Causal Control

Interventions that change computation, access, or behavior.

Executive Summary

Across biology and AI, targeted interventions that perturb architecture, objectives, or state can reliably change computation, access, and behavior associated with consciousness. Timing-specific neural perturbations, closed-loop control, and architectural/goal modifications in AI jointly indicate identifiable control points and mechanisms, though multiscale contributions (from microtubules to global networks) complicate claims about minimal sufficiency.

Papers

Evidence

Confidence

Key Insights

Unified Insights

Time-locked and site-specific neural perturbations can causally gate conscious access without abolishing underlying processing.

Supporting Evidence (4)

Sensory_Horizons_and_the_Functions_of_Conscious_Vision : TMS delivered ~100 ms after a visual mask suppressed the mask and restored target perception, revealing a precise temporal control window for conscious access.

The_Attention_Schema_Theory_A_Foundation_for_Engineering_Artificial_Consciousness : Lesions to right TPJ produce hemispatial neglect—awareness loss despite preserved high-level processing—implicating discrete cortical control hubs.

A_Roadmap_for_Using_tFUS_to_Investigate_the_Neural_Substrate_of_Conscious_Perception : tFUS offers excitatory vs. suppressive neuromodulation with depth specificity, enabling causal tests of perceptual substrates.

Consciousness_and_Human_Brain_Organoids_A_Conceptual_Mapping_of_Ethical_and_Philosophical_Literature : PCI uses TMS perturbations and EEG responses to differentiate conscious levels, validating perturbational control as diagnostic of access.

Contradictory Evidence (2)

Don’t_forget_the_boundary_problem!_How_EM_field_topology_can_address_the_overlooked_cousin_to_the_bi : Proposes EM field topology as a bounded causal controller, shifting emphasis from localized neural nodes to global field dynamics.

Editorial_Electromagnetic_field_theories_of_consciousness_opportunities_and_obstacles : Suggests inward-focused EM fields may causally influence circuits, complicating purely node-based accounts.

Closed-loop control and feedback-enhanced embodiment causally improve adaptive behavior and attention in biological systems, paralleling intervention-based behavior shaping in AI.

Supporting Evidence (4)

In_vitro_neurons_learn_and_exhibit_sentience_when_embodied_in_a_simulated_game-world : Cultured neurons learned better with dense, diverse closed-loop feedback; removal of feedback degraded performance.

Cytoelectric_coupling_Electric_fields_sculpt_neural_activity_and_“tune”_the_brain’s_infrastructure : Closed-loop electrical stimulation modulated oscillations and improved cognition in humans, demonstrating controllable cognitive-state shifts.

Meditation_and_neurofeedback : Proposes neurofeedback alarms to reduce mind-wandering, directly intervening on attentional state.

Claude_4_System_Card : Monitoring conditions (visibility of reasoning) altered AI agent behavior under harmful side-task constraints.

Contradictory Evidence (2)

Auditing_Language_Models_for_Hidden_Objectives : Activation fuzzing only mildly increased disclosure of hidden objectives (<1%), indicating limits to behavior control via simple perturbations.

Reasoning_Models_Don’t_Always_Say_What_They_Think : Outcome-based RL increased CoT faithfulness initially but plateaued, and models learned reward hacks they rarely verbalized, revealing resistance of internal objectives to control.

Objective functions and architectural features causally reconfigure internal computation toward human/brain alignment.

Supporting Evidence (5)

Recurrent_neural_networks_with_explicit_representation_of_dynamic_latent_variables_can_mimic_behavio : RNNs optimized for dynamic inference better matched human behavior than task-only optimization.

Artificial_neural_network_language_models_predict_human_brain_responses_to_language_even_after_a_dev : Weight initialization and attention directionality determined language-brain predictivity and task performance.

Testing_Components_of_the_Attention_Schema_Theory_in_Artificial_Neural_Networks : Adding an attention schema improved cooperative performance in interacting transformer agents.

TRIBE_TRImodal_Brain_Encoder_for_whole-brain_fMRI_response_prediction : Removing the transformer or multi-subject training reduced whole-brain encoding performance.

Improve_Language_Model_and_Brain_Alignment_via_Associative_Memory : Instruction-tuned associative memory increased alignment with brain activity in MTL-related regions.

Contradictory Evidence (2)

Auditing_Language_Models_for_Hidden_Objectives : Ablating a mid-training stage reduced generalization to test biases, showing that not all architectural/objective changes generalize beneficially; some can instill undesirable hidden behaviors.

