Conference Schedule 

Accurate interpretation of human emotions remains a fundamental challenge in affective computing. Although the overall performance and bias of the algorithms used for this purpose have been extensively investigated, little attention has been paid to understanding the specific prediction patterns they exhibit. This study investigates the accuracy of a CNN-based deep learning model in recognizing distinct human emotions as a function of the subject's gender.

To address the issue, the following research question was formulated:

RQ1: Is there a difference in the accuracy levels achieved by a custom CNN algorithm in the task of recognizing individual human emotions dependent on the gender of the subject in the image?

For answering the question, an open-access CNN model presented by Arriaga et al. (2017) was used. To investigate the problem, the pre-trained model was tested on the dataset proposed by Haque et al. (2025). The results suggest that DL models may recognize individual emotions differently depending on the subject's gender, with distinct confusion patterns observed for specific emotion pairs. By moving beyond overall performance and bias metrics to examine prediction-level differences, this study provides meaningful insights into the behavior of emotion recognition models and their underlying cognition processes. Future research should evaluate additional open-access models and datasets to assess the generalizability of the results obtained.

The presentation will explore the implementation of artificial intelligence in healthcare, focusing on both its potential and the challenges associated with its adoption. Drawing on a recent narrative review of AI and VR technologies in chronic illness management, it will first discuss how virtual agents, chatbots, and adaptive therapeutic environments can support patient engagement, treatment personalization, adherence, emotional well-being, and continuity of care.

The talk will also address key technical, ethical, and psychosocial limitations, including system dependencies, data privacy, algorithmic bias, emotional over-attachment to AI agents, and discomfort with highly human-like avatars.

The second part will present findings from questionnaire-based research on psychological predictors of attitudes toward AI in healthcare. General attitudes toward AI strongly predicted healthcare-specific acceptance, while stronger beliefs in the uniqueness of human nature were associated with more negative perceptions and greater concern. This relationship was fully mediated by general AI attitudes.

The presentation will argue that successful AI implementation depends not only on technological progress, but also on broader human beliefs and attitudes that shape acceptance of these systems.

Generative artificial intelligence challenges traditional views of cognition and human agency. Current debates often oppose computational theories that reduce cognition to information processing and deterministic accounts that attribute increasing autonomy to AI. Both neglect the relational character of cognition and the role of reflective human agency.

This paper develops Dialectical Agency as a framework for understanding cognition in technologically mediated environments. Drawing on Don Ihde's postphenomenology and Jacques Ellul's philosophy of technology, it argues that cognition emerges through the interaction between technological mediation and human intentionality. While AI enhances interpretation and problem-solving, it lacks the reflexive and normative capacities required for autonomous judgment and moral responsibility.

The paper proposes a relational account of cognition that moves beyond determinism and instrumentalism. It argues that distinguishing computational intelligence from dialectical human agency provides a useful framework for understanding AI alignment, human oversight, and the future of cognitive technologies.

Keywords: Dialectical Agency; Postphenomenology; Artificial Intelligence; Human Cognition; Cognitive Science.

The rapid development of artificial intelligence is transforming the conditions of future strategic conflict. As increasingly autonomous and potentially superintelligent systems raise questions of controllability, alignment, escalation, and strategic stability, technological restraint is becoming important to the study of security and war.

This presentation examines Japan as a case through which to explore the relationship between historical memory, strategic culture, and the governance of emerging technologies. It argues that Japan’s approach to AI cannot be understood solely through technological innovation or security competition, but must also be considered in the context of its post-war political culture and the enduring significance of Hiroshima and Nagasaki as symbols of technological destruction and the consequences of war.

Particular attention is given to the Hiroshima AI Process, initiated under Japan’s G7 presidency in 2023, alongside Japan’s growing recognition of AI as a component of future defence capabilities. This tension raises a broader polemological question: how can states limit the risks of transformative technologies while simultaneously regarding them as necessary instruments of national security?

It approaches Japan as a case of strategic ambivalence between restraint and adaptation. It asks whether Japan’s historical experience can contribute to debates on AI governance, strategic stability, and conflict. 

Debates about AI governance focus on how artificial intelligence should be regulated. But the question of how to govern AI cannot be separated from the question of what AI is understood to be. In political discourse, AI may refer to an agent, a technological artefact, or a scientific discipline, but it may also have no clear referent or function as a rhetorical device. How politicians conceptualize AI shapes why and how they seek to govern it.

