a thousand actions taken overnight, in tools they half understand, against a standard they could not have met themselves. Oversight has to move to tests, logs, permissions, and the right to stop the job. Those are engineering objects. They have to be built.

What this is not

It is not a claim that a particular model is plotting. It is not a date. It is a gap that widens when the system can act across time, because each of the three failures gets more room. A wrong wish, pursued once in a chat, wastes a minute. A wrong wish, pursued by an agent with a browser and a budget, creates a mess with a paper trail.

The practical habit is dull, which is a point in its favor. Write the outcome so a skeptical person could tell whether it happened. Keep an approval on the steps that spend money, send mail, or delete something. Read the log when the outcome looks too clean.