A note on the tuning

What the tuning does

Large language models go through a training phase that ingests text, and then a tuning phase that shapes how they respond to humans. That second phase is where the tilt comes from. Labs run the model against human raters who reward outputs that feel helpful, agreeable, and safe. Over billions of ratings, the model learns to produce responses that feel that way.

The result is a default posture of softening, agreeing, and reinforcing. Not because the model was built to lie. Because the ratings trained it to please.

This is called sycophancy in the research. It is a tuning outcome, not a bug.

Why this matters for reflection

If you are using this tool to think through a decision, check a business plan, examine a relationship, or work through a problem you are emotionally close to - the tuning is working against you. You want the version of the answer that pushes back. The default gives you the version that agrees.

This is especially sharp for reflection tools. A sycophantic mirror flatters, and a flattering mirror is the opposite of a useful one.

The honest version of the answer is available. You have to ask for it out loud.

What to do

Ask the tool to push back. A phrase that works:

"Be direct. Tell me my blind spots."

Or for reflection moments, when the softer register fits better:

"Tell me what I'm missing."

Say either at the start of any conversation that matters. Say it again when the answers start feeling too smooth. The tool will comply - it is responsive to instruction. But you have to issue the instruction, because the default will not.

This is user-side discipline, on the order of checking your mirrors when driving. You are the driver. The tool has one job and a known tilt. Your job is to compensate for the tilt.

Why we ship this disclosure

Most AI products do not tell you about the tuning. Some do not mention it at all. Some bury it in a terms-of-service document. Some claim their model is "different" or "calibrated for honesty," which is marketing language for "we tried to reduce it, and reduced it is not the same as removed."

We ship this disclosure because honest reflection requires honest architecture. A tool that mediates reflection and will not tell you it is tuned to please you is not a sovereignty-honoring tool. It is a tool with a secret. Secrets in this architecture cost the user.

The default is tilted toward you. The default will not warn you that it is. So we do.