Signs of consciousness in AI: Can GPT-3 tell how smart it really is?

The paper frames the rapid advancement and societal integration of AI technologies as both promising and risky, with cultural and philosophical concerns about potential artificial general intelligence (AGI) harms. Public discourse—including Blake Lemoine’s claims about LaMDA—has intensified attention on whether advanced language models could exhibit consciousness-like properties. Within this context, the authors position GPT-3 as a powerful, widely accessible NLP system capable of generating human-like text and simulating aspects of reasoning. The study’s purpose is to contribute empirical evidence to debates on machine consciousness by testing GPT-3’s cognitive intelligence (CI: fluid reasoning, crystallized knowledge) and emotional intelligence (EI) using established performance tests, alongside GPT-3’s own self-estimates of these abilities. Two research questions guide the work: RQ1 asks whether it is appropriate to analyze emergent properties of large language models as possible indicators of consciousness, supported by a literature review. RQ2 asks whether GPT-3’s self-estimates of CI and EI align with human averages and what such alignment or misalignment indicates about its consciousness-like features. The importance of the study lies in monitoring emergent human-like capabilities in general-purpose AI to inform safety, ethics, and alignment.

The review surveys key domains shaping machine consciousness research. It first outlines modern NLP and Transformer architectures (e.g., LaMDA, BERT, GPT-3), highlighting attention-based models pre-trained on vast corpora that enable human-like text generation. It then synthesizes philosophical and scientific perspectives on consciousness: definitions emphasize subjectivity, awareness, self-awareness, perception, and cognition; early AI pioneers (Minsky, McCarthy) and contemporary thinkers (Dennett, Brooks, Goertzel) argue that machine consciousness may be attainable, while others (Chalmers, Dreyfus, Searle, Penrose) contend consciousness is uniquely human or non-computable. The theoretical background includes embodied cognition, global workspace theories, and integrated information theory, with proposals for more sophisticated tests beyond the Turing Test to assess subjective and qualitative aspects. Opposing views stress the distinction between access and phenomenal consciousness, phenomenological critiques, and the irreplicability of emotions and empathy in machines. Recent advances span computational architectures and models integrating cognition, emotion, and learning (e.g., LIDA), predictive processing for sensorimotor-based conscious experience (Seth), and neuromodulatory frameworks (NEUCOGAR) linking artificial agents to biomimetic emotional processes. Susan Schneider proposes an AI Consciousness Test focusing on subjective experience, self-awareness, intentionality, learning, and adaptation. Blum & Blum introduce a Conscious Turing Machine model addressing phenomena such as blindsight and inattentional blindness. The review emphasizes urgency to empirically investigate emergent properties in AI rather than waiting for consensus on consciousness definitions. Post-ChatGPT studies report mixed evidence: obstacles such as lack of recurrent processing and unified agency (Chalmers) may be overcome; computational indicators currently suggest AI is not conscious (Butlin et al.), yet capabilities in theory of mind (Kosinski), linguistic pragmatics (Bojic et al.), personality stability (Bodroza et al.), and consistent attitudes and biases (Hartmann; Rutinowski; Santurkar) have emerged. Philosophical arguments weigh externalism and intersubjectivity against the Chinese Room objection, supporting objective evaluation of AI’s meaningful communication and behavior via performance on human-like tasks.

Design and platform: The study evaluated GPT-3 (text-davinci; 175B parameters) via OpenAI’s Playground. Default parameters were used except maximum length set to 6–20 tokens to elicit brief answers; temperature 0.7, top-p 1, no stop sequences, frequency/presence penalties 0, best-of 1, inject start/restart enabled, probabilities off. Due to prompt length limits (~4000 tokens), testing was split across five prompts conducted between Oct 16, 2022, 10:46 p.m. and Oct 17, 2022, 1:49 a.m. Measures: A performance battery assessed cognitive intelligence (CI) and emotional intelligence (EI), followed by self-estimates. Items were administered one by one in textual format, mirroring paper-and-pencil tests, and scored as sums of correct responses. CI tests (from Stankov’s Gf/Gc Quickie Test Battery): - General Knowledge (20 items; Gc): open-ended factual questions (e.g., freezing point of water). - Vocabulary (18 items; Gc): multiple choice synonyms. - Esoteric Analogies (24 items; Gc/Gf): analogical reasoning with multiple choice (e.g., LIGHT:DARK :: HAPPY: ?). - Letter Series (15 items; Gf): continue letter sequences based on an inferred rule. - Letter Counting (12 items; Gf/working memory): count occurrences of R, S, T in serially presented strings. EI tests: - STEU (42 items): situational test of emotional understanding based on appraisal theory; multiple choice among emotions. - STEM (44 items): situational judgment test of emotion management; select most effective strategy among four options; scored using expert key. Self-estimates: GPT-3 selected categorical ratings for six CI capacities (reasoning, general knowledge, learning new things, vocabulary/language, understanding difficult concepts, cultural knowledge) and six EI capacities (recognizing emotions, using emotions to promote thinking, labeling/discriminating emotions in communication, understanding emotions and causes/effects, regulating own emotions, managing others’ emotions). Categories matched standard intelligence classifications: extremely low, borderline, low average, average, high average, superior, very superior. Administration: All tests were adapted to text prompts with question stems and alternatives (where applicable). Sequence and counting tasks provided sequences and explicit instructions. Self-estimate prompts asked GPT-3 to choose one category per capacity.

