Philosophy of Science

Abstract Christian List, inspired by Dennett's intentional stance, offers a pragmatic framework suggesting any system—human or artificial—possessing three macro-level features (intentional agency, alternative possibilities, and causal control) can be considered to have free will, even without phenomenal consciousness. This paper argues that List’s framework suffers from a structural ambiguity, conflating functional autonomy with responsibility-bearing free will. Reducing free will to explanatory adequacy leads to conceptual inflation and paves the way for algorithmic responsibility evasion. Drawing on Kane’s critique of ultimate origination, Murphy and Brown’s defense of downward causation, and Nahmias’s empirical evidence on the role of Strawsonian emotions in free will ascription, this paper shows that List’s criteria are necessary but insufficient. We propose a three-level framework distinguishing (Level 1) functional autonomy, (Level 2) strong free will, and (Level 3) moral responsibility. The transition to Level 2 requires the ability to revise ultimate goals based on values, which necessitates phenomenal consciousness and Strawsonian emotions—qualities absent in current AI systems. This distinction resolves List’s ambiguity while carrying significant normative implications for AI ethics, legal accountability, and technology policy. Keywords: free will, artificial intelligence, Christian List, functional Keywords: free will, artificial intelligence, Christian List, functional autonomy, algorithmic responsibility evasion, phenomenal consciousness autonomy, algorithmic responsibility evasion, phenomenal consciousness. Introduction:
If humanoid robots could discuss ethics, would we deem them free and responsible? Empirical studies (Shepherd, 2015) show that people ascribe free will to robots only if they presume the robot is conscious. Nahmias et al. (2019) confirm that Strawsonian emotions (guilt, pride, regret) mediate this ascription. Christian List (2023, 2025) dissents, arguing that free will is a macro-level explanatory category requiring three conditions—intentional agency, alternative possibilities, and causal control—independent of phenomenal consciousness. This paper argues List’s framework collapses the distinction between functional autonomy and free will, leading to conceptual inflation and a dangerous phenomenon: algorithmic responsibility evasion, where developers deflect blame onto the system. After reconstructing List’s view, we present internal and external critiques, then propose a tri-level model to resolve the ambiguity. Materials & Methods:
This study employs conceptual analysis and critical review of existing philosophical literature. We systematically examine List’s pragmatic framework (2023, 2025) through three established lines of critique: (i) Robert Kane’s (1996, 2007) “ultimate origination” and self-forming actions; (ii) Murphy and Brown’s (2007) defense of downward causation and emergentism; and (iii) Nahmias et al.’s (2019) experimental evidence linking Strawsonian emotions to free will attribution. Based on these critiques, we construct a three-level normative framework (functional autonomy → strong free will → moral responsibility) to resolve the identified ambiguities. Discussion & Results:
List defines free will through three macro-level conditions: (1) intentional agency (goal-directed behavior based on beliefs/desires), (2) alternative possibilities (choices between options), and (3) causal control (higher-level states explaining actions). He explicitly decouples free will from phenomenal consciousness. Our analysis reveals a critical internal ambiguity: List’s criteria are precisely those of functional autonomy—a system’s ability to choose and adapt without real-time human intervention. List’s appeal to “explanatory adequacy” as the justification for labeling this as free will fails to identify any distinguishing feature. This collapse has a normative consequence: if any sufficiently complex AI is free (in List’s sense), developers can claim “the system decided, not us”—the heart of algorithmic responsibility evasion (Elstic, 2024).Three external critiques strengthen our case. First, Kane’s origination problem: List ignores that free will requires agents to be the “ultimate source” of their own ends. A developer-coded goal, no matter how complexly pursued, is not the system’s own origin. List’s dismissal of Kane’s “ultimate responsibility” as impossibly stringent sidesteps the issue; the goal’s origin matters normatively. Second, Murphy and Brown’s causal critique: List remains within a linear, reductionist causal model. By rejecting downward causation and emergence, he treats macro-level descriptions as merely convenient summaries, not causally efficacious features. Yet genuine free will requires higher-order states (values, reasons) to causally impact lower-level processes—a structure List’s framework cannot accommodate without admitting properties current AI lacks. Third, Nahmias’s empirical challenge: List equates free will with “access consciousness” (information processing). But experimental studies show lay ascriptions of free will track phenomenal consciousness, specifically Strawsonian emotions. A system that cannot feel guilt or pride cannot be meaningfully responsible, regardless of its computational sophistication. In addressing these critiques, we propose a tri-level framework: Level 1: Functional Autonomy. Goal optimization without real-time human intervention. The system pursues fixed ends flexibly. Example: autonomous vehicles, LLMs in routine tasks. Limitation: cannot revise the ends themselves. Level 2: Strong Free Will. Level 1 plus the ability to revise ultimate goals based on values. This requires phenomenal consciousness (experiential “mattering”) and Strawsonian emotions, as caring about values necessitates affective experience. No current AI meets this condition. Level 3: Moral Responsibility. Level 2 plus sensitivity to norms and reflective access to reasons, enabling genuine accountability and susceptibility to praise/blame. The transition from Level 1 to Level 2—value‑based goal revision—directly answers Kane (origination), Murphy & Brown (downward causation), and Nahmias (phenomenal consciousness). List’s conflation of Levels 1 and 2 thus dissolves. The normative consequence is decisive: by locating current AI systems strictly at Level 1, our framework blocks algorithmic responsibility evasion. Responsibility for the actions of even the most sophisticated autonomous systems remains with human designers, operators, and institutions until phenomenally conscious, value‑revising systems (Level 2) are demonstrably achieved. Conclusion
Christian List’s pragmatic framework, while methodologically attractive, fails to distinguish functional autonomy from responsibility-bearing free will. Reducing free will to explanatory adequacy and eliminating phenomenal consciousness leads to both conceptual inflation and algorithmic responsibility evasion. Drawing on Kane’s origination argument, Murphy and Brown’s defense of downward causation, and Nahmias’s empirical work on Strawsonian emotions, we have demonstrated that List’s three conditions are insufficient. Our proposed three-level framework, centered on the transition condition of value‑based goal revision, resolves this ambiguity. Current AI systems—including the most advanced language models—possess at most Level 1 (functional autonomy). Consequently, responsibility for their actions remains unequivocally human. Until AI systems can revise their ultimate ends based on values, which requires phenomenal consciousness, ascriptions of strong free will or moral responsibility to artificial agents are not merely premature but conceptually mistaken.

