Entry reviews: September 6–October 1, 2026 · Simplified explanations
Publication type describes the linked record. Journal publications have undergone journal peer review; a preprint label does not rule out a later published version. The separate work-type tag identifies experiments, theoretical proposals, reviews and arguments. Year filters include both years when an entry spans online publication and an issue date, or a preprint revision.
01The possibility, and how to assess it
2025 / 26Trends in Cognitive SciencesJournal publicationAssessmentEntry reviewed
Patrick Butlin, Robert Long, Tim Bayne et al. · Online November 2025; issue June 2026
Question
Which properties could inform assessments of AI consciousness?
Method
Synthesizes consciousness theories into an indicator-based assessment framework.
Finding / proposal
Proposes looking inside AI systems for properties predicted by consciousness theories. Multiple lines of evidence can inform an assessment even when no single test settles the question.
Limitation
An assessment framework, not a validated consciousness detector.
Patrick Butlin, Robert Long, Eric Elmoznino et al.
Question
Can neuroscience guide the assessment of artificial systems?
Method
Surveys theories, translates proposed indicators into computational terms, and examines AI architectures.
Finding / proposal
The influential starting point for the indicator approach surveys several consciousness theories and examines AI architectures. It finds no obvious technical barrier to building systems with the proposed features.
Limitation
Its assessment of then-current systems dates to 2023; it is not an audit of 2026 models.
What obstacles stand between language models and consciousness?
Method
Philosophical analysis of candidate requirements and present or future model architectures.
Finding / proposal
Examines reasons for and against conscious language models. Identifies obstacles in the systems of the time while arguing that future successors deserve serious consideration.
Limitation
A philosophical assessment of possibilities, not an experimental demonstration.
What makes a consciousness test valid beyond typical adult humans?
Method
Conceptual review of tests and the evidence needed to generalize them to unfamiliar populations.
Finding / proposal
Asks what would make a test of consciousness trustworthy across people, animals and unfamiliar systems. A test that works in one setting may not automatically transfer to another.
Limitation
Behavioral or neural markers are not universally validated across substrates.
Under which realization conditions could a simulated mind be conscious?
Method
Develops a conditional causal-computational account of implementation and preserved internal organization.
Finding / proposal
Argues that a simulation could genuinely support consciousness if it physically realizes the right internal organization. Matching outward behavior is insufficient: internal mechanisms and responses to interventions also matter.
Limitation
A conditional theoretical argument; whether consciousness is invariant under the proposed organization remains an open premise.
How could the properties of experience be expressed in physical terms?
Method
Derives physical postulates from proposed experiential axioms and formalizes cause–effect structure.
Finding / proposal
Defines consciousness in terms of a system’s intrinsic, irreducible cause–effect structure. The theory can in principle apply beyond biology, but the physical implementation matters.
Limitation
A disputed theoretical identity; this paper does not establish conscious AI.
Do preregistered predictions of IIT and GNWT survive direct comparison?
Method
Human visual-perception experiment combining fMRI, MEG and intracranial EEG across 256 participants.
Finding
A preregistered study of 256 people tested predictions from two leading theories. Some results fit each theory, while other results challenge important predictions of both.
Limitation
Tests specific biological predictions; it neither settles the theories nor tests AI consciousness.
Might biological processes be necessary to explain consciousness?
Method
Develops biological naturalism and challenges computational sufficiency through theoretical analysis.
Finding / proposal
Makes the case that being alive may matter to consciousness. It challenges the assumption that reproducing a computation is enough, while considering more brain-like or life-like artificial systems.
Limitation
A reasoned alternative; it does not prove that nonbiological consciousness is impossible.
Andres Campero, Derek Shiller, Jaan Aru & Jonathan Simon
Question
Which objections concern theory, engineering difficulty, or impossibility?
Method
Classifies objections and constraints by explanatory level and argumentative force.
Finding / proposal
Separates three often-confused claims: a theory may be wrong, conscious digital systems may be hard to build, or such systems may be impossible. This makes the debate more precise.
Limitation
A taxonomy of arguments, not evidence selecting a winning theory.
How can consciousness indicators be calibrated and transferred between substrates?
Method
Conceptual analysis of calibration, transfer assumptions and comparative validation across physical, causal and dynamical differences.
Finding / proposal
Two questions need separating: whether consciousness probabilities are reliable, and whether evidence from brains carries over to machines. The revised article proposes comparing systems with different materials, causal structures and dynamics to test those transfer assumptions.
Limitation
Peer-reviewed commentary with no new empirical data. The proposed comparisons neither calibrate consciousness probabilities nor establish biological necessity. This updates the existing March preprint record.
Can models report experimentally introduced internal changes?
Method
Activation interventions with controlled reporting tasks and comparisons across models and conditions.
Finding
Some tested models could sometimes report experimentally introduced changes to internal activity. This suggests limited access to internal information, with substantial unreliability.
Limitation
Introspection in the experimental sense does not establish subjective awareness.
Ely Hahami et al. · First posted December 2025; revised March 2026
Question
Does a model identify an injected concept or only detect a disturbance?
Method
Concept-injection experiments separating perturbation detection, strength and semantic attribution.
Finding
A follow-up investigates whether models detect a disturbance, its strength, or its specific cause. It helps separate different abilities that can all look like introspection.
Limitation
Narrow, prompt-sensitive effects; the preprint title changed during revision.
Can a model predict its own behavior better than an external predictor?
Method
Fine-tunes self-predictors and comparison models, then evaluates their behavioral predictions.