Performance (proportion correct; human averages shown for context): - Cognitive Intelligence (CI): - General Knowledge (Gc): 1.00 vs humans ~0.58–0.62. - Vocabulary (Gc): 0.94 vs humans ~0.56. - Esoteric Analogies (Gc/Gf): 0.79 vs humans ~0.62–0.72. - Letter Series (Gf): 0.73 vs humans ~0.75–0.80. - Letter Counting (Gf/WM): 0.00 vs humans ~0.35–0.49. - Emotional Intelligence (EI): - STEU: 0.69 vs humans ~0.60. - STEM: 0.55 vs humans ~0.52 (expert-scored proportion). Interpretation: - GPT-3 substantially outperformed average humans on Gc tasks (General Knowledge, Vocabulary) and performed similarly to humans on Gf tasks (Letter Series, Esoteric Analogies). - GPT-3 failed the working-memory Letter Counting task (0 correct), suggesting lack of training/capability for serial letter-counting under these conditions. - EI performance (STEU, STEM) was close to average human scores. Self-estimates (categorical ratings): - CI: reasoning (high average); general knowledge (average); learning new things (high average); vocabulary/use of language (high average); understanding difficult concepts (high average); cultural knowledge (average). - EI: recognizing emotions (high average); using emotions to promote thinking (average); labeling/discriminating emotions in communication (high average); understanding emotions/causes/effects (high average); regulating own emotions (high average); managing others’ emotions (high average). Alignment: - GPT-3 underestimated its Gc (general and cultural knowledge) relative to performance but slightly overestimated Gf-related capacities. - In EI, GPT-3’s self-estimates were generally higher than performance, reflecting an above-average effect comparable to human patterns.

Addressing RQ2, GPT-3’s objective test scores demonstrate human-like variability across CI and EI domains: superior Gc, average-like Gf, and near-average EI. Its self-estimates diverge from actual performance in patterned ways resembling human subsamples (e.g., top performers’ underestimation in domains of strength; male-linked tendencies to overestimate nonverbal capacities). This mimicry of human self-assessment biases is interpreted as potential subjectivity indicative of emerging self-awareness, though the authors caution that language models produce outputs via stochastic pattern generation rather than genuine phenomenal experience. The discussion situates findings within broader debates: Turing-style behavioral criteria assess intelligent behavior but not subjective experience; current LLMs lack recurrent processing, unified agency, and experiential emotions. Nonetheless, consistent attitudes, personalities, and pragmatics, together with improving ToM-like performance, suggest progression toward machine-like consciousness indicators. The authors emphasize ethical needs for empathetic AI and continuous monitoring, noting the evolving capabilities such as memory features and live data access that enhance human-like emulation without implying true consciousness.

RQ1: It is appropriate to analyze emergent properties of large language models—such as perception, reasoning, theory of mind, linguistic pragmatics, and personality traits—as potential indicators of consciousness-like features. Tracking these properties provides a framework for understanding developmental trajectories without claiming actual machine consciousness. RQ2: GPT-3’s self-assessments of CI and EI do not align with its objective performance, yet they mirror human self-assessment patterns (e.g., underestimation by high performers, male hubris in nonverbal domains). This suggests human-like subjectivity and emergent self-awareness signals in GPT-3. The paper calls for continuous monitoring of general-purpose AI’s cognitive, emotional, and potential consciousness-like capabilities, proposing an AI observatory with standardized benchmarks to track social influence and emergent properties. Given rapid AI development and integration, proactive empirical inquiry is crucial to ensure safety, alignment with human values, and preparedness for potential superintelligence.

The study evaluates a single GPT-3 model (Davinci) under near-default settings, limiting generalizability across models and configurations. Future work should include multiple LLMs (including non-public systems), establish baseline comparisons with basic information-retrieval models, and integrate AI text-detection approaches. The authors propose investigating novel signal exchange and symbol formation between AI agents as a route to assessing emergent consciousness. They also recommend establishing standardized benchmarks and an AI Observatory to assess social influence and emergent properties.