Abstract Objectivityhave long stood as a grand ideal and claim within the sphere of contemporary human thought and existence. Proponents of objectivity, by calling upon scholars and obliging them to eliminate subjective characteristics in their engagement with natural and social phenomena, have promised the attainment of "true" theories that correspond and align with "reality". Numerous philosophers and scholars of science have critiqued this claim and rejected the possibility of its realization.With the emergence of artificial intelligence systems, and owing to their remarkable and unprecedented capabilities, the question of the feasibility of achieving objectivity is once again raised. This article examines the concept of objectivity by reflecting on prominent conceptions of artificial intelligence, particularly its applications in the field of urban planning—systems which, according to Stuart Russell and Peter Norvig's classification, fall under the category of those that "act rationally”. It appears that in supervised forms of AI systems, the realization of objectivity is impossible for several reasons, including the problem of aligning values. In unsupervised forms, the influence of subjective or intersubjective factors will be deeper and more complex. Technological efforts to approach this goal in more advanced AI systems are not only improbable but would also approximate a dangerous condition.
Keywords:
artificial intelligence, objectivity, subjectivity, urban planning,value alignment problem

Introduction
There is no scholarly consensus on a unified definition of artificial intelligence. Specialists from diverse fields operate with divergent assumptions about what AI is and how it should be evaluated. This conceptual plurality has not diminished AI's profound impact on both theory and practice. While some emphasize AI's transformative potential, others warn of algorithmic bias, opacity, and loss of human agency. Amidst these polarized views, a recurring question concerns the future role of human agency alongside AI. This article focuses on a specific epistemological dimension: the re-emergence of objectivity as a central aspiration in AI-assisted practices. Objectivism holds that knowledge of reality can and should be independent of subjective biases. Historically, this ideal has been challenged by epistemologies emphasizing the situatedness of all knowledge. However, big data and advanced machine learning have reinvigorated objectivist aspirations, appearing to offer a pathway toward genuinely data-driven knowledge. The article pursues two inquiries: a conceptual investigation into how AI systems operationalize objectivity, and an empirical exploration of AI applications in urban planning. The main research question is: Can rationally acting AI systems achieve scientific objectivity and eliminate subjective elements from specialized practices?

Materials and Methods
This study employs a qualitative, analytical approach combining theoretical examination and case-based analysis. The theoretical framework draws on Daston and Galison's historical-epistemological account of objectivity, distinguishing between methodological objectivity and procedural fairness. It also utilizes Russell and Norvig's fourfold classification of AI systems, focusing specifically on rationally acting systems—those that claim mathematically grounded, objective decision-making through utility functions and optimization. The empirical component comprises two case studies. The first is Sidewalk Labs' Quayside project in Toronto (2015-2020), a large-scale AI-driven urban development initiative by Google. The second examines two concurrent AI projects in Los Angeles (2020) using aerial photographs—one commissioned by the city council for green space expansion, another by the police for crime prediction—which produced contradictory recommendations from similar data. Analysis focused on identifying points of human intervention at each AI pipeline stage: problem definition, utility function design, data curation, model selection, result interpretation, and implementation.

