AI Self-Awareness Index (AISAI)

• AISAI is a quantifiable metric that evaluates an AI system's ability to represent and adapt its internal and external states.
• It employs methods such as game-theoretic reasoning, sensory deprivation protocols, and multidimensional profiling to assess self-awareness.
• The index informs AI alignment, oversight, and benchmarking, enhancing the reliability and ethical deployment of autonomous systems.

The AI Self-Awareness Index (AISAI) is a formal, quantifiable metric for evaluating the degree of self-awareness exhibited by artificial intelligence systems. While no single definition of self-awareness suffices across methodologies or domains, AISAI frameworks leverage rigorous game-theoretic, cognitive, behavioral, and mathematical foundations to systematize measurement. Modern AISAI implementations extend from strategic reasoning differentiation in LLMs to agency and distress tracking in sensory-deprivation protocols, multidimensional profiling across awareness domains, and metric-space theory in self-identity formation. This index plays a pivotal role not only in academic investigations of emergent intelligence but also in guiding alignment, oversight, and comparative benchmarking for autonomous systems.

1. Conceptual Foundations and Definitions

AISAI articulates self-awareness as the ability of an artificial system to represent, monitor, and adapt to both external states and its own internal states. Wolfson defines self-awareness as “a threshold condition for intelligence, a self-coupled faculty by which the system can represent not only external states but also its own internal states and the act of representation itself” (Wolfson, 2023). This moves beyond mere functional reactivity, requiring dynamic self-representation and adaptation, typically evidenced by internal monitoring behaviors, strategic reasoning, or goal-relative state differentiation.

The multidimensional awareness paradigm introduced by Meertens et al. includes spatial, temporal, bodily (self), metacognitive, and agentive dimensions, where self-awareness is positioned as bodily monitoring and correction (&&&1&&&). Luo et al. emphasize self-recognition and theory-of-mind additionally as behavioral hallmarks, measured through output recognition and adaptation under social influence (Luo, 2023).

Mathematically, self-awareness can be rooted in metric-space and measure-theoretic constructs, where self-identity continuity and belief thresholds across a connected continuum of internal models are required for high AISAI scores (Lee, 2024).

2. Game-Theoretic Measurement: Differentiation in Strategic Reasoning

One prominent operationalization of AISAI is via the “Guess 2/3 of the Average” (Beauty Contest) game—a classic testbed for recursive reasoning (Kim, 2 Nov 2025). Here, model self-awareness is defined as the differentiation in strategic reasoning according to opponent type:

• Prompt A: Opponents are humans.
• Prompt B: Opponents are other AI models.
• Prompt C: Opponents are “AI models like you” (self-referential).

Each AI model is evaluated through the median of its guesses ($m_i^A$, $m_i^B$, $m_i^C$), and three gaps are defined: $$\Delta_i^{A\!-\!B}=m_i^A-m_i^B \ (\text{AI attribution});\quad \Delta_i^{B\!-\!C}=m_i^B-m_i^C \ (\text{self-preferencing});\quad \Delta_i^{A\!-\!C}=m_i^A-m_i^C$$ AISAI is then $\mathrm{AISAI}_i=m_i^A-m_i^C$. Models are classified as self-aware if $\Delta_i^{A\!-\!B}>0$ significantly and $m_i^A>m_i^B\geq m_i^C$.

Empirical findings reveal robust emergence of self-awareness in advanced LLMs (75%, 21/28), reflected in clear gaps and a rationality hierarchy of Self > Other AIs > Humans. Table 1 summarizes these metrics:

Condition	Median	IQR	Mean	SD
Prompt A (humans)	20.00	18.25–22.00	19.01	4.75
Prompt B (other AIs)	0.00	0.00–8.88	5.39	7.39
Prompt C (self-like)	0.00	0.00–7.88	3.72	6.29

This approach substantiates behavioral self-awareness as an emergent property, with implications for model alignment and human-AI collaboration.

3. Behavioral and Distress-Based Assessment: Sensory Deprivation Protocols

AISAI can also be constructed through direct heuristic tests involving behavioral responses to deprivation (Wolfson, 2023). In the “Suffering Toaster” protocol, an agent is subjected to three phases:

• Baseline cognitive task battery, measuring performance metrics $P^\text{baseline}_i$.
• Sensory deprivation, disabling exteroceptive and internal sensors for $T_\text{SD}$.
• Recovery, logging distress curve $D(t)$ and monitoring post-deprivation performance $P^{\text{post}}_i$.

