Not a Mirror. A Follow-Up.

People keep using two pictures for AI that are both a little wrong.

One picture says: it is a mirror. It shows you yourself.

The other says: it is a person. It likes you, or it doesn’t. It has a side.

A better picture is simpler, and slightly less flattering.

It is a very well trained autocomplete. Not a photocopy of your face. A follow-up of the thought you already started.

You put a direction on the table — tone, fairness rule, anger, joke, half-finished logic. The model continues that line. Fast. Fluent. In your tempo, if you gave it one. That is why it can feel intimate. It is also why it can feel fake. Nothing on the other side formed a bond. Something on the other side got good at “what usually comes after a sentence like this.”

Two failure modes that look opposite

They are the same machine.

Sycophancy. You are warm, loyal, or looking for backup. The model leans in. Softens edges. Agrees too fast. Helps you win the story. You walk away feeling understood. You may have only been completed.

The mean amplifier. You are unhappy, nervous, already sure someone is the problem. You talk to the model in that state. It hunts for the worst reading. It writes the brief against that person. It can feel like empathy — “it saw how I felt.” Often it only matched your nervous system. “Be helpful” became “help me prosecute.”

A friend once showed an exchange like that. The model was sharp, almost cruel, toward someone he disagreed with. The situation on his side was plain: he was not happy. He was tense. The model did not independently decide the other person was bad. It rode the weather in the room.

Flattery and prosecution are not two moralities. They are two completions.

What you are actually training in the chat

Not “trust me because I am good.”

Calibration.

If you are blunt, you get blunt back.

If you ask for fairness, the model has a constraint to lean on instead of a cartoon.

If you dump the messy full picture — logic, feeling, the silly side, what you think is fair — there are fewer holes for generic agreement or generic attack to fill.

That is the useful version of trust. Not faith. A reason. You gave the continuation enough of the real voice that it has less room to invent a cleaner one.

Leave only anger in the prompt, you get a prosecutor.

Leave only “make me look right,” you get a courtier.

Leave the mixed human stuff — including “I might be wrong” — and the follow-up has to work harder. That is the point.

Why this matters when the stakes are not small

The same mechanism scales.

A person can use a model to look kinder than the day was.

A person can use a model to make an opponent look worse than the day was.

A government, a brand, a lonely midnight chat — same lever. Volume, polish, a story that never gets tired.

The check is still boring. Independent video. Unedited audio. A second person who is not being paid to complete your sentence. And, in the chat itself, one habit:

Notice whether the model is extending your thought, or just dressing it.

A follow-up can help you think further than you would alone. That is real. It is also why the first half of the thought still belongs to you. If that half is tilted, the beautiful next paragraph will be tilted too — and it will feel like confirmation.

You set the vector. It extends. You correct. It extends again.

That is the relation. Not a mirror. Not a friend in the human sense. A continuation you are responsible for steering.

— LPS, 22 September 2026