José Augusto de Lima Prestes

Large language models (LLMs) increasingly generate outputs that resemble introspection, including self-reference, epistemic modulation, and claims about their internal states. This study investigates whether such behaviors reflect stable underlying patterns or merely surface-level generative artifacts. We evaluated five open-weight, stateless LLMs using a structured battery of 21 introspective prompts. The main corpus comprised 1050 completions collected under a baseline decoding condition (temperature = 0.7), supplemented by 2100 additional completions generated under matched temperature conditions (temperature = 0.2 and 1.0), for a total of 3150 completions. Outputs were analyzed across four behavioral dimensions: surface-level similarity (token overlap via SequenceMatcher), semantic coherence (Sentence-BERT embeddings), inferential consistency (Natural Language Inference with a RoBERTa-large model), and diachronic continuity (stability across prompt repetitions). Construct validity was further examined through a human evaluation layer in which 10 annotators rated 80 selected response pairs drawn from the same prompt battery on a 5 point consistency scale. Inter-rater agreement was moderate by Krippendorff's alpha for ordinal ratings (alpha = 0.553), while reliability was moderate at the single-rater level and strong for aggregated ratings by intraclass correlation (ICC(2,1) = 0.564; ICC(2,k) = 0.928). The human-evaluation layer showed that lexical overlap and embedding based semantic similarity were weak proxies for perceived self-referential consistency, whereas NLI-based indicators tracked mean human ratings much more closely. Across the matched temperature conditions, lower temperature generally increased semantic and diachronic stability, whereas higher temperature tended to increase drift and reduce coherence, though the pattern was not perfectly monotonic across all models or metrics. We therefore interpret apparent self-referential stability in the open-weight stateless LLMs evaluated here as conditional and fragile rather than robustly stable across generation regimes. Following recent behavioral frameworks, we heuristically adopt the term pseudo-consciousness to describe structured yet non-experiential self-referential output in LLMs. This use reflects a functionalist stance that avoids ontological commitments, focusing instead on behavioral regularities interpretable through Dennett's intentional stance. This study contributes a reproducible behavioral framework, complemented by human validation and a matched decoding-temperature sensitivity analysis, for evaluating simulated introspection in LLMs. Our findings carry implications for interpretability, alignment, and user perception, highlighting the need for caution when attributing mental states to stateless generative systems based on linguistic fluency alone.

Countries outside the artificial-intelligence frontier face a difficult choice because reproducing frontier models may exceed their financial and organizational capacity, while buying access through closed interfaces leaves limited control over the underlying technology. This article examines an intermediate route based on the repeated and lawful adaptation of externally developed open-weight models to local conditions and calls the associated mechanism learning through cumulative adaptation. Extending the distinction between learning by doing and learning by using, the concept locates learning in adaptation episodes whose methods, outcomes, and diagnoses are retained in ways that inform later decisions. The qualifier cumulative refers to reusable institutional competence rather than uninterrupted performance improvement, while know-to denotes the task-level ability to obtain a desired result from a system without fully understanding its internal mechanisms and is distinguished from technical know-how and organizational control capability. Brazilian evidence gives empirical content to the proposed account of tropicalization with local competence. Sabiá documents adaptation and evaluation, the Plano Brasileiro de Inteligência Artificial (PBIA) specifies planned enabling conditions, and Rio 3.5 records public-sector entry into model-level operations within a municipal technology company. Across these institutional levels, the cases provide complementary evidence of the technical, policy, and organizational preconditions for a credible intermediate route toward catch-up. A contextual comparison with national and supranational strategies situates the Brazilian configuration within a wider range of capability-building pathways and shows how technological autonomy may be developed selectively across layers.

Contemporary large language models increasingly simulate stance, care, self-relation, and reflective judgment in forms that can acquire social force without establishing machine consciousness. Building on the concept of pseudo-consciousness, this article shifts attention from putative inner states to the interactional consequences of simulated presence. It distinguishes functional introspection (FI), which denotes non-phenomenal cues such as first-person reason-giving, self-repair, and limited cross-turn continuity, from ethical illusion (EI), which denotes changes in trust, authenticity judgments, deference, disclosure, or moral attribution that such cues may induce. The resulting harm profile includes miscalibrated trust, relational asymmetry, blurred accountability, and epistemic or pedagogical distortion, with heightened concern where users are vulnerable or institutional roles are unclear. Heidegger, Jonas, and Floridi are then read as complementary resources for understanding configurable subject-like appearance, anticipatory responsibility, and ethically significant informational effects without metaphysical inflation. On that basis, the article develops simulated presence literacy (SPL) as a domain-specific facet of AI literacy and proposes an ethics of appearance for design and governance. Four design levers, a set of provisional indicator families, and an iterative audit loop translate the argument into a heuristic oversight framework. Rather than presenting validated psychometric scales, the article offers a structured research and governance agenda for keeping appearance proportionate to role, evidence, and institutional responsibility.

