---
schema: 1
type: term
title: _smuggled intelligence_
description: >-
  When evaluating how smart an LLM is, the unrecognized leakage of the
  evaluator's own intelligence into the test — through who prompted it, what
  inputs they gave it, and how it was set up — so the measured capability is
  partly the evaluator's, not the model's. Coined by Alex Komoroske as the
  mirror image of _smuggled infinities_.
canonical: >-
  https://liquidgold.cc/channel/everyinc/explore/term/smuggled-intelligence-f29b8d32b891
markdown: >-
  https://liquidgold.cc/channel/everyinc/explore/term/smuggled-intelligence-f29b8d32b891.md
llms: null
parent: 'https://liquidgold.cc/channel/everyinc/llms.txt'
channel: everyinc
updated: '2026-10-23T22:12:35.346Z'
language: null
children: []
---


# \_smuggled intelligence\_

Canonical: https://liquidgold.cc/channel/everyinc/explore/term/smuggled-intelligence-f29b8d32b891

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

## Definition

When evaluating how smart an LLM is, the unrecognized leakage of the evaluator's own intelligence into the test — through who prompted it, what inputs they gave it, and how it was set up — so the measured capability is partly the evaluator's, not the model's. Coined by Alex Komoroske as the mirror image of _smuggled infinities_.

**_Smuggled intelligence_** is Alex Komoroske's coinage — the mirror image of his friend Ben Mathysse's frame of *smuggled infinities* [^bb2c42aa-478f-5c9a-ac26-e83a54c63127@4009]. Where a *smuggled infinity* breaks an argument by importing something impossible (like an agent that is *perfect*), a *smuggled intelligence* corrupts an evaluation by importing something unaccounted for: **the evaluator's own intelligence**.In the interview, the term names a trap to avoid when judging **how smart an LLM is**: beware smuggling your own intelligence into the test — *who prompted it, what they gave it, and how it was set up* all shape the output [^bb2c42aa-478f-5c9a-ac26-e83a54c63127@4086]. Komoroske emphasizes that the most important input to an LLM's output is the **user**, and that prompting skill is a real, distinct intuition with little to do with technical savviness. The practical upshot he draws: if you are not actively seeking **disconfirming evidence**, the model will happily hand you confirming evidence for anything.- _Smuggled intelligence_ = your own intelligence leaking into an LLM evaluation via prompting, setup, and inputs.- It pairs with *smuggled infinities* as a two-part guard against over-crediting agent capabilities.- Prompting skill is real and distinct from technical savviness — it is part of what you are measuring.- Seek **disconfirming evidence**, because the model will not volunteer it.