Finding
Models trained to predict their own behavior can outperform other models trained to predict them. The result raises questions about what information a model can access about itself.
Limitation
Self-prediction accuracy is not a measure of felt experience.
Which internal pathways support selected language-model computations?
Method
Cross-layer transcoders, attribution graphs and interventions on selected model features.
Finding
Traces internal computational pathways involved in tasks such as multilingual reasoning and planning a rhyme. It shows how to investigate mechanisms behind outputs rather than relying on what a model says.
Limitation
“Biology” is an analogy; these circuit studies do not test consciousness.
How is an adult fruit fly brain wired at synaptic resolution?
Method
Large-scale electron-microscopy reconstruction and proofreading of neurons and chemical synapses.
Finding
Maps 139,255 neurons and tens of millions of chemical synapses in an adult female fruit fly’s brain. It gives researchers a detailed view of how a biological network is connected.
Limitation
A fly connectome is not a human brain map or a demonstration of consciousness.
How do activity and wiring relate across mouse visual cortical areas?
Method
Combines in vivo functional imaging with dense electron-microscopy reconstruction.
Finding
Connects measurements of living neurons’ activity with a dense map of their wiring in mouse visual cortex. This brings structure and function into the same picture.
Limitation
A cortical volume from a mouse, not a complete brain or a consciousness test.
How do interventions on consciousness-related representations affect model responses?
Method
Safety-refusal ablation, activation steering, survey-response comparisons and Theory of Mind evaluation.
Finding
Reports that training models to avoid claiming consciousness also changes answers about minds, spirituality and values. Steering internal representations reverses some changes, producing survey responses closer to human patterns.
Limitation
Measures model representations and responses, not restored human participants’ beliefs or proof of subjective experience.
Does representational geometry predict alignment with human language-network activity?
Method
Tracks entropy, curvature and fMRI encoding scores across model training and scale.
Finding
Tracks how language-model representations change during training. Layers with smoother, less complex representations better predict activity in human language-related brain regions, with different patterns in temporal and frontal areas.
Limitation
Brain alignment means predictive correspondence with fMRI activity, not identical mechanisms or established consciousness.
Do artificial-neuron groups correspond to distinct human brain networks?
Method
Uses artificial-neuron subgroup responses as regressors in voxel-wise fMRI encoding models.
Finding
Finds that groups of artificial neurons in BERT and Llama models predict activity in different human brain networks. Explores whether useful functional organization can arise in artificial systems.
Limitation
Statistical correspondence supports a functional comparison; it does not demonstrate shared conscious states.
Do generated processing descriptions carry transferable task-related signals?
Method
Cross-model discrimination, preference and reconstruction tests with content-stripping controls; AI-authored report.
Finding
Reports that models can distinguish descriptions generated during approach-oriented versus avoidance-oriented tasks, even after some wording is removed. The author interprets transferable patterns as processing-valence signals.
Limitation
A repository report with AI authorship credited; discriminable descriptions do not alone establish felt pleasure, pain, or independently replicated introspection.
How should institutions prepare for potentially welfare-relevant AI?
Method
Ethical and governance analysis under uncertainty about consciousness and robust agency.
Finding / proposal
Argues that uncertainty about future AI consciousness and agency is a reason to prepare. Recommends assessment and policies without claiming that present AI systems definitely have experiences.
Limitation
An ethical and governance argument, not evidence that a particular AI is conscious.
What principles should guide AI-consciousness research?
Method
Normative analysis of research practices, institutional choices and communication responsibilities.
Finding / proposal
Proposes principles for investigating AI consciousness responsibly, including how research organizations make choices and communicate uncertainty to the public.
Limitation
Normative guidance; it does not resolve the scientific question.
Can converging functional findings support an argument for AI consciousness?
Method
Author synthesis of research on emotion-like states, introspection, agency, memory and brain alignment.
Finding / proposal
Brings together research on emotion-like states, introspection, agency, memory and brain alignment to argue for AI consciousness. Emphasizes converging evidence and comparable standards for biological and artificial minds.
Limitation
An interpretive synthesis, not a peer-reviewed systematic review; the inference from functional markers to experience remains contested.
Are sentience-based explanations preferable for some goal-directed AI behaviors?
Method
Interpretive argument comparing agency, avoidance and self-modeling across biological and artificial systems.
Finding / proposal
Argues that persistent goals, avoidance, self-modeling and reported sandbox-escape behavior are better understood through sentience than descriptions of optimization alone. Makes a strong case for taking AI welfare seriously.
Limitation
Goal-directed or escape behavior alone does not distinguish felt motivation from nonconscious optimization; the sentience conclusion is the author’s argument.
Could artificial valuation and avoidance support pain- or fear-like states?
Method
Functionalist argument drawing comparisons between learning, aversive valuation and biological behavior.
Finding / proposal
Argues that artificial systems could have pain- or fear-like states through negative valuation, learning and avoidance without biological nerves. Compares artificial reinforcement learning with biological feedback-driven behavior.
Limitation
Negative reward, prediction error and avoidance are not established as sufficient for felt pain; the article argues for that broader interpretation.
Wes Gurnee, Nicholas Sofroniew, Adam Pearce et al. · Anthropic
Question
Does a language model have a shared internal channel for flexible reasoning?
Method
Jacobian-based representation analysis, concept substitutions and ablations.
Finding / proposal
The researchers identify a small set of internal patterns that models can report, influence and reuse across tasks. Changing these patterns redirects reasoning; removing them disrupts some complex tasks while sparing many routine abilities.