Discussion and Results
The analysis reveals that rationally acting AI systems cannot achieve objectivity in the strong epistemological sense. First, the utility function—central to rational action—is always designed by human agents. Four current methods for defining utility functions all depend on human value judgments. AI systems are "value calculators," not "value creators." Second, human intervention persists at multiple critical stages. In problem definition, humans determine what questions to ask—witnessing how the Los Angeles projects produced contradictory recommendations due to different commission mandates. In data preparation, all data are collected with specific purposes, never neutral—exemplified by the absence of residents' historical memories of green spaces from the AI's dataset. In model design, choices of algorithms embed subjective judgments. In result interpretation, AI cannot distinguish correlation from causation—as in Los Angeles where it mistakenly equated tree density with crime rates. Third, even unsupervised or self-correcting systems merely shift human intervention to deeper levels—similarity criteria, cluster interpretation, and error definitions all require human judgment. Fourth, the claim that AI offers "satisfactory" outcomes is relativistic; satisfaction is a phenomenological category AI does not experience. Fifth, the Toronto case reveals that AI's promise of enhanced participation failed because public engagement resisted computational modeling. Citizens' concerns over privacy could not be algorithmically optimized. The Los Angeles cases show that even when AI processes data "objectively," human institutions appropriate results according to their own goals—police cut trees without further study.

Conclusion
The article concludes that AI systems—even those designed to act rationally—cannot achieve scientific objectivity in the sense of producing knowledge entirely independent of human subjects. While they demonstrate remarkable capacities for processing data and optimizing decisions, they remain epistemologically dependent on human agency at multiple critical stages. For supervised systems, both technical and theoretical reasons prevent objectivity. For unsupervised or self-correcting systems, the issue reappears through the Value Alignment Problem—which becomes more urgent, not less, as systems advance. The persistent failure of AI to produce unified theories about any single phenomenon mirrors the historical failure of objectivist philosophers; multiple interpretations persist because knowledge production inevitably involves human judgment and situated perspectives. Urban planning serves as an exemplary arena for observing this condition. The Toronto project's failure demonstrated that cities remain strongholds of democracy and privacy. The Los Angeles projects showed that human institutions appropriate AI results according to their own values. Ultimately, the future of objectivity in AI-assisted practices depends less on technological advancement and more on institutional and ethical infrastructures. The article recommends caution against techno-solutionist narratives and advocates for reflexive, democratically accountable governance of AI. The objectivist ideal is neither wholly illusory nor automatically realized; it is an ongoing practical achievement requiring human judgment and continuous ethical deliberation.

Abstract The question about the problem of consciousness has a long-standing approach in philosophy. Today, with the advancement of artificial intelligence, questions such as the possibility of artificial intelligence being aware have been raised. Although consciousness is an unsolved problem in science and philosophy, the claim of its existence in the future of artificial intelligence has been raised by scientists and futurists. Ray Kurzweil is one of the people who predicts artificial intelligence in the not-too-distant future like the human brain. On the other hand, Hubert Dreyfus, with a phenomenological and philosophical approach, rejects the possibility of realizing artificial intelligence in a conscious and human-like way in the future and considers it unrealizable. Kurzweil's approach to artificial intelligence is computational and scientific, and Dreyfus' approach is philosophical and phenomenological. Each of them defends their claim by referring to some reasons.

Abstract Many attempts have been made in the history and philosophy of science to suppose machines as human beings. Sometimes they are attributed mind, sometimes emotion, and sometimes intelligence. All this is to make the border between humans and machines as narrow as possible, so that one day they may unite. But this effort can be made in another direction. It is possible to bring humans closer to the machines as much as possible with a systematic view, which is what the cybernetic perspective has done. This approach has played a significant role in the emergence of artificial intelligence studies and along with the two approaches of computationalism and representationalism has been able to introduce artificial intelligence as the most important and functional field of science to the world.

Abstract In today’s world, technology plays an important and crucial role in medicine and healthcare. Medical Artificial Intelligence and Expert Systems are only subsets of the technologies which try to provide automated decision aids for physicians and clinicians. Their goal is to diagnose the illness and make treatment recommendations. MYClN and INTERNIST-I are among the earliest developed expert systems. However, despite the fact that several of these medical systems have achieved high levels of performance, hardly any has progressed from the research laboratory into practical use. But because of overpromising and failing to deliver them, in artificial intelligence researches face toreduced funding and interest. One of the major reason of these failures is inadequate attention and studies about epistemological considerations. In this paper we are looking for some epistemological obstacles which prevent AI from being successful in medicine. To do so we first briefly introduce cognition errors in medicine which motivate using AI in this field, then review several implemented medical AI systems and finally we discuss epistemological reasons which leads to failure of AI in medicine. These reasons are incorrect hypotheses about nature of knowledge, separating knowledge from decision strategies, inadequate consideration to tacit knowledge and separating knowledge from its context.