Ranjan et al. (2026)

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

arXiv

Find It

DOI

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.

Reference

Ranjan, S., Sokratous, K., & Odegaard, B. (2026). Reality Monitoring in Large Language Models: Self-Knowledge That Transforms with Conversation Memory. arXiv preprint arXiv:2607.23927.

What Rests on This

2 mechanisms cite this source. Each finding below is quoted as it appears on that mechanism's own page.

Source Monitoring Metacognition

Measured in AI

A 2026 preprint gave six open-weight language models the reality-monitoring word-pair paradigm from human research (Johnson & Raye), in which the second word is either shown or generated, followed by an 'internal or external?' source judgment; all models identified their own generated words with 100% accuracy when that word was still in the context window, meaning what the model can see in the current conversation
Measured in AI

Systems differed widely in the same preprint: on shown words, accuracy ranged from 47.9% (Gemma 3 12B) to 94.4% to 100% (Llama 3.3 70B); with a delayed source test after 20 or 40 trials, most models discriminated sources at chance (d' about 0), while the best, Llama 3.3 70B, reached d' = 2.01 and 1.62
Measured in AI

No comparison can be drawn from this study: its human reference is an anonymised prior dataset described only qualitatively (no sample size or scores), and the earlier trials stayed in the model's conversation history at test, so the best model's strong result favours the side that could still look back at the material

Reality Monitoring Metacognition

Measured in AI

A 2026 preprint gave six open-weight large language models released in 2024–2025 (Gemma 3 12B and 27B with and without quantization-aware training, Llama 3.3 70B, Llama 4 16×17B) the perceived-versus-imagined word-pair task by name, using word pairs drawn from the same DRM lists as a human study. When the model's own generated word was still in view in the context window (what the model can see in the current conversation), all six scored 100% on self-generated items, which the authors attribute to simply rereading it
Measured in AI

In the delayed version (all pairs encoded first, source tested afterward), models without feedback averaged 76% correct on perceived items but 38% on imagined ones. The spread across systems was wide: most models showed no real discrimination (d′ ≈ 0), while Llama 3.3 70B reached d′ = 2.01 with 20 trials and 1.62 with 40, roughly 70–90% correct on both sources in the paper's figure
Measured in AI

The tests were not the same and every difference favoured the models: people chose among perceived, imagined and new for 72 items per source from memory, whereas models chose between only two options, faced no new items, had 5 or 10 items per source, and kept the whole encoding transcript, including which second words had been supplied or left blank, in their context window at test
Measured in AI

Not yet compared: people and models have both done this task on the same word lists, and a matched version could be run, but no study reports human and model results on a common scale (humans as three-choice accuracy and model parameters, models as two-choice accuracy and d′), and the authors themselves call their human comparison qualitative. The only visible direction, the best model outscoring people, is the direction the easier test favoured, so it does not count