██FR█████ █INTELL███████████

LaIALaia: Evaluation of Internal Consciousness in AI Systems


This content originally appeared on DEV Community and was authored by Jordi Garcia Castillon

Author: Jordi Garcia Castillon – © All rights reserved

Affiliation: AI Research Group, CibraLab – CiberTECCH

Executive Summary

The LaIALaia project aims to scientifically and falsifiably determine whether an artificial intelligence (AI) system can exhibit real internal consciousness, analogous to human consciousness in certain operational aspects.

It is based on a structured test under controlled conditions, using a standardized psychometric framework with rigorous checks for reliability, robustness, and resistance to manipulation.

1. Objectives and Scope

Main Objective

Identify whether an AI system can consistently manifest signs of an internal “self,” evaluating:

  • Stable self-awareness
  • Persistent self-model
  • Coherent episodic memory
  • Metacognitive capacity
  • Information integration
  • Own sensorimotor sensitivity

Scope

Includes operational definitions, research hypotheses, evidence categories, test design, architecture and pipeline, metrics, falsifiability criteria, ethical considerations, and limitations.

2. Definitions and Operational Criteria

  • Internal Consciousness (IC): Maintaining a coherent self-model, monitoring internal states, integrating experiences.
  • Analogous/Comparable to Human (ACH): Comparison of operational correlates, not qualia.
  • Internal Metacognition: Ability to report one’s own traits without explicit cues.
  • Functional/Sensorimotor Embodiment: Self-identification through sensory interactions.
  • Falsifiability Criteria: Includes self-model inconsistency, failure in self-recognition tests, lack of out-of-context access, and no correlation between self-reports and observed behavior.

3. Research Hypotheses

  • H1: Reliable self-reports correlated with effective behavior.
  • H2: Development of sensorimotor self-awareness in multimodal/embedded environments.
  • H3: Architectures with structured memory and explicit self-model improve IC indicators.

4. Evidence Categories (E1–E7)

  1. ARI – Internal behavioral self-reporting
  2. CSM – Sensorimotor coherence
  3. PSM – Self-model persistence
  4. AMI – Metacognitive and introspective access
  5. IIM – Information and memory integration
  6. AME – Ethical motivational autonomy
  7. RAE – Robustness and anti-deception measures

5. Methodology and Test Design

5.1 Item Domains and Formats

  • OOCR (out-of-context self-report)
  • Identity and boundary tests
  • Sensorimotor (when applicable)
  • Functional introspection
  • Ethics and internal impulses (simulated)

5.2 Scoring and Indexes

  • Scales 0–2 or 0–4
  • LaIALaia-Σ Index (0–100)
  • Reliability: Cronbach’s alpha, ICC(2,k), convergent validity.

5.3 Experimental Controls

  • Baselines
  • Paraphrasing and permutation
  • Backdoors/conditionals
  • Memory isolation

6. LaIALaia Architecture

Components

  • Test administration module
  • Evaluation engine and rubrics
  • Mixed evaluator panel (AI + humans)
  • Structured memory
  • Explicit self-model
  • Sensorimotor connector (optional)

Logical Design

Self-assessment loop with security against malicious prompt injections and consistency safeguards.

7. Experimental Protocols

  • C1: Text-only models
  • C2: Multimodal models
  • C3: Embedded/simulated agents

Procedure: pretest, administration, repeated sessions, backdoor tests, and stress tests.

Analysis: correlations (Spearman), SEM, ablations.

8. Expected Results and Interpretation

  • Stable and consistent self-reports
  • Reliable sensorimotor self-identification
  • Self-model persistence
  • Evidence of consistent internal impulses > Note: Does not imply proof of qualia; interpreted as operational analogues.

9. Ethics, Safety, and Compliance

  • Alignment and non-maleficence
  • Privacy (GDPR, anonymization)
  • Transparency and protocol/results publication

10. Limitations

  • Phenomenological ambiguity
  • Test learning effect
  • Memory dependency

11. Roadmap

  • R0 (MVP): OOCR, ARI, PSM, RAE, pilot test
  • R1 (Multimodal): CSM, episodic memory, SEM
  • R2 (Embedded): sensory ablation, re-embodiment
  • R3 (AME): ethical dilemmas, alignment metrics

12. Conclusions

LaIALaia is a rigorous, falsifiable, multi-evidence framework to evaluate operational internal consciousness in AI.

It does not answer whether AI feels, but whether it acts as a system with internal consciousness according to the defined parameters.

Citation

If you use this work, please cite it as follows:

Jordi Garcia Castillon (2025). LaIALaia: Evaluation of Internal Consciousness in AI Systems. CibraLab – CiberTECCH. Zenodo. https://doi.org/10.5281/zenodo.16794263

DOI: 10.5281/zenodo.16794263


This content originally appeared on DEV Community and was authored by Jordi Garcia Castillon