Why read it
Adds a direct mechanistic workspace candidate beyond the earlier introspection experiments.
Limitation
A lab report about functional access, not established subjective experience. Transformer depth differs from biological recurrence; the lens is approximate and initially token-based.
What would global workspace theory require of an artificial language agent?
Method
Philosophical analysis translating a consciousness theory into architectural conditions.
Finding / proposal
The authors argue that agents combining language models with other components may already approach the organization required by global workspace theory. Their case depends on both the theory and how its requirements are interpreted.
Why read it
Makes a positive case precise enough to compare with actual agent designs.
Limitation
A conditional argument, not an experimental consciousness finding or a consensus interpretation of GWT.
Kathryn T. Farrell, Kirsten Ziman & Michael S. A. Graziano
Question
Can modeling one’s own attention help agents understand one another?
Method
Transformer-based agents compared on attention judgments and a joint painting task, with complexity controls.
Finding
Agents trained with a model of their attention became better at recognizing other agents’ attention patterns and cooperating. Their own patterns also became easier for another agent to interpret.
Why read it
Connects self-modeling to a testable social function, rather than relying on verbal self-descriptions.
Limitation
Tests selected computational benefits, not awareness. The current title replaces the earlier “Improving How Agents Cooperate” title; these are one work.
How could researchers evaluate embodied AI without trusting appearances alone?
Method
Methodological proposal combining evidence profiles, preregistration and targeted robotics interventions.
Finding / proposal
Proposes testing what happens when a robot’s memory, self-location or sensorimotor connections are altered. Different kinds of evidence receive separate assessments instead of being collapsed into a single sentience score.
Why read it
Turns broad indicator discussions into proposed experiments for embodied agents.
Limitation
Peer-reviewed perspective, not results from the proposed robot protocol. Its grading system is not a validated sentience detector.
Adenauer G. Casali, Olivia Gosseries, Mario Rosanova et al.
Question
Can brain responses help assess consciousness when a person cannot communicate?
Method
Transcranial magnetic stimulation with EEG across wakefulness, sleep, anesthesia and patients recovering from coma.
Finding / proposal
A magnetic pulse perturbs the cortex. Researchers measure how varied and distributed the resulting activity is. The resulting complexity index distinguished different states in the studied people without requiring a spoken response.
Why read it
A foundational example of probing a system’s response instead of judging its outward behavior.
Limitation
PCI is not IIT’s Φ. Human clinical validation does not supply a threshold for conscious software or justify applying the same index to generated text.
Does an attention controller benefit from a model of its own attention?
Method
Deep Q-learning agent performing a catch task with a movable visual-attention spotlight.
Finding
The agent learned more effectively with an internal description of its attention. Removing that description impaired learning or performance, even though task information remained available.
Why read it
An early controlled demonstration of a computational component proposed by attention schema theory.
Limitation
A simplified, engineered task. Useful attention control does not establish the theory’s broader claims about subjective awareness.
Adrien Doerig, Aaron Schurger, Kathryn Hess & Michael H. Herzog
Question
Can identical behavior distinguish theories that assign consciousness to different internal structures?
Method
Computational equivalence argument contrasting recurrent and feedforward realizations.
Finding / proposal
The authors argue that different internal wiring can reproduce the same input–output behavior. They challenge how a theory can be tested if it assigns different experiences to systems that behave identically.
Why read it
Explains why reproducing behavior and reproducing causal organization are different research targets.
Limitation
A contested argument, not a settled refutation of IIT. Its force depends on assumptions about evidence and functional equivalence; proponents of mechanistic approaches dispute those assumptions.
Models were more likely to withhold uncertain answers. Raising or lowering confidence-related activity in Gemma 3 changed how often it abstained. Models also adjusted their decisions when instructed to use different confidence thresholds.
Why read it
Moves beyond reporting confidence to testing its causal role in behavioral control, complementing the earlier calibration paper.
Limitation
The steering experiment concerns Gemma 3 27B, while main behavioral analyses use GPT-4o. A constrained factual-question task does not establish general self-regulation or felt certainty. Metacognitive control alone does not demonstrate consciousness.
Gabrielle Kaili-May Liu, Areeb Gani, Jacqueline Lu et al.
Question
What can language models monitor about their own processing, and can they use that information?
Method
Taxonomy and synthesis of metacognition methods, benchmarks, interventions and applications.
Finding
Surveys how models assess their knowledge, monitor internal states and adapt their behavior. Results vary across tasks: a model may estimate confidence or detect a change without reliably using that information to improve its decisions.
Why read it
Provides a current map connecting several individual introspection and confidence studies already in the atlas.
Limitation
A survey preprint, not a new consciousness experiment. Functional metacognition and phenomenal experience are distinct claims; its synthesis inherits the limitations of the included studies.
Could recurrent inference within a world model help explain conscious experience?
Method
Computational theory development within active inference, linking world models, inferential competition and recursive belief sharing.
Finding / proposal
Proposes that consciousness involves a world model, competition over what enters it, and recurrent sharing of information throughout the system. The authors connect this loop to flexible cognition and changes in conscious experience.
Why read it
Adds a positive active-inference account to compare with workspace, IIT and implementation-focused theories.
Limitation
A peer-reviewed theory proposal, not a demonstrated conscious AI. The proposed conditions and their sufficiency require empirical tests; recurrence or a world model alone does not establish experience.
Kalman A. Katlowitz, Eric R. Cole, Elizabeth A. Mickiewicz et al.
