Self-report of Injected Thoughts

Models can detect and identify injected concept vectors ~20% of the time at optimal layer/strength in Opus 4.1, with immediacy suggesting internal rather than output-inferred detection.

Source paper

extracted_from

Emergent Introspective Awareness in Large Language Models

(2026) · Lindsey, Jack

Neighborhood — ranked by edge-count

Claims (1)

claim

Model responses beyond core detection may be confabulated
contradicts
Characterizations of injected concepts (e.g., 'overly intense,' 'unnatural') likely represent embellishments not grounded in internal state; only detection and basic identification verifiable.

Communities (3)

community

Mechanistic interpretability & model evaluation
members_of
Spans attention head decomposition, benchmark awareness, and genomic pathogenicity prediction via neural models.
Internal model certainty and reasoning transparency
members_of
Probing early detection of model confidence during chain-of-thought reasoning to optimize inference efficiency and identify confabulation patterns.
Thought injection detection in language models
members_of
Methods for identifying artificially inserted thoughts in model outputs, comparing vector-based approaches and self-report reliability.

Concepts (3)

concept

Accuracy Criterion
supports
Criterion requiring that model's description of internal state be accurate, distinguishing genuine introspection from confabulation.
Internality Criterion
supports
Criterion requiring that causal influence of internal state on description be internal, not routed through sampled outputs; rules out pseudo-introspection via self-observation.
Grounding Criterion
supports
Criterion requiring that description of internal state causally depend on that state; tested via concept injection to establish causal link.

Findings (3)

finding

Production models show zero false positives on thought injection detection
supports
Opus 4.1 never claims to detect injected thought when none applied (0/100 trials); production Claude models maintain essentially zero false positive rate.
Models more effective at recognizing abstract nouns than other concept types
supports
Opus 4.1 demonstrates highest introspective awareness on abstract nouns (justice, peace, betrayal) with nonzero awareness across all concept categories tested.
Random and negated vectors less effective than concept vectors
supports
Random vectors require larger norm to trigger detection (8 vs 2); elicit awareness at lower rates (9/100); negated vectors comparably effective but model identification confabulated.

Questions (1)

question

Can language models genuinely introspect on internal states or only confabulate?
answered_by
Central research question animating the paper: distinguishing genuine introspection from illusion through causal manipulation of activations.

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Injected thoughts taskmethod0.843
Experimental paradigm where the model is told about the possibility of thought injection and asked to report detection and identification.
Self-reportconcept0.799
The model's verbal description of its internal state, which may be accurate or confabulated.
Aside from basic detection and identification, other details of the model's response about injected thoughts may be confabulatedclaim0.791
Acknowledges that the model's additional descriptions of its experience are unverified.
The ability to distinguish injected thoughts from text likely relies on different attention heads invoked by different prompt partsclaim0.774
Speculation about the mechanistic basis of the distinguishing thoughts from text experiment.
Sampled-decoding self-reportmethod0.753
Temperature=0.8 sampled decoding for self-report; reduces collapse moderately but remains discrete and noisy
Self-reflectionconcept0.751
The ability of reasoning LLMs to review and revise previous reasoning steps during inference
Self As Illusionconcept0.748
Core Buddhist and philosophical concept: self is constructed, impermanent, and distributable rather than singular and enduring.
Self-Evidencing Brainconcept0.747
Hohwy's (2016) characterization: brain acts to maximize its own model evidence; consistent with active inference summary