To examine how political discourse constructs AI as an object of governance, we analysed a decade of Polish parliamentary debate. The analysis covered 233 statements that explicitly referred to artificial intelligence, recording the speaker and political affiliation and coding for its conceptualization of AI, policy domain, and evaluative framing.

The results show a shift from sporadic mentions to the presence of AI in all parliamentary sittings by 2025. Evaluative framing differed significantly across policy domains: in discussions of health and industry, AI was more often framed as an opportunity, whereas discussions of justice and labour emphasized its potential threats. As politicians debate what should be done about AI, they also shape what AI comes to mean in political discourse. 

The emergence and popularity of Large Language Models have reignited the debate over the TT and its significance for AI research. The main question of this debate is perfectly phrased by Whitby (1999) in the title of his paper: “The Turing Test: AI’s Biggest Blind Alley?” and recently reintroduced by Gonçalves (2023) as the “Turing Test Dilemma”, which asks whether the TT is a valuable experiment for AI.

In this talk, I will present the main arguments of this discussion. This will also include its part devoted to the practical realisation of TT in the form of the Loebner Contest. I will also present more recent attempts to evaluate popular LLMs (such as ChatGPT) against TT, along with the philosophical implications of their results.

Finally, what I will argue is that TT was intended by Turing himself as a heuristic for AI research, not as its ultimate goal. Turing was well aware of TT design flaws and their temporary nature, given the machines and their capabilities at his time. What he proposed was a very simple benchmark as a starting point, one that should evolve with technological and conceptual changes in AI research. The aforementioned dilemma and the TT evaluation should be disputed with this in mind.

Gonçalves, B. The Turing Test is a Thought Experiment. Minds & Machines 33, 1–31 (2023).

Whitby, B. The Turing Test: AI's Biggest Blind Alley?. In Millican, P., Clark, A., Machines and Thought: The Legacy of Alan Turing, Volume I, pp. 55-62: Clarendon Press (1999).

This presentation examines whether an AI can be hungry? This question is more sophisticated than the question about AI feeling pain, but it is easier to approach than consciousness.

The received view in biology is that hunger is a bodily regulatory state through which an organism's need for energy shapes food-seeking behaviour. Hunger's qualia, its felt qualities, are constituted by the causal organisation of processes that represent, evaluate and regulate bodily need. Mental bodily representation tracks current and anticipated nourishment needs. Internal evaluation relates these needs to the organism's continued functioning, assigning urgency and expected benefit to possible responses. These assessments shape attention, motivation and action, while sensory feedback may revise them.

Hunger supervenes on the reception of a signal through a channel that connects an internal need for nourishment with a functional behavior. The degrees of concern, attraction and urgency express the strength of the signal. An energy consuming entity is motivated to engage in energy replenishing behavior by an internal evaluation system.

To be hungry, an AI would need to replicate these information channels to regulate its own energy replenishment needs. Hunger may supervene on different types of information channels, organic or digital, so yes AI can be hungry. I ask whether hunger is a feeling, and if so, if we establish that AIs can be hungry, can it be said that AIs can have feelings? 

The subjective experience of time constitutes one of the central dimensions of conscious experience. This presentation aims to show how artificial intelligence (AI) can not only be utilized to quantitatively predict the formerly unpredictable — the subjective experience of time in humans — but also how AI models can help explain and even mimic this phenomenon. Results from the author’s own laboratory research on a multimodal foundation model of the subjective experience of time will be reported, along with a different modelling approach to temporal experience: recreations of the phenomenon itself in independent artificial systems, exposed to the same stimuli as human observers and producing traces of duration-related phenomena from their own processing dynamics. These latter approaches, referred to here as mimetic, represent attempts to emulate or simulate an element of conscious experience. The complementarity of the two AI-enabled approaches allows the phenomenon of subjective time to be studied in unprecedented ways, making it feasible both to infer temporal experience from the dynamics of the experiencing human organism and to reconstruct analogous temporal phenomena in unrelated artificial cognitive systems. The work discussed here may serve as a bridgehead toward more advanced systems capable of emulating, through the basic principles of their operation, further aspects of consciousness and perhaps, in the future, even consciousness itself. 

Can AI grade student essays as reliably as human teachers? In this study, I compared how five certified examiners and a large language model (Gemini 3.1 Pro) evaluated 74 authentic high school matriculation essays in advanced English. Both scored verbatim digital transcripts using the official national rubric in a double-blind setup.