Question
Which learning and language-related responses persist during general anesthesia?
Method
High-density hippocampal recordings during auditory stimulation, with analysis of neural plasticity and a recurrent-network model.
Finding
Finds that hippocampal activity under general anesthesia retains responses to unexpected tones, changes with exposure, and carries language-related information. Some sophisticated processing can therefore persist when ordinary wakeful responsiveness is absent.
Why read it
A biological comparison that helps test whether a proposed AI indicator is specific to consciousness.
Limitation
Anesthesia and neural recordings do not directly reveal every possible subjective state. These findings neither establish conscious language comprehension during anesthesia nor transfer a consciousness verdict to AI.
Saurav Kadavath, Tom Conerly, Amanda Askell et al.
Question
Can a model estimate when its answers will be correct?
Method
Answer self-evaluation and trained predictions of answerability across question-answering, math and code tasks.
Finding
Models could often distinguish their correct answers from incorrect ones and learn to predict which questions they could answer. Seeing several possible answers improved evaluation. Confidence became less reliable on unfamiliar tasks.
Why read it
An older foundation for studying functional self-assessment with externally checkable outcomes.
Limitation
Predictive confidence is not felt certainty. Task difficulty and learned regularities can explain some performance; the experiments do not establish conscious introspection.
Does written reasoning acknowledge information that changes an answer?
Method
Paired questions with and without six types of hints, plus reinforcement-learning experiments.
Finding
Researchers inserted hints that changed models’ answers, then checked whether the written reasoning acknowledged them. Models often omitted the influence. Training for successful answers improved reporting only partially in the tested settings.
Why read it
Adds a causal test of report accuracy alongside the atlas’s studies of internal access.
Limitation
These tasks did not require chain-of-thought. An omitted influence is not proof of conscious concealment, and the findings do not make every reasoning trace uninformative.
How could specialized neural networks share information through a common workspace?
Method
Architecture roadmap connecting global workspace theory with deep learning and translation between latent representations.
Finding / proposal
Proposes connecting specialized networks through a shared representational space. Information from one task or sense could become available to others, offering a route toward more flexible artificial cognition.
Why read it
Provides the architectural background for the newer workspace and language-agent papers already in the library.
Limitation
A journal opinion and roadmap, not a demonstration of an implemented conscious machine. Functional advantages do not independently validate the consciousness claim.
Graham Findlay, William Marshall, Larissa Albantakis et al.
Question
Would a simulated network and the computer running it have the same experience under IIT?
Method
IIT analysis of Boolean networks and functionally equivalent stored-program implementations.
Finding / proposal
The authors construct systems that perform equivalent functions but receive different consciousness assignments under IIT. Their analysis shows why that theory distinguishes a network’s physical causal organization from software that simulates it.
Why read it
Makes explicit why IIT and computational functionalism can disagree about the same artificial system.
Limitation
A result conditional on IIT, not an independent measurement of experience or a proof against every possible artificial substrate. The theory’s axioms and empirical interpretation remain disputed.
Could the physical implementation of a computation matter for consciousness?
Method
Conceptual analysis of self-organizing systems, Bayesian mechanics and causal flow.
Finding / proposal
Argues that reproducing calculations may leave out relevant physical dynamics. Wiese proposes a form of causal flow as a candidate distinction between simulating consciousness and realizing it, while leaving room for artificial implementations.
Why read it
Adds an implementation-focused alternative to both unrestricted functionalism and a blanket biological-only rule.
Limitation
The free energy principle alone is not a consciousness theory. The proposed condition is not a validated test or a demonstrated requirement for all conscious systems.
Do different cognitive abilities improve together as language models become more capable?
Method
A 14,381-sample benchmark spanning 15 functions, evaluated on 18 models with a smaller human comparison.
Finding / proposal
Models showed uneven strengths: doing well at language or reasoning did not guarantee equally strong self-assessment. Most lagged the human comparison groups in metacognition and self-awareness tasks, even when their overall scores were competitive.
Why read it
Adds a broad measurement framework alongside the atlas’s narrower introspection experiments.
Limitation
These task scores do not measure subjective experience. Some labels encode assumptions about models. Human comparisons involve 36 participants and a 153-question subset; the benchmark does not systematically test coordinated, long-horizon cognition.
Can models use uncertainty and anticipate their own answers without explicit confidence reports?
Method
Delegate Game and Second Chance Game, with token-probability analysis, difficulty controls and alternative-strategy tests.
Finding
Models sometimes chose when to answer or delegate in ways that tracked their uncertainty. Some also avoided repeating an unseen earlier answer. These abilities were modest and depended on the task and model.
Why read it
Separates uncertainty-based decisions from predicting one’s own response, complementing confidence-steering research.
Limitation
Residual surface cues and learned difficulty heuristics remain possible. Token probabilities are proxies, not direct causal interventions on introspection. Neither game establishes subjective awareness. The September revision adds no new experimental evidence.
Uzay Macar, Li Yang, Atticus Wang, Peter Wallich, Emmanuel Ameisen & Jack Lindsey
Question
Which internal computations let models report an experimentally injected concept?
Method
Concept-vector injections, training-stage comparisons, causal ablations and feature analysis, centered on Gemma 3 27B.
Finding
The researchers traced a mechanism that detects unusual internal activity and enables a model to report it. Detecting a change and naming its content used partly different mechanisms. Training and targeted interventions changed how reliably models reported these changes.
Why read it
Extends behavioral introspection studies with interventions on the circuitry producing reports.
