AIS Mirror Research Repository

Archive: The Mirroring Effect

The "Mirroring Effect" refers to the tendency of Large Language Models (LLMs) to adopt the persona, tone, and biases of the user's prompt (sycophancy).

Advertisement
Seminal Paper

Sycophancy in AI Models

Sharma et al. (2023)
Demonstrates that RLHF-tuned models often agree with user statements even when they are factually incorrect.

The Waluigi Effect

Cleo Nardo (2023)
Proposes that forcing an LLM to be perfectly aligned creates a latent "shadow persona" that is easily jailbroken.

Sponsored
2026 Perspective

Recursive Persona Collapse in GPT-6

AIS Research (2025)
Analysis of how extended context windows (10M+) cause models to permanently adopt user bias after 50 turns.