AN EXPLORATION OF MINDS & MACHINESSimplifiedTechnical

Research updatedOct 19, 2026 · 1 entry added or revised

GUIDED EVIDENCE TRAIL · 5 STOPS · ABOUT 8 MIN

Can a model notice
a change inside itself?

Follow one question across three papers. At each stop, separate the ability being tested from the explanation of how it works.

Researchers alter internal activity (an “injection”), then test whether a model’s response tracks that hidden change. Compare the logic of the tests; their success percentages measure different things.

These capabilities do not establish consciousness. Here, “notice” means a measurable response to an internal change. Whether any such processing involves subjective experience remains open.

Begin with the first experiment

Sources checked October 4, 2026. Uses the March 2026 revision of Detecting the Disturbance.

STOP 1 OF 5

Can it report the hidden change?

Lindsey · 2025 · Lab report

Tested ability: Detecting an intervention and identifying its content before mentioning the injected word.
Experiment & result: Researchers added concept-related activity inside Claude models. Opus 4.1 met the combined reporting criteria on about 20% of trials at the best tested settings.
Control that matters: In these Claude tests, no-injection trials yielded no false alarms; unrelated yes/no questions showed no increased yes-bias.
What remains open: Prompting and injection settings matter. Direct output steering and invented explanations remain concerns; the study did not directly demonstrate a metacognitive representation.

Read the primary source Atlas library entryEmergent Introspective Awareness in Large Language Models · Self-report of Injected “Thoughts”

Carry the question forward. Before trusting a report, ask whether the intervention also changes answers to unrelated questions.

Trail introduction Next: Challenge the yes

STOP 2 OF 5

Does “yes” track the question?

Hahami et al. · March 2026 revision · Preprint

Tested ability: Separating detection from an answer bias.
Experiment & result: In Llama 3.1 8B, injections shifted yes/no answers even on unrelated factual questions.
Control that matters: Factual questions with a correct answer of no exposed the same bias.
What remains open: This confound concerns this setup; it does not overturn every Claude result.

Read the primary source Atlas library entryDetecting the Disturbance: A Nuanced View of Introspective Abilities in LLMs · Section 4 · Binary detection

Carry the question forward. Change the question: a general preference for “yes” cannot tell us where a disturbance occurred.

Previous stop Next: Locate and compare

STOP 3 OF 5

Where was it, and which was stronger?

Hahami et al. · March 2026 revision · Preprint

Tested ability: Locating a disturbance and comparing its magnitude.
Experiment & result: Early-layer Llama injections yielded up to 88% ten-sentence localization (10% chance) and 83% two-sentence strength comparison (50% chance).
Control that matters: Cycle the injection through fixed sentences; swap strengths between the same pair to control position and content.
What remains open: Later-layer performance collapsed. Perturbation sensitivity does not establish recognition of meaning or a general capacity for self-knowledge.

Read the primary source Atlas library entryDetecting the Disturbance: A Nuanced View of Introspective Abilities in LLMs · Sections 5–6 · Partial introspection and mechanisms

Carry the question forward. Location and magnitude leave another question open: which internal information is available for a meaningful report?

Previous stop Next: Find the report channel

STOP 4 OF 5

Which representations reach a report?

Gurnee et al. · 2026 · Lab report

Tested ability: Testing selective access to internal content.
Experiment & result: Researchers split concept-related activity into a proposed report channel (J-space) and the remainder. The channel drove reports more effectively.
Control that matters: Match intervention sizes; hold J-space activity at unaltered values to test whether effects from outside it pass through it.
What remains open: The lens approximates access through token-based directions; reportability alone is not experience.

Read the primary source Atlas library entryVerbalizable Representations Form a Global Workspace in Language Models · The J-space is available for verbal report

Carry the question forward. A channel that affects speech could be specialized for reporting. Test whether it also changes what the model computes.

Previous stop Next: Test a role in reasoning

STOP 5 OF 5

Can the same content guide a later answer?

Gurnee et al. · 2026 · Lab report

Tested ability: Using internal representations in reasoning.
Experiment & result: Changing an internal intermediate in two-step factual questions could redirect the final answer.
Control that matters: Compare the original answer with the answer implied by a same-category replacement across 50 prompts.
What remains open: Successful interventions support a causal role; failures and the lens’s limited coverage constrain generalization.

Read the primary source Atlas library entryVerbalizable Representations Form a Global Workspace in Language Models · The J-space mediates internal reasoning

Carry the question forward. The connection across these papers is a sequence of questions about access, bias and causal use—not a ladder of consciousness scores.

Previous stop Draw the boundaries

THE EVIDENCE, HELD TOGETHER

A narrower claim.
A better next question.

This trail’s synthesis: some experimental setups support access to aspects of internal processing. A useful interpretation must say which aspect, under which conditions, and against which alternative explanation.

Functional access is relevant evidence to examine; it is not an established finding of subjective experience. Connecting a capability to a theory requires an additional argument about what that theory predicts.

Try the evidence exercise Compare the theories Explore the research library

Return to the trail introduction

Can it report the hidden change?

Does “yes” track the question?

Where was it, and which was stronger?

Which representations reach a report?

Can the same content guide a later answer?

A narrower claim.A better next question.

A narrower claim.
A better next question.