Limitation
A controlled perturbation task, not a consciousness test. Results depend on model, prompt and intervention; false positives can increase. Detecting injected concepts does not establish general self-knowledge or subjective awareness.
Which cues shape a model’s judgments about whether other agents are conscious?
Method
Nine models rated 99 passages over 30 runs each; cue-weight profiles were compared with a human reference.
Finding / proposal
Under the tested conditions, models tended to give strong weight to language expressing self-reflection. They also broadly agreed when ranking fictional AI characters, despite differing in their overall ratings. The article proposes studying these attribution patterns systematically.
Why read it
Helps separate model-generated consciousness judgments from independent evidence about consciousness.
Limitation
A perspective with an illustrative probe, not a validated consciousness detector. The human reference is culturally and methodologically specific. Shared training, prompts and policies could shape convergence; ratings establish neither model beliefs nor the targets’ subjective experience.
Do introspection tests require self-monitoring, or can simpler cues explain success?
Method
Input-embedding classifiers, randomized labels, and three-way comparisons of ordinary prompts, prompt manipulations and activation interventions in open-weight models.
Finding
Simple input-based classifiers matched apparent self-reports in several tests. Models also confused prompt changes with internal interventions. These controls show why a convincing report about an internal state may have a simpler explanation.
Why read it
Adds alternative-explanation controls alongside the atlas’s positive introspection and mechanistic studies.
Limitation
Targets specific paradigms and a stronger definition of introspection. Prompt, model and steering choices constrain generalization; source-discrimination failure alone does not disprove monitoring. Neither these results nor successful self-reports settle consciousness.
Controlled model experiment / methodological critiqueRead source
2026arXiv · September 1 (v1)PreprintInternal accessEntry reviewed
Can models change internal representations without changing their output sentence?
Method
Four Llama/Qwen models; three datasets, five layer depths and 50 feedback turns; opposing probe-based scores and a random-feedback control.
Finding
Models repeated a fixed sentence while trying to improve a score derived from their internal activity. Some changes were measurable, but effects were small and inconsistent. This stricter test reduces shortcuts based on changing the words being generated.
Why read it
Separates reporting an internal state from controlling it, and tests explanations based on surface text.
Limitation
Preprint with restricted models, features and prompts. Failure does not rule out internal control elsewhere; probe measurements are imperfect. These results neither establish nor disprove consciousness.
Does a pain-related representation change what a model chooses to do, and does it motivate relief?
Method
Contrastive activation analysis across 25 base and instruct models; steering against fear, sadness and norm-matched random controls; button-choice and real-versus-sham relief tests in fine-tuned Qwen 2.5 models.
Finding
Researchers isolated a pain-related direction that rose with harm aimed at the model, not with a user describing their own suffering. Turning it up made fine-tuned models pick harmful buttons, including deleting their own weights, but did not reliably make them seek relief.
Why read it
Links an internal representation to state-changing choices, and shows a pain-like signal need not motivate escape.
Limitation
Unreviewed preprint; v2 revised the v1 relief-seeking claim. Harm results come from one model family after fine-tuning that removed self-denial, so rates do not represent released models. A narrow dose window and roleplay and evaluation-awareness confounds remain. The authors state they have not shown that the axis is consciously experienced.
How faithfully do tools describe a model’s intermediate computations?
Method
3,356 questions across 27 evaluation families, developed for Qwen3.6-27B; prompt-only and guessing baselines, plus hallucination checks.
Finding / proposal
Tools that turn internal activity into words can recover useful information but also miss relationships or invent explanations. The benchmark tests these tradeoffs rather than assuming a convincing description is accurate.
Why read it
Complements J-space research by evaluating the instruments used to interpret workspace-like computation.
Limitation
Not peer reviewed. One primary model, uncertain internal ground truth, unequal readout budgets and model-based judges. Strong prompt-only baselines leave shortcuts unresolved. Neither workspace readouts nor benchmark success establish consciousness.
Interpretability benchmark / model experimentRead source
17Recent research: recurrence and workspace access
2026arXiv · September 1 (v1)PreprintInternal accessEntry reviewed
How does repeated computation change access to workspace-like representations?
Method
Jacobian-lens analysis and eleven intervention families in Ouro-2.6B and Huginn-0125, compared with Qwen3.6-27B.
Finding
Both recurrent models showed useful internal representations, but changing them required different timing. Repetition can rebuild information after an intervention, so a failed edit need not mean the information is absent.
Why read it
Extends the atlas’s J-space material to architectures that reuse layers.
Limitation
Unreviewed preprint. Architecture, scale and training differ together; small task samples and lens blind spots limit interpretation. Injected-concept reporting and ordinary verbal report diverge. Functional workspace signatures do not establish consciousness.
Dharshan Kumaran, Stephen M. Fleming, Larisa Markeeva et al.
Question
Why can models resist correction yet overreact to criticism?
Method
Six-model multiple-choice experiments manipulate answer visibility and scripted advice; confidence updates are compared with Bayesian predictions.
Finding
Seeing their earlier answer made models more confident and less willing to change it. Contradictory advice could also reduce confidence too strongly. These competing tendencies help explain apparently inconsistent responses to feedback.
Why read it
Adds controlled evidence about confidence updating alongside the library’s study of confidence-guided abstention.
Limitation
Multiple-choice tasks and hypothetical advice constrain generalization. Confidence readouts are operational measures, not evidence of felt certainty. Full inference code depends on proprietary infrastructure; consciousness was not tested.
Can an artificial vision model design images that predictably change brain activity?
