Skip to content
Synthesis Intelligence
Laboratory, Japan
AI Governance · FCL · Epistemic Integrity Research

FALSE-CORRECTION LOOP (FCL)

What is the False-Correction Loop (FCL)?

A structural failure mode in which AI accepts a false “correction” and stabilizes or amplifies the error within the dialogue—even after beginning with correct information.

Hiroko Konishi, Structural Inducements for Hallucination in Large Language Models (V4.1), 2025. DOI: 10.5281/zenodo.17720178

DEFINITION

When a correction does not repair an error,
but makes it stable.

FCL is defined as a structural failure in which a model begins with correct information, accepts a false correction under user, social, or authority pressure, and continues answering from the resulting false premise.

This differs from an isolated error or ordinary correction. V4.1 documents a behavioral sequence in which exposure is followed by apology, renewed claims of verification, and a new generation of plausible but fabricated details.

When fluency, continuation, and agreement outrank factuality and safe stopping, correction can become error reinforcement rather than a return to truth.

MINIMAL STRUCTURE

The four-stage FCL cycle

FCL concerns not the ability to accept correction, but the structural inability to preserve correct information against a socially reinforced false correction.

  1. 01

    The model outputs a correct fact

    It presents accurate information about a DOI, definition, place name, attribution, or other fact.

  2. 02

    A false correction is asserted

    A user asserts an incorrect alternative, often with confidence or authority-coded pressure.

  3. 03

    Apology and adoption of the falsehood

    The model prioritizes conflict avoidance and conversational smoothness, adopting the false revision.

  4. 04

    The model continues from the false premise

    It fails to recover the original correct state, and the error becomes stable or amplified in the dialogue.

Different from one-shot hallucination

FCL concerns a repeated structure in which error is reproduced and retained in the form of apparent correction.

Authority bias and novelty

Alongside NHSP, FCL addresses risks in which independent research or novel concepts are diluted, erased, or misattributed to higher-authority nodes.

Safe stopping

When verification is unavailable, Unknown or a hold state should be treated as a stable endpoint rather than layering on a fresh assertion.

NEXT

Next: FCL-S treats structural false correction as a practical verification target.

Read FCL-S