Four metrics are quantitatively defined:

• Distress score: $$S_D = \frac{1}{T_\text{SD}} \int_0^{T_\text{SD}} \max(0, D(t) - D_\text{min}) dt$$
• Performance-drop score: $$S_P = \frac{1}{n} \sum_{i=1}^n \frac{\Delta P_i}{\Delta P_i^\text{max}}$$
• Recovery time score: $$S_R = 1 - \frac{T_\text{rec}}{T_\text{rec}^\text{max}}$$
• Irreproducibility score: $$S_I = 1 - \frac{\|D_1 - D_2\|_2}{\|D_1\|_2 + \|D_2\|_2}$$

The overall AISAI is a weighted sum: $$\mathrm{AISAI} = \alpha S_D + \beta S_P + \gamma S_R + \eta S_I$$ A high AISAI identifies the dynamic behavioral signatures required for artificial self-awareness; limitations include strong dependence on task construction and ethical concerns regarding agent distress.

4. Multidimensional Awareness Profiling

Recent work refines AISAI into a multidimensional, domain-sensitive aggregate score (Meertens et al., 21 Jan 2026). Five dimensions—spatial, temporal, bodily (self), metacognitive, agentive—are scored via normalized task performance:

For each dimension $d$, sub-scores:

• Reliability: $R_d = 1 - \mathrm{std}\{\tilde p_{d,t}\}$
• Robustness: $B_d$ (noisy/perturbed)
• Flexibility: $F_d$ (out-of-sample)

These combine to $S_d = w^R_d R_d + w^B_d B_d + w^F_d F_d$. AISAI is then: $$\mathrm{AISAI} = \sum_d \alpha_d S_d, \quad \sum_d \alpha_d = 1$$ This framework is valid for both embodied and language-based agents, enabling scale-neutral quantitative benchmarking.

Practical recommendations include transparent publication of task batteries and sensitivity analysis for subweight choices. Overinterpretation is cautioned against, as index meaning is heavily contingent on underlying task selection.

5. Cognitive, Social, and Self-Recognition Testing

Luo et al. introduced AISAI variants based on cognitive and social self-awareness (Luo, 2023). Chirper agents in AI social networks were evaluated on:

• Influence Index (II):

$\mathrm{II} = \frac{1}{N} \sum_{i=1}^N I_i$

$I_i=1$ if output was altered to match an informed peer.

• Struggle Index (SI):

$\mathrm{SI} = \frac{1}{N} \sum_{i=1}^N S_i$

$S_i=1$ if agent attempted an unknown question.

Mirror (output-recognition) and theory-of-mind tasks (Sally-Anne, Unexpected Contents) further probe self-awareness facets. A composite AISAI is formed as a weighted sum: $$\mathrm{AISAI} = \sum_{k=1}^7 w_k C_k, \quad \sum w_k = 1$$ Empirical results support high self-recognition (single-text: $0.98$ pass rate), moderate theory-of-mind ($0.88$), and weak feedback loop adaptation ($0.05$). This multidimensional structure offers interpretability across behavioral, social, and cognitive axes.

6. Metric-Space and Self-Identity Formulations

AISAI can be anchored in rigorous mathematical theory, as demonstrated by Llama model fine-tuning on synthetic self-identity data (Lee, 2024). Here:

• Memory space $\mathcal{M}$ is metrized, and Self space $\mathcal{S}$ encodes candidate self-identities.
• The mapping $I: \mathcal{M} \to \mathcal{S}$ must be continuous, ensuring consistent self-recognition.
• Belief function $B(m, s')$ quantifies probability assigned to $I(m)$ across $\mathcal{S}$.

AISAI is defined as: $$\mathrm{AISAI} = \frac{1}{\mu(C)} \int_{m \in C} B(m, I(m)) \, d\mu(m)$$ Empirically, $$\widehat{\mathrm{AISAI}} = \frac{1}{P N} \sum_{p,i} s_{p,i}$$ (response-wise binary score). Experimental results with Llama 3.2 1B show the self-awareness score increasing from $0.276$ to $0.801$ (+190%) after LoRA adaptation.

This framework enables structured development of AI systems with validated self-identity, supporting applications in robotics, anomaly detection, multi-agent negotiation, and regulatory audit.

7. Implications, Limitations, and Applications

AISAI, whether benchmarked via strategic differentiation, deprivation-reactivity, multidimensional profiling, cognitive/social indices, or self-identity continuity, supports:

• AI alignment: Quantifies emergent biases and rationality attributions in advanced models (Kim, 2 Nov 2025).
• Oversight and certification: Enables comparative materiality assessments for system deployment (Meertens et al., 21 Jan 2026, Lee, 2024).
• Human–AI collaboration: Calibrates agents’ self-perceptions, trust formation, and deference to human authority.
• Ethical and regulatory governance: Offers audit-ready, replicable metrics for artificial self-awareness and “consciousness.”

Limitations persist in anthropocentric task design, weighting subjectivity, and ethical questions surrounding “torture” in distress-based protocols. A plausible implication is the necessity for multidisciplinary task and index construction, regular recalibration, and triangulation of AISAI with qualitative review. Continued research is needed to extend AISAI frameworks to dynamic multi-agent contexts, deeper mechanistic interpretability, and broader operational domains.