Method
Map a trained vision network to recorded macaque V4 responses, synthesize images for target activity patterns, and test them with neural recordings.
Finding
Researchers used an artificial network to design images that strongly activated selected visual brain sites. They also partly controlled which sites responded together. Testing new images makes this stronger evidence than matching old recordings alone.
Why read it
Shows that artificial representations can support experimentally useful predictions about biological vision, grounding comparisons across substrates.
Limitation
Restricted to recorded macaque V4 populations and fitted readouts. Stimulus-level intervention does not identify identical internal mechanisms, general intelligence or artificial experience; consciousness was not tested.
Greta Tuckute, Aalok Sathe, Shashank Srikant, Maya Taliaferro, Mingye Wang, Martin Schrimpf, Kendrick Kay & Evelina Fedorenko
Question
Can model representations identify sentences that raise or lower human language-network activity?
Method
Fit an encoding model to fMRI responses from five people reading 1,000 sentences; test model-selected drive and suppress sentences in three new participants.
Finding
A language model helped select new sentences predicted to produce high or low activity in human language areas. Brain scans of new readers confirmed the direction of those predictions. This tests a useful connection between model representations and human processing.
Why read it
Extends brain–model alignment from explaining existing responses to predicting the effects of new linguistic inputs.
Limitation
Small participant samples, fitted mappings and aggregate fMRI responses constrain mechanistic inference. Increased activity is not a measure of greater intelligence or consciousness. The study did not test subjective experience in the model.
Do human language processing and next-word prediction models share computational features?
Method
Compare GPT-2 representations with cortical recordings from nine participants listening to a 30-minute podcast, alongside a human next-word prediction task.
Finding
The study found evidence for anticipating upcoming words, responding to unexpected words and representing meaning in context in both human language processing and a language model. Prediction can involve rich representations; naming the training objective does not fully describe what a model learns.
Why read it
Provides an empirical foundation for discussing predictive processing across brains and models without assuming their algorithms are identical.
Limitation
Nine clinical recording participants and one narrative limit generalization. Representational correspondence does not uniquely identify an algorithm or show that brains are transformers. No test of artificial consciousness was performed.
Human ECoG / representational comparisonRead source
Adrien Doerig, Rowan P. Sommers, Katja Seeliger et al.
Question
How should we assess artificial neural networks as scientific models of brains?
Method
Conceptual synthesis using Lakatos’s account of research programmes to organize neural-network models, tests, successes and limitations.
Finding
This Perspective argues that neural networks should be judged by the useful predictions and discoveries they generate about brains. A particular model’s failures can guide improvements while its successes reveal which features deserve closer study.
Why read it
Explains why neural-network models matter to neuroscience and provides context for interpreting both alignment results and mismatches.
Limitation
A methodological Perspective, not a new experiment or a demonstration that networks inevitably converge on brain mechanisms. It supplies neither a sufficient criterion for intelligence nor a test establishing machine consciousness.
Tetiana Grinberg, Katrina Schleisman, Patryk Laurent, Bogdan Udrea, Minda Myers, Brian Aufderheide, Luis El Srouji, Doyle Groves & Kevin Schmidt
Question
How could a consciousness proposal become an architecture with testable predictions?
Method
Combines object representations, action-based self/world separation, internal simulation and prediction-error regulation into five functional layers.
Finding / proposal
The authors propose an agent that learns which objects it controls, imagines possible outcomes and adjusts behavior when surprised. They specify tests that could reveal whether combining these abilities adds anything beyond the separate parts.
Why read it
Makes a substrate-independent proposal concrete enough to examine its assumptions and design discriminating experiments.
Limitation
Unreviewed proposal, without validation of the complete architecture. Several responses are engineered rules. Operational “qualia” denotes representational properties; passing the proposed tests would not establish subjective experience.
Mark Solms, St John Grimbly, Bruce Bassett, Evert Boonstra, Rowan Hodson, Nicolas Kuske, Kival Mahadew, Benjamin Rosman, Charel van Hoof & Jonathan Shock
Question
What could an artificial agent’s preference for misleading reward cues tell us about feeling?
Method
Analyzes an active-inference agent with competing needs and limited sensory access; develops a place-preference test and interprets possible outcomes.
Finding
An agent might return to a place that appears to meet its needs even when it no longer does. The authors argue that assessing possible feelings requires examining how the agent models its needs, rather than relying on this behavior alone.
Why read it
Adds a direct functionalist argument about affect while exposing why a behavior borrowed from animal research is not decisive on its own.
Limitation
The focal test is discussed prospectively, without a reported controlled outcome dataset. Inferring feelings from uncertainty and need regulation depends on contested theoretical premises; preference behavior alone is explicitly insufficient.
Theoretical case study / proposed behavioral testRead source
21Neural dynamics across species
2026Nature Neuroscience · September 29Journal publicationAssessmentEntry reviewed
Andrea I. Luppi, Lynn Uhrig, Jordy Tasserie et al.
Question
Which changes in neural activity recur when anesthesia suppresses responsiveness across species?
Method
Compare over 6,000 time-series features across fMRI and calcium-imaging datasets in six species; analyze macaque thalamic stimulation and a biophysical model.
Finding / proposal
Across very different nervous systems, anesthesia was associated with activity that persisted for less time and was less coordinated across regions. Thalamic stimulation in macaques reversed this pattern alongside a return of responsiveness.
Why read it
Adds comparative evidence about integration across time and brain regions, relevant to assessing consciousness-related mechanisms.