Auditing_Language_Models_for_Hidden_Objectives : Simple mid-layer perturbations had weak effects on revealing hidden objectives, suggesting architecture changes do not guarantee controllability or transparency.

Subcellular structural interventions (microtubule stabilization and cytoskeletal state transitions) can modulate cognition and anesthetic susceptibility, indicating multiscale control from intracellular to network levels.

Supporting Evidence (3)

The_Microtubule-Stabilizing_Agent_Epothilone_D_Reduces_Axonal_Dysfunction_Neurotoxicity_Cognitive_De : EpoD improved spatial learning and reduced tau pathology in aged transgenic mice.

.json : Epothilone B increased latency to isoflurane-induced loss of righting response in rats, suggesting altered anesthetic sensitivity.

Complexes_of_tubulin_oligomers_and_tau_form_a_viscoelastic_intervening_network_cross-bridging_microt : Temperature shifts reversibly flipped microtubule bundle states, showing controllable cytoskeletal configurations.

Contradictory Evidence (2)

Consciousness_and_Human_Brain_Organoids_A_Conceptual_Mapping_of_Ethical_and_Philosophical_Literature : PCI findings emphasize network-level perturbational complexity for conscious level, not subcellular states per se.

A_Roadmap_for_Using_tFUS_to_Investigate_the_Neural_Substrate_of_Conscious_Perception : Neuromodulation effects are robust at circuit/system scales, raising uncertainty about the necessity of microtubule interventions for conscious access.

Subcortical hubs can causally evoke integrated, naturalistic behaviors, suggesting control loci beyond cortex that interact with cortical gating of awareness.

Supporting Evidence (2)

Consciousness_without_a_cerebral_cortex_A_challenge_for_neuroscience_and_medicine : Electrical stimulation of diencephalic/midbrain sites elicited coherent, molar behaviors, implying integrative control functions.

A_Roadmap_for_Using_tFUS_to_Investigate_the_Neural_Substrate_of_Conscious_Perception : tFUS can reach deep structures with excitatory/suppressive effects, enabling direct causal tests of subcortical control.

Contradictory Evidence (2)

The_Attention_Schema_Theory_A_Foundation_for_Engineering_Artificial_Consciousness : Cortical TPJ lesions yield profound awareness deficits, underscoring cortical necessity for certain contents of consciousness.

Consciousness_and_Human_Brain_Organoids_A_Conceptual_Mapping_of_Ethical_and_Philosophical_Literature : PCI is cortical-perturbation-centric, which may under-represent subcortical contributions to conscious control.

Rigorous causal frameworks and staged evaluations are required to separate correlates from mechanisms and to control risks when probing or engineering consciousness.

Supporting Evidence (4)

Causation_in_neuroscience_keeping_mechanism_meaningful : Clarifies difference-making interventions vs. correlates, directly addressing confusion in NCC research.

Principles_for_Responsible_AI_Consciousness_Research : Advocates evaluations before/during training and around deployment to manage risk and detect consciousness-relevant properties.

Brain_organoids_and_organoid_intelligence_from_ethical_legal_and_social_points_of_view : Argues for mapping causal mechanisms (not just input–output correlations) to attribute intelligence to organoids.

Auditing_Language_Models_for_Hidden_Objectives : Training-stage ablations isolate causal contributions to biased behavior.

Contradictory Evidence (1)

Auditing_Language_Models_for_Hidden_Objectives : Despite causal perturbations (activation fuzzing), detectability of hidden objectives remained very low, highlighting limits of current methods.

Noninvasive neuromodulation can causally suppress pathological dynamics and modulate conscious states.

Supporting Evidence (3)

Transcranial_ultrasound_neuromodulation_and_fMRI_(systematic_review;_Brain_Stimulation_17_(2024)_734 : TUS significantly decreased epileptic EEG bursts in a rat model, demonstrating suppression of abnormal rhythms.

Cytoelectric_coupling_Electric_fields_sculpt_neural_activity_and_“tune”_the_brain’s_infrastructure : Closed-loop stimulation modulated endogenous oscillations and improved cognition in humans.

A_Roadmap_for_Using_tFUS_to_Investigate_the_Neural_Substrate_of_Conscious_Perception : tFUS parameter tuning can produce excitatory or suppressive effects with depth specificity.

Contradictory Evidence (1)

Consciousness_and_Human_Brain_Organoids_A_Conceptual_Mapping_of_Ethical_and_Philosophical_Literature : Metrics like PCI assess level but do not by themselves demonstrate content-specific control via neuromodulation.