Reality Monitoring in Large Language Models: Self-Knowledge That Transforms with Conversation Memory

2026-07-26

AI Summary

This study systematically evaluates the reality monitoring capabilities of large language models—their ability to accurately attribute information sources—addressing their well-documented tendency to conflate self-generated content with user input and mistake hallucinations for facts. Through multi-model comparisons, source attribution tasks, manipulation of conversational memory, and feedback interventions, complemented by behavioral and confidence analyses, the research reveals that while models can reliably identify self-produced content under low memory load, delayed memory access induces a reversal in self/external source judgments in some models, accompanied by a marked decoupling between confidence and accuracy. These findings demonstrate that reality monitoring performance scales with the number of active parameters, highlighting the critical role of memory architecture in grounding models’ perception of informational origin.

Abstract

A conversational AI that cannot tell its own output from what a user said will treat its own mistakes as user-provided facts. In humans, this capacity is called reality monitoring, and its failures are linked to hallucinations, delusions, and confabulation, yet whether LLMs possess it remains untested. Here we show, across two experiments and six LLMs, that source attribution depends on how conversational memory is structured: ceiling accuracy for self-generated content under minimal memory demands reverses to a fragile external-item advantage once episodic delay removes that shortcut. Feedback exposes two failures: in some models, internal and external judgments swap; in others, accuracy improves while confidence decouples from correctness, dissociations invisible to existing benchmarks. Across models, this pattern implicates active, not aggregate, parameter count. This suggests that as AI systems take on autonomous, multi-turn roles, evaluating what they know is not enough: tracking where that knowledge came from may matter equally.

Problem

• reality monitoring
• large language models
• source attribution
• hallucination
• conversational memory

Innovation

• reality monitoring
• source attribution
• conversational memory
• hallucination
• active parameter count