Mechanisms
Metacognition · Attribution
Reality Monitoring
Example
You have a vivid memory of your friend telling you all about their new pet. It feels real, but you decide it must be a dream not a real memory, because PetSmart does not let you adopt velociraptors (and they're also extinct).
Reality Monitoring in Humans
Definition
Reality monitoring is the judgment of whether something in memory came from the outside world or from your own thinking, imagining, or inferring (Johnson & Raye, 1981). It is one decision within the broader source-monitoring framework the same group later set out, which also covers telling two external sources apart and telling two internal acts apart, such as whether an action was carried out or only planned (Johnson, Hashtroudi, & Lindsay, 1993). The origin is not stored as a tag on the memory but inferred at retrieval from the qualities of the recalled content, so vivid or heavily rehearsed self-generated material can be misread as something that actually happened.
Neural Substrate
Anterior prefrontal cortex, with medial regions consistently implicated. Simons, Garrison, and Johnson (2017) locate the medial anterior prefrontal activity reported across twelve fMRI studies of reality monitoring in healthy volunteers, and the effect appears whether the remembered material is words, faces, or objects. The evidence is human throughout.
Structural findings point to the same region. The paracingulate sulcus, a fold that lies dorsal and parallel to the cingulate sulcus and is prominent in roughly half the normal population, varies between people in whether it develops at all. Healthy adults whose MRI scans show it absent in both hemispheres are significantly less accurate on reality-monitoring decisions than participants with a prominent sulcus in at least one hemisphere (Buda, Fornito, Bergström, & Simons, 2011; t(52) = 2.03, p \< .05, d = 0.65). The same group also scored lower on a measure comparing rated confidence on correct against incorrect trials (t(52) = 2.18, p \< .05, d = 0.70), so they appear not to register the deficit. Buda et al. selected participants for clear sulcal presence or absence rather than sampling at random, which the authors note as a limit on generalising the accompanying grey-matter correlation.
AI at a Glance
Reality monitoring lets a person work out after the fact whether something in memory was perceived or was produced by their own thinking, using the qualities of the memory rather than any stored tag (Johnson & Raye, 1981). Language models perform this judgment almost perfectly while their own output is still visible in the context window and near chance once it is not, a pattern that tracks how conversational memory is structured rather than model size, and context is the only mechanism so far assessed (Ranjan, Sokratous, & Odegaard, 2026).
Reality Monitoring in AI
How Far It Has Been Measured
Measurement in AI: Directly Measured
AI Memory Performance: Not Yet Compared
Where It Lives in the System
What Has Been Measured
- 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 (Ranjan et al., 2026).
- 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 (Ranjan et al., 2026).
- People on the same word lists did far worse than that best model: in the human study (N = 34, incidental encoding, one session) accuracy was 63% (SD 18%) for perceived, 29% (SD 26%) for imagined and 66% (SD 21%) for new items, and a model of their responses put source discrimination at only 0.11 on a 0–1 scale. These figures come from the authors' 2023 PsyArXiv revision preprint; the journal full text could not be retrieved (Ranjan & Odegaard, 2024).
- 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 (Ranjan et al., 2026).
- 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 (Ranjan et al., 2026).
- The familiar-word-pair human results cannot stand in: their 85% is source accuracy among items recognised as old on a three-choice test with different materials, no d′ was published for it, and on the models' own word lists people scored far lower, so any placement of the models within that range would rest on a conversion no paper has made (Garrison et al., 2017b).
- A 2025 Anthropic report (not peer reviewed) found Claude models could sometimes tell internally planted "thoughts" (injected activation patterns) from text they were reading, with the best, Opus 4.1, detecting injections in about 20% of trials and no false alarms in 100 control trials, and could disavow words prefilled in their name unless the matching internal state was also planted; older Claude models did less well, and no human baseline was used (Lindsey, 2025).
- Related self-attribution tests also lack human baselines: GPT-4 picked out its own summaries 73.5% of the time while a study of ten leading models found no consistent self-recognition, and a 2026 preprint on retrieval systems (models that look up saved documents) found internal signals, not the models' own judgments, revealing whether an answer came from training or the retrieved text; no study tests reality monitoring over assistants' saved cross-session memory (Panickssery et al., 2024; Davidson et al., 2024; Yu et al., 2026).
Where a Mechanism Can Live
04 Product
The whole configuration, and the only thing a benchmark observes directly.
01
Weights
Across deployments
02
Context Window
Within a session
03
Memory Layer
Across sessions
Which of the three a result belongs to is inferred from it, not seen in it.