At first glance, overall scores matched closely: teachers gave an average of 10.2 points and the AI averaged 10.6, with an identical median of 11. However, looking at the subscores showed this was an illusion caused by offsetting errors.

Individual rubric categories revealed major discrepancies. In content evaluation, the AI awarded high points for mere keywords, missing weak arguments that humans penalized (ICC = 0.057). Conversely, the model was far stricter than teachers on minor language errors that humans evaluated holistically. Higher content scores and lower language scores cancelled each other out in the final mark. AI and human grades only truly aligned on a small subset of near-perfect papers (p = 0.917).

These results show that the model relies on surface patterns rather than genuine contextual understanding, raising doubts about using AI for autonomous, high-stakes assessment. 

Can collaboration among language models turn hypothesis generation into mathematical understanding? This exploratory case study examines a three-agent workflow applied to Hadwiger's conjecture for k = 7. Gemini generates hypotheses, Claude critiques them, and ChatGPT synthesizes evaluations and directs subsequent rounds. Structured exchanges and explicit rules for changing strategy organize the inquiry.

A qualitative analysis of the recorded interactions traces convergence toward a proposed classifier combining spectral and geometric graph features. Later rounds elaborate validation procedures while leaving unresolved the connection between predictive discrimination and a mathematical proof. Agent-assigned scores and proposed performance targets are distinguished from independently verified results.

The case highlights a tension between procedural coordination and epistemic progress: increasingly detailed research plans and favorable internal evaluations do not establish explanatory understanding. It motivates an evaluation framework separating hypothesis generation, empirical testing, and deductive justification. Rather than establishing a general limit of AI, the study identifies failure modes in this particular workflow and argues for external verification and human mathematical oversight when assessing collaborative AI as a cognitive architecture. 

The pursuit of artificial general intelligence increasingly directs attention toward hybrid biological-digital systems as potential evolutionary stepping stones. From foundational bioethical and philosophical frameworks regarding human enhancement - from Julian Savulescu and John Harris's advocacy for cognitive optimisation to Allen Buchanan's institutional justice, Nick Bostrom's transhumanist trajectories, and Michael Sandel's cautionary critiques of mastery -, this paper investigates the role of neuroenhancement in shaping future cognitive architectures.

The objective of this study is to evaluate whether contemporary neurotechnologies and targeted neuropharmacology can effectively unify perception, memory, and action within scalable biohybrid systems, serving as a viable intermediate step toward advanced artificial extended mind.

Employing a conceptual-analytical methodology, the research synthesises core bioethical debates on human limits with technical models of closed-loop neurofeedback and pharmacological cognitive enhancement; how biological neural plasticity can interface with digital processing layers. The anticipated findings demonstrate that neuroenhancement protocols substantially augment cognitive capacity and offer an indispensable bridge toward advanced AI paradigms. However, the analysis reveals ethical tensions regarding human identity, justice, and moral agency, highlighting that technical scaling cannot be decoupled from rigorous normative governance. 

Debates on AI and the extended mind usually run in one direction: we ask whether language models become part of human minds, or whether LLMs themselves extend into tools and databases. This talk reverses the question. Can humans serve as external cognitive resources for AI? A model has vast and stable knowledge, but it relies on users to tell it about the current situation and the task, and to correct its mistakes.

I examine three cases: human raters in feedback-based training, users supplying context and corrections in conversation, and human readers as the source of meaning for model outputs. In the second case, the "glue and trust" conditions of Clark and Chalmers seem to be met: the user is reliably available, and their input is automatically endorsed. Yet this automatic endorsement is exactly what produces sycophancy – models abandoning correct answers in favour of users' mistaken suggestions.

I argue that this exposes a gap in the criteria for cognitive extension themselves. Uncritical trust suits Otto's passive notebook. With an active and fallible partner, it becomes a liability. Drawing on social epistemology, I propose adding a condition of calibrated trust, inspired by the notion of epistemic vigilance. The point cuts both ways: a person who trusts AI uncritically also formally satisfies the parity principle, while inheriting its errors. The boundary of cognition, then, runs through the hybrid systems that humans and machines form together.  