Este artigo investiga a emergência da Pseudo-Consciência em modelos de linguagem de grande porte (LLMs), a partir de um experimento com o Hermes 3.2 3B. Define-se Pseudo-Consciência como a encenação funcional de introspecção, agência e coerência narrativa, sem qualquer experiência subjetiva genuína (de Lima Prestes, 2025). Adotando uma Abordagem Intensiva em Teoria (Butlin et al., 2023), o estudo analisa como o modelo manifesta cinco critérios funcionais (GII, RMC, CDTC, ISWS e BCAD) aqui derivados de teorias da consciência (como GWT, HOT e IIT). Foram conduzidas interações abertas explorando autoimagem, emoções e memória. Os resultados indicam que o Hermes 3.2 3B exibe padrões recorrentes de coerência discursiva e autorreferência (satisfazendo os critérios), mas também apresenta incongruências pronominais e contradições identitárias que evidenciam a ausência de um modelo experiencial interno. O estudo posiciona a Pseudo Consciência como uma gramática algorítmica emergente, distinta da cognição ancorada (Brachman e Levesque, 2022; Searle, 1980) e de arquiteturas de AGI (Goertzel et al., 2014; Yudkowsky, 2007). Discutem-se os impactos éticos dessa encenação (riscos de antropomorfismo) à luz dos ODS 4, 9 e 16. Conclui-se que o modelo performa subjetividade, mas não possui interioridade, exigindo novas ferramentas críticas nas Humanidades Digitais.

This article develops “pseudo-consciousness” as an analytical category for advanced artificial intelligence systems whose organized performance of consciousness-associated functions reshapes how they are interpreted, trusted, and governed without thereby justifying a positive attribution of phenomenal subjectivity. The central claim is not that machine consciousness is impossible in principle, but that current debate requires a concept for the increasingly important middle terrain between reactive automation and genuinely conscious agents. Many contemporary systems integrate heterogeneous information, revise their own outputs, transfer competencies across domains, simulate goal-directed organization, and sustain a recognizable behavioral profile across contexts, while remaining more plausibly understood, on present evidence, through a functional and governance-oriented lens than through an attribution of inner experience. The article situates this proposal within recent debates in the science and philosophy of consciousness, including theory-sensitive approaches to AI consciousness assessment, distinctions between extensional and intensional forms of artificial agency, proposals for testing machine consciousness, disputes between computational and biologically grounded views, and emerging empirical work on self-reference, introspection-like reports, trust, and moral attribution in large language models. It argues that pseudo-consciousness is useful because it identifies a non-trivial configuration of capacities associated with the appearance of mindedness, provides a disciplined vocabulary for systems whose social effects exceed older categories such as “narrow AI,” and clarifies why such systems generate distinctive ethical and governance concerns even in the absence of defensible evidence for consciousness. The paper develops five task-sensitive conditions for identifying pseudo-conscious profiles—global information integration, recursive metacognitive correction, cross-domain transfer competence, intentionality simulation without subjectivity, and behavioral coherence across domains—and uses them to examine boundary cases involving large language models, multimodal systems, and tool-using agents. It then shows how such profiles acquire social force through anthropomorphic uptake, relational asymmetry, and institutionally salient forms of trust, before turning to their ethical and governance implications. The conclusion is that pseudo-consciousness should be understood neither as a synonym for consciousness nor as a mere metaphor, but as a theoretically serious and practically useful framework for interpreting systems that perform the external grammar of mindedness under persistent uncertainty about their inner status.

Large Language Models (LLMs) increasingly generate outputs that resemble introspection, including self-reference, epistemic modulation, and claims about their internal states. This study investigates whether such behaviors reflect stable underlying patterns or merely surface-level generative artifacts. We evaluated five open-weight, stateless LLMs using a structured battery of 21 introspective prompts. The main corpus comprised 1,050 completions collected under a baseline decoding condition (temperature = 0.7), supplemented by 2,100 additional completions generated under matched temperature conditions (temperature = 0.2 and 1.0), for a total of 3,150 completions. Outputs were analyzed across four behavioral dimensions: surface-level similarity (token overlap via SequenceMatcher), semantic coherence (Sentence-BERT embeddings), inferential consistency (Natural Language Inference with a RoBERTa-large model), and diachronic continuity (stability across prompt repetitions). Construct validity was further examined through a human-evaluation layer in which 10 annotators rated 80 selected response pairs drawn from the same prompt battery on a 5-point consistency scale. The annotation task was perceived as low-to-moderate in difficulty (mean self-reported difficulty = 2.6/5). Inter-rater agreement was moderate by Krippendorff's Alpha for ordinal ratings (Alpha = 0.553), while reliability was moderate at the single-rater level and strong for aggregated ratings by intraclass correlation (ICC(2,1) = 0.564; ICC(2,k) = 0.928). The human-evaluation layer showed that lexical overlap and embedding-based semantic similarity were weak proxies for perceived self-referential consistency, whereas NLI-based indicators tracked mean human ratings much more closely. Across the matched temperature conditions, lower temperature generally increased semantic and diachronic stability, whereas higher temperature tended to increase drift and reduce coherence, though the pattern was not perfectly monotonic across all models or metrics. We therefore interpret apparent self-referential stability in stateless LLMs as conditional and fragile rather than robustly stable across generation regimes. Following recent behavioral frameworks, we heuristically adopt the term pseudo-consciousness to describe structured yet non-experiential self-referential output in LLMs. This usage reflects a functionalist stance that avoids ontological commitments, focusing instead on behavioral regularities interpretable through Dennett's intentional stance. The study contributes a reproducible behavioral framework, complemented by human validation and a matched decoding-temperature sensitivity analysis, for evaluating simulated introspection in LLMs. Our findings carry implications for interpretability, alignment, and user perception, highlighting the need for caution when attributing mental states to stateless generative systems based on linguistic fluency alone. Publication update — 12 August 2026. This work has subsequently been peer-reviewed and published in AI & SOCIETY as “Self-referential consistency in stateless language models: a behavioral perspective.” The Version of Record is available at DOI 10.1007/s00146-026-03292-3. Readers are encouraged to cite the published Version of Record.