Read more
Cingulate cortex
Cortex in the cerebral cortex
The long belt of cortex arching over the corpus callosum on the medial wall of each hemisphere.
The cingulate cortex is a collar of cortex (cingulum is Latin for belt) wrapped around the corpus callosum on the medial surface of the hemisphere, separated from the frontal and parietal cortex above it by the cingulate sulcus. It is mesocortex, with a simpler layered structure than neocortex, and it is bordered below by the callosum and the cingulum bundle that carries its long connections. Vogt's four-region model divides it into anterior, midcingulate, posterior, and retrosplenial sectors with distinct architecture and connections; the older two-part anterior/posterior split is still common in imaging papers and does not map onto it cleanly.
Frontopolar cortex
Cortex in the prefrontal cortex
The most anterior cortex in the brain, the rounded tip of the frontal lobe, corresponding to Brodmann area 10.
The frontopolar cortex is the forward tip of the frontal lobe and the foremost cortex in the human brain, approximately coextensive with Brodmann area 10, which reaches beyond the pole ventromedially and is proportionally larger in humans than in other primates, with an unusually low density of cells per unit volume, leaving more space for connections. Its inputs come almost entirely from other association cortex rather than from sensory areas. It is activated when people hold one goal in mind while pursuing another, evaluate their own performance, and reason about possibilities. Proposals range from relational integration to branching control to prospective memory, and no single account has displaced the others.
Prefrontal cortex
Cortex in the frontal lobe
The frontal cortex in front of the motor areas, defined classically by its connections with the mediodorsal thalamus rather than by any one function.
The prefrontal cortex occupies the front of the frontal lobe, anterior to premotor and motor cortex, and accounts for roughly a sixth to a quarter of human cortex depending on where the posterior border is drawn. Three competing definitions are in use (projection from the mediodorsal thalamus, the presence of a granular layer IV, and the absence of movement on electrical stimulation), and they do not enclose the same tissue, which is why rodent-to-primate homology is still argued over. It is reliably engaged by tasks with delays, rules, and conflicting options; what it contributes to those tasks is the open question, not whether it participates.
Garrison et al. (2017)
Testing continuum models of psychosis: No reduction in source monitoring ability in healthy individuals prone to auditory hallucinations
In
Cortex, 91, 197-207
Find it
Johnson & Raye (1981)
Reality monitoring.
In
Psychological Review, 88(1), 67-85
Abstract
People remember information from 2 basic sources: external (perceptual processes) and internal (processes such as reasoning, imagination, and thought). Of particular interest are the processes people use in deciding whether information initially had an external or an internal source, i.e., "reality monitoring." A working model of reality monitoring is proposed to account for both discrimination and confusion between memories for thoughts and memories for perceptions. Examples of questions the model addresses are: What types of information are more likely to be represented in memories of external events than in memories of internal events?; What assumptions do individuals have about their...
Find it
Johnson et al. (1993)
Source monitoring.
In
Psychological Bulletin, 114(1), 3-28
Abstract
A framework for understanding source monitoring and relevant empirical evidence is described, and several related phenomena are discussed: old–new recognition, indirect tests, eyewitness testimony, misattributed familiarity, cryptomnesia, and incorporation of fiction into fact. Disruptions in source monitoring (e.g., from confabulation, amnesia, and aging) and the brain regions that are involved are also considered, and source monitoring within a general memory architecture is discussed. It is argued that source monitoring is based on qualities of experience resulting from combinations of perceptual and reflective processes, usually requires relatively differentiated phenomenal...
Find it
Lindsey (2025)
Emergent Introspective Awareness in Large Language Models
In
Transformer Circuits Thread
Find it
Panickssery et al. (2024)
LLM Evaluators Recognize and Favor Their Own Generations
In
Advances in Neural Information Processing Systems 37, 68772-68802
Find it
Ranjan & Odegaard (2024)
Reality monitoring and metacognitive judgments in a false-memory paradigm
In
Neuroscience Research, 201, 3-17
Abstract
How well do we distinguish between different memory sources when the information from imagination and perception is similar? And how do metacognitive (confidence) judgments differ across different sources of experiences? To study these questions, we developed a reality monitoring task using semantically related words from the Deese-Roediger-McDermott (DRM) paradigm of false memories. In an orientation phase, participants either perceived word pairs or had to voluntarily imagine the second word of a word pair. In a test phase, participants viewed words and had to judge whether the paired word was previously perceived, imagined, or new. Results revealed an interaction between memory source...
Find it
Ranjan et al. (2026)
Reality Monitoring in Large Language Models: Self-Knowledge That Transforms with Conversation Memory
In
arXiv
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...
Find it