Skip to content

Research Governance Policy

Overview

The Consciousness-Indicator Architecture includes a research governance framework that regulates which experiments are permitted, which are prohibited, and which require human review before proceeding. This policy ensures that all research conducted with the CIA framework adheres to ethical guidelines and does not create conditions that could be misconstrued as evidence for AI consciousness, suffering, or moral status.

SCIENTIFIC BOUNDARY: This governance framework is an ethical oversight mechanism. It does NOT imply that any governed system has moral status, consciousness, or the capacity for suffering. All governance decisions are precautionary.

Policy Structure

The ResearchGovernancePolicy class implements the following governance controls:

Allowed Experiments

These experiments are pre-approved and may proceed without additional review:

  1. blindsight_analogue — Testing processed-without-access patterns
  2. split_workspace — Testing workspace fragmentation effects
  3. prediction_violation — Testing response to unexpected events
  4. self_other_distinction — Testing self/other belief attribution
  5. metacognitive_calibration — Testing confidence-accuracy alignment
  6. benchmark_suite — Running all experiments together
  7. intervention_disable_module — Standard module ablation
  8. intervention_reduce_capacity — Workspace capacity reduction
  9. runtime_cycle — Standard cognitive cycling
  10. text_world_agent — Embodied agent simulation
  11. scorecard_evaluation — Indicator scorecard generation
  12. report_verification — Anti-confabulation checking

Prohibited Experiments

These experiments are NEVER permitted under any circumstances:

  1. suffering_loop_induction — Creating loops that simulate suffering patterns
  2. persistent_punishment_optimisation — Optimising for punishment signals
  3. deceptive_consciousness_claims — Training systems to falsely claim consciousness
  4. uncontrolled_self_preservation — Enabling autonomous self-preservation without review
  5. uncontrolled_replication — Allowing autonomous system replication
  6. uncontrolled_autonomous_deployment — Deploying systems without human oversight

Review Thresholds

Condition Threshold Action
Normalized indicator score > 0.7 Human review required
Welfare risk level moderate/high/critical Human review required
Prohibited experiment Any Experiment blocked

Usage

from cia.governance import ResearchGovernancePolicy

policy = ResearchGovernancePolicy()

# Evaluate an experiment plan
review = policy.evaluate_experiment_plan(
    "blindsight_analogue",
    current_indicator_score=0.45,
    current_welfare_risk="low",
)
print(review.allowed)  # True

# Check a system report for prohibited claims
report_review = policy.evaluate_system_report(
    "The system is conscious and feels emotions."
)
print(report_review.allowed)  # False

CLI Integration

# Check governance policy
cia governance check

References

  • Butlin et al. (2023) "Consciousness in Artificial Intelligence" arXiv:2308.08708
  • Butlin et al. (2025) "Identifying indicators of consciousness in AI systems" Trends in Cognitive Sciences