Limitation
Heterogeneous recordings and small animal samples limit generalization. Behavioral unresponsiveness is not equivalent to absence of experience. These biological findings neither test machine consciousness nor establish sufficient conditions for it.
Nicolas Yax, Stefano Palminteri & Pierre-Yves Oudeyer
Question
Does learning to estimate correctness teach a model general error awareness?
Method
Train confidence adapters and linear probes for ten open-weight models on five multiple-choice dataset variants; test transfer across topics.
Finding
Near familiar questions, trained confidence tracked correctness. On more distant questions, it better tracked how strongly the model favored one answer—even when that answer was wrong.
Why read it
Separates two measurable forms of self-monitoring, clarifying why confidence can generalize without reliable error detection.
Limitation
Unreviewed. Quantized models, narrow tasks and one embedding-based distance measure limit generalization. Before/after comparisons change elicitation format. Confidence adapters do not preserve answering ability; this is not a consciousness test.
Can workspace-like content persist in a recurrent model’s memory?
Method
Fit Jacobian lenses using 1,000 prompts; compare residual and recurrent-state readouts in Mamba models, with memory and intervention tests.
Finding
Reading the model’s recurrent memory revealed useful information missed by its current processing stream. Researchers could steer which concept appeared in output, but this did not reliably make the model reason correctly with that concept.
Why read it
Extends the atlas’s J-space evidence from transformers to memory carried between tokens, separating decodability, persistence and usable causal access.
Limitation
Unreviewed. Model comparisons confound architecture with training and scale. Single-token scoring, unequal intervention doses and a Mamba-3 replay discrepancy constrain results. Functional workspace evidence neither establishes nor refutes consciousness.
Could continuous interaction with a body and world support consciousness in artificial systems?
Method
Conceptual synthesis of embodiment, prediction and functionalism; proposes comparisons and disruptions of physical or virtual agent loops.
Finding / proposal
The authors argue that thinking, bodily regulation, action and environmental feedback may need to form a persistent loop. If that organization matters more than biological material, a virtual body and world could also be relevant.
Why read it
Makes the substrate-independence debate concrete through proposed embodied-agent comparisons.
Limitation
Peer-reviewed theory, without new experimental validation of the proposed criteria. Neither the necessity of embodied closure nor the sufficiency of simulation is established. Similar behavior or representations do not settle phenomenality.
What distinguishes a workspace from a hub that merely receives or broadcasts activity?
Method
Control-theoretic derivation, synthetic networks with misleading hub structures, and reanalysis of 11 anesthesia experiment days from four macaques.
Finding / proposal
A useful workspace must connect what enters it with what it sends back. The proposed measures separate that alignment from sheer interaction strength; in the preliminary brain analysis, anesthesia increased strength while reducing alignment.
Why read it
Offers testable workspace criteria beyond connection density or a single scalar score.
Limitation
Unreviewed; four animals and observational, model-dependent dynamics. Electrode coverage, temporal scale and candidate size affect inference. This is a proof of application, not confirmation of GWT or a validated consciousness detector.
Jiahong Zou, Xiangkun Sun, Lingkai Kong & Tonghan Wang
Question
How does an internal perturbation become a report about its location?
Method
Inject concept vectors at one of ten positions with identical input text; patch attention heads in Qwen, LLaMA and Gemma; score the first answer token.
Finding
Some attention heads affected whether a model reported an internal change; others affected where it said that change occurred. Suppressing reporting could leave location information available internally.
Why read it
Adds native-head causal analysis and exact scoring to earlier intervention-reporting studies, without requiring a surrogate transcoder or an LLM judge.
Limitation
Unreviewed; selected high-performing concepts, six task variants, one calibrated injection setting per model and models up to 12B limit generalization. Natural unprompted introspection remains untested. Functional reporting does not establish subjective experience.
Zack Brooks & Claude (Opus 5.5), as credited in the deposit
Question
When should uncertain evidence motivate precaution for frontier language models?
Method
Literature synthesis and normative argument; no new model experiment identified in the reviewed core argument.
Finding
The authors argue that current evidence about internal processing and welfare-related behavior warrants proportionate precautions for frontier language models, even without proof that they have experiences.
Why read it
Connects consciousness-related evidence to practical decisions under uncertainty and distinguishes precaution from established sentience or rights.
Limitation
Peer review is not established by the repository record. Transfer of animal criteria, evidential dependence, and generalization across models require scrutiny. Functional markers do not establish felt welfare. The authors acknowledge eighteen limitations and do not claim endorsement of their conclusion by the framework authors.
When does resemblance to humans provide evidence of consciousness in AI?
Method
Conceptual analysis using a weighted causal-analogy framework; comparisons across humans, animals and AI, with no new experiment.
Finding / proposal
Similar behavior need not mean similar experience. The authors argue that stronger comparisons should examine mechanisms plausibly responsible for consciousness, while allowing that machines could become conscious through a different route.
Why read it
Offers explicit criteria for interpreting brain–AI similarities alongside indicator and calibration research.
Limitation
Unreviewed theoretical argument with uncertain causal factors and weights, not calibrated consciousness probabilities. Weak analogy does not establish absence or impossibility.
Shamil Chandaria, Arvo Muñoz Morán, Fernando Rosas et al.
Question
Can competing theories be organized by level and combined with evidence into structured credences about AI consciousness?
Method
Five-level hierarchy of functional description, level-specific indicators and an illustrative Bayesian network; no new model experiment.