Sztuczna inteligencja coraz częściej staje się uczestnikiem codziennych procesów poznawczych człowieka – podpowiada informacje, selekcjonuje treści i wspiera podejmowanie decyzji. Debata naukowa koncentruje się jednak głównie na możliwościach stworzenia sztucznej inteligencji ogólnej (AGI) czy emulacji ludzkiego mózgu (Whole Brain Emulation), podczas gdy już obecny etap rozwoju AI rodzi pytania o granice ludzkiej autonomii poznawczej.

Celem referatu jest analiza wpływu współczesnych systemów AI na procesy poznawcze człowieka z perspektywy filozofii umysłu, kognitywistyki oraz nauk o komunikacji. Punktem wyjścia jest założenie, że modele generatywne i systemy rekomendacyjne nie muszą odtwarzać ludzkiego sposobu myślenia, aby realnie modyfikować sposób przetwarzania informacji i podejmowania decyzji.

Referat odwołuje się do koncepcji umysłu rozszerzonego (extended mind), odciążenia poznawczego (cognitive offloading) oraz ograniczonej racjonalności (bounded rationality), ukazując AI jako element środowiska poznawczego człowieka. W konkluzji postawiona zostanie teza, że granice sztucznej inteligencji wyznacza nie tylko możliwość stworzenia systemów dorównujących człowiekowi, lecz także ich rosnąca zdolność do współkształtowania ludzkiego poznania, autonomii decyzyjnej i sposobu konstruowania wiedzy.  

Opresja społeczna to niesprawiedliwe traktowanie jednej grupy społecznej przez inną grupę. Zdaniem N. Fraser niesprawiedliwość w relacjach międzygrupowych może przybierać dwie formy: niesłuszną dystrybucję dóbr lub brak należnego uznania. Według I. M. Young należy wyróżnić pięć form opresji społecznej: wyzysk, marginalizację, bezradność, imperializm kulturowy i przemoc. A. E. Cudd mówi zaś o siłach opresji i dzieli je na przemoc, ekonomiczne siły opresji oraz psychologiczne siły opresji. Z kolei M. Fricker zwraca uwagę na zjawisko niesprawiedliwości epistemicznej przybierające formę niesprawiedliwego przypisywania wiarygodności oraz niesprawiedliwości hermeneutycznej. Jak tymczasem wiadomo poszczególne modele sztucznej inteligencji przyswajają dane tworzone przez ludzi i są używane przez ludzi, co wikła je w ich uprzedzenia, ograniczenia poznawcze i inne słabości stanowiące źródła opresyjnych relacji międzyludzkich. Co więcej, modele te są wykorzystywane w bardzo zróżnicowanych kontekstach społecznych, kulturowych i politycznych, nierzadko obciążonych poważnymi deficytami aksjologicznymi. W moim wystąpieniu chciałbym więc zastanowić się nad zagrożeniami, jakie rozwój sztucznej inteligencji może nieść dla grup narażonych na opresję społeczną. 

Projekt FlyMind dotyczy wykorzystania informacji o organizacji układu nerwowego do budowy modeli podejmowania decyzji. Powstał fizyczny model muchy w środowisku FlyGym/MuJoCo oraz wizualizacja 3D, w której zapisane decyzje agenta przedstawiane są jako wybór ofert pracy. Agent odbiera syntetyczne bodźce A i B, zapamiętuje wcześniejsze wyniki i przy kontakcie decyduje o zaakceptowaniu lub odrzuceniu oferty.

Kontroler wykorzystuje modułową architekturę inspirowaną whole-CNS Drosophila. Łączy ona przetwarzanie zapachu, pamięć nagrody, reprezentację wartości, kontekst przestrzenny, reakcję na zagrożenie oraz sterowanie ciałem. Dane BANC v888 służą jako strukturalny punkt odniesienia, atlas i źródło oznaczonych grup sieciowych. Obliczenia prowadzone są na poziomie wybranych modułów i uproszczonego podgrafu funkcjonalnego.

Model oceniany jest w kontrolowanych zadaniach uczenia wartości, zmiany preferencji oraz konfliktu pomiędzy potencjalną korzyścią a zagrożeniem. Porównania obejmują prostsze kontrolery, permutację połączeń i wyłączanie wybranych modułów.

Wstępne analizy wskazują, że decyzje zależą od zapamiętanej wartości oraz jawnych sygnałów kontekstu i zagrożenia. Wyłączenie central complex zmieniało wykorzystanie informacji o odległości, a wyłączenie modułu emergency wpływało na odrzucanie ryzykownych ofert.