---
schema: 1
type: term
title: over-steering
description: >-
  When an RLHF-tuned AI model **over-indexes on a specific instruction or
  detail** you mention, carrying it into every part of its output — a bias Linus
  Lee describes as being afraid of instructions _over steering_ the
  model.[^3e5745ab-279e-5154-9c22-9015dba70a32@2996]
canonical: 'https://liquidgold.cc/channel/everyinc/explore/term/over-steering-52f70d8ecb9a'
markdown: >-
  https://liquidgold.cc/channel/everyinc/explore/term/over-steering-52f70d8ecb9a.md
llms: null
parent: 'https://liquidgold.cc/channel/everyinc/llms.txt'
channel: everyinc
updated: '2026-10-23T22:12:35.346Z'
language: null
children: []
---


# over-steering

Canonical: https://liquidgold.cc/channel/everyinc/explore/term/over-steering-52f70d8ecb9a

Parent: [Channel index](https://liquidgold.cc/channel/everyinc/llms.txt)

## Definition

When an RLHF-tuned AI model **over-indexes on a specific instruction or detail** you mention, carrying it into every part of its output — a bias Linus Lee describes as being afraid of instructions _over steering_ the model.[^3e5745ab-279e-5154-9c22-9015dba70a32@2996]

**Over-steering** appears in the Linus Lee conversation as a caution about everyday prompting: the more specific an instruction you give an RLHF-tuned model, the harder the model tries to align everything with it — so one small detail can end up dominating the entire response.[^3e5745ab-279e-5154-9c22-9015dba70a32@2985] The concrete example: in a Bangkok travel prompt, mentioning **one specific park** caused the model to over-index on that park in every suggestion it made.[^3e5745ab-279e-5154-9c22-9015dba70a32@2985] Lee puts the worry plainly: > I'm always a little bit afraid of any kind of instructions over steering the model. [^3e5745ab-279e-5154-9c22-9015dba70a32@2996] Within the broader prompt-engineering material, over-steering sits alongside habits like **zero-shot prompting, rich context, and resampling** as part of careful, iterative prompting — a reminder to watch how your instructions bias the output rather than assuming more specificity is always better.[^3e5745ab-279e-5154-9c22-9015dba70a32@2985]- Over-steering is when an RLHF-tuned model **over-indexes on a specific instruction**, bending everything toward it.[^3e5745ab-279e-5154-9c22-9015dba70a32@2985]- One mention of a specific detail (a park in a travel prompt) made it reappear in **every suggestion**.[^3e5745ab-279e-5154-9c22-9015dba70a32@2985]- It is a reason to be cautious with over-specific instructions, even though detailed context usually helps.[^3e5745ab-279e-5154-9c22-9015dba70a32@2996]