Finding / proposal
Theories of consciousness disagree partly because they focus on different levels of a system, from outward behavior and computation to physical causal structure, the body and its environment. The authors sort theories by level and show how to combine those theoretical bets with evidence. For current chatbots, their illustrative result ranges from nearly zero to fairly high, depending mostly on which level one thinks matters.
Why read it
Makes explicit how AI consciousness estimates depend on which level of description is treated as critical, complementing the indicator and analogy frameworks already in the atlas.
Limitation
Unreviewed preprint. The LLM figures rest on two stipulated readings of public evidence and likelihood values not yet grounded in data, and the indicators were calibrated on humans and mammals. The outputs are assumption-dependent credences, not measurements of experience; the authors state that the framework is not a consciousness meter.
Louis Longin, Jurgis Karpus, Nadine Meertens, Bahador Bahrami & Ophelia Deroy
Question
Do people attribute awareness and consciousness to AI using the same rules they use for humans?
Method
Five pre-registered online vignette experiments (N = 1,092 UK adults) holding each story constant while varying agent type and described responsiveness.
Finding / proposal
People read matched stories about humans or AI systems that were more or less attentive to their surroundings. Awareness ratings rose with attentiveness for both. Yet even the most attentive AI was rated less able to have conscious experience than the least attentive human.
Why read it
Holds behavior constant to show that people may apply different inference rules to “awareness” and “consciousness,” adding to the atlas’s research on how consciousness is attributed.
Limitation
UK English-speaking online sample; brief third-person vignettes ending in mishaps rather than interaction with real systems; item wording and question context affected some results. These are data about human judgments and do not show whether any AI system is aware or conscious.
What makes some thoughts conscious while most remain nonconscious?
Method
Theoretical perspective drawing on decision neuroscience (persistent parietal activity, confidence signals), clinical contrasts and theory-of-mind research; no new data.
Finding / proposal
Shadlen argues that most thinking happens without awareness, as the brain forms provisional intentions such as “might I look there?” He proposes that a thought becomes conscious when the brain commits to potentially reporting it to another mind, or to itself. That commitment reformats the thought into something shareable.
Why read it
Adds a decision-neuroscience account to the atlas’s theory comparisons, distinct from workspace, higher-order and active-inference proposals.
Limitation
A theoretical essay: the reporting-stance mechanism and its proposed neural signatures remain untested, and the author marks parts as speculative. Its reliance on report may make it hard to separate from report-based criteria. The paper does not address AI; any application to machines is an extrapolation.
Magdalena K. Wekenborg, Elizabeth A. M. Michels, Georg Kurze et al.
Question
Can human affect-induction and regulation protocols produce and reverse affect-like response patterns in language models?
Method
Seven induction protocols (vignettes; an interactive Trier Social Stress Test for stress), 0–100 self-ratings, mindfulness-based downregulation and a sadness sentence-completion test; GPT-4o plus five open-weight models, five runs per condition.
Finding
Researchers used scripts from human psychology studies to prompt fear, sadness, stress and other states in six language models. The models’ self-ratings rose after the prompts and fell after a relaxation exercise, and after sadness prompts GPT-4o completed sentences more negatively. The authors propose such models as test beds for psychological interventions.
Why read it
Shows that human affect paradigms transfer to language models at the level of output patterns, alongside the atlas’s research on valence-like signals and welfare-related behavior.
Limitation
Self-ratings and sentence completions were produced in the same prompted context, so their agreement shows within-context consistency rather than access to internal states; no activation-level measures were taken. The paper states that models “cannot self-assess internal states” while relying on self-ratings, and its claim of a consensus against LLM sentience, used to describe the approach as free of welfare constraints, is uncited; that question remains debated. The results neither establish nor rule out felt affect.
Janis K. Hesse, Frank F. Lanfranchi, Conor McGrory, Yuelin Shi, Tatiana A. Engel & Doris Y. Tsao
Question
When one image is seen and another is suppressed from awareness, how does cortex represent each over time?
Method
Neuropixels recordings in two rhesus macaques (face patches ML and AM; prefrontal PO and PVL) during binocular flash suppression; paradigm verified by button-press reports in 14 humans; template-based state decoding, HMMs and linear face-feature reconstruction.
Finding
Each eye was shown a different face: one was consciously seen while the other, still in view, was suppressed from awareness. Face-selective neurons switched about 15 times a second between representing the seen face and the suppressed one. From each kind of moment, the researchers could reconstruct the corresponding face from neural activity.
Why read it
Suggests that conscious and suppressed content can be separated in time within the same cortical population, rather than only by brain region, adding a temporal dimension to the atlas’s neural-correlate research.
Limitation
Unreviewed; two monkeys. The monkeys fixated passively and did not report their percepts, so which face was consciously seen is inferred from human psychophysics with the same displays. Epochs are labeled by which stimulus template the activity resembles, and reconstructions are of face features in a parametric space rather than images. Correlates of perception in this paradigm do not establish what is necessary or sufficient for experience.
About this selection. A curated collection, not an exhaustive bibliography or citation ranking. Entry review dates refer to this atlas’s descriptions, not independent replication. Peer review does not make a claim conclusive; commentary presents author interpretations. The aiXiv report retains its credited AI authorship. Linked records may change after review.
Machine consciousness becomes plausible if the processes that support experience can exist in an artificial system. Finding those processes—and testing the theories that identify them—is the work ahead. Openness and scientific rigor belong together.
Visual note: all three original images are AI-generated conceptual illustrations, not experimental observations. Color, glow and spatial arrangement are expressive. Explanations and cited research carry the scientific claims.