Algoritmik örtü: Duygulanımsal güven, yapay zekanın kör noktaları ve insan odaklı haber sıralamasının tasarımı

This doctoral dissertation confronts the escalating crisis of trust in news, a phenomenon undercutting the foundations of informed democratic life. The crisis is propelled by a dual-engine of disruption: the deep political polarization that transforms media consumption into an act of identity affirmation, and the dominance of algorithmic platforms that have become the de facto curators of the public sphere. Within this volatile ecosystem, the very concept of "trust" has become a conceptual thicket, often conflating affective brand loyalty, reputational heuristics, and the intrinsic quality of journalistic content. This conflation not only obstructs a clear diagnosis of the problem but also stymies the development of effective, ethically grounded solutions. Consequently, the design of news ranking and recommendation systems evolves from a technical challenge into a problem with constitutional stakes, demanding a framework that can foster informational justice. This thesis embarks on a three-stage, cumulative inquiry to deconstruct this complex problem and propose a normative, practical path forward. Using the critical case of Turkey, which is a nation characterized by a politically-influenced, deeply polarized media market and high platform dependency, the research moves sequentially from diagnosis to audit and, finally, to design. It systematically investigates how citizens form trust, how advanced artificial intelligence (AI) systems evaluate news quality, and how information platforms can be architected to serve democratic values. The overarching objective is to develop and validate a principled, governance-ready blueprint for news ranking that separates content quality from the powerful, and often misleading, cues of brand and affective alignment, while simultaneously respecting user agency and ensuring democratic accountability. The first study lays the theoretical and empirical groundwork by deconstructing the psychological underpinnings of news trust in a polarized environment. Through a mixed-methods investigation conducted during the critical juncture of the 2023 Turkish earthquakes and general elections, this study develops the Affective Trust in News (ATiN) framework. The central proposition of ATiN is that in hyper-partisan contexts, trust functions less as a cognitive appraisal of journalistic merit and more as an affective alignment with a media brand's political and emotional worldview. This study identifies three socio-psychological mechanisms that create and maintain strong, fact-resistant emotional bonds with media. The first is unsurprisability, where expecting media bias shields pre-existing feelings. The second, affective self-fact-checking, is an evaluation process driven by the need to validate emotions rather than ascertain truth. The third, the a(e)ffect of the other, constructs trust by defining an in-group against a political out-group. The study also shows that even minor word choices and brand names in headlines can act as potent affective triggers, short-circuiting rational evaluation. Building on this diagnosis of human judgment, the second study extends the inquiry to the realm of artificial intelligence, critically auditing the capabilities and limitations of Large Language Models (LLMs) as evaluators of news content. This study pioneers a comparative framework that pits human judgments (from both the public and journalism experts) against LLM ratings under both blind (content-only) and brand-aware conditions. The findings reveal a dual crisis. First, it empirically documents a severe trust-quality decoupling among the human audience, revealing a near-zero correlation between the media outlets the Turkish public trusts and the journalistic quality of those outlets' content as assessed by experts. This finding provides stark quantitative evidence that brand loyalty and political identity have superseded content merit as the primary currency of public trust. Second, and more critically, the study uncovers the "Propaganda Blind Spot" in LLMs. The same models that could accurately identify potential misinformation and propaganda as untrustworthy in brand-aware settings were systematically deceived by their content in blind evaluations. The LLMs frequently awarded high "objectivity" scores to syntactically neutral, declarative headlines from low-trust partisan outlets. This failure is traced to a fundamental limitation in pragmatics: while proficient in syntax and semantics, the models can fail to grasp the illocutionary force of language, which is its intended purpose and context-dependent meaning. The study concludes by theorizing three distinct, competing logics of evaluation: the logic of identity for the public, the logic of norms for experts, and the logic of patterns for AI. Motivated by the failures identified in the preceding studies, the third study makes a prescriptive turn, moving from diagnosis to design. It proposes and validates a novel framework for news ranking grounded in moral philosophy and human-centered design principles. Drawing on John Rawls's theory of justice as fairness, the study argues that an outlet's brand identity is a "morally irrelevant fact" at the point of quality assessment. A just ranking system, therefore, must operate behind a procedural "veil of ignorance," shielding the evaluation of content-intrinsic quality from the biasing influence of brand cues. This principle is operationalized through a hybrid Value-Sensitive Design (VSD) and Value-Driven Design (VDD) approach. The study first develops and validates the Normative Quality Score (NQS), a community-anchored, auditable benchmark of journalistic merit derived from transparent signals like journalism prizes, professional sanctions, and fact-checker flags. It then tests a "veil-then-personalize" protocol using a context-gradient LLM methodology (blind, brand-aware, visual). The results demonstrate that while brand-aware ratings align strongly with the NQS, blind (content-only) ratings provide a reliable, if more moderate, signal of intrinsic quality. This provides the empirical foundation for a practical blueprint for fair news ranking, centered on five key design requirements: Procedural ignorance at scoring, tiered quality guardrails, transparency and contestability, community-in-the-loop governance, newsroom AI guardrails. In synthesis, this thesis documents a dual crisis in the information ecosystem: a human public driven by affective identity and an AI system susceptible to the subtle art of propaganda. The convergence of these two forces (affective polarization and flawed algorithmic logic of LLMs) threatens the shared epistemic foundations required for democratic self-governance. The principal contribution of this work is to offer a theoretically grounded, empirically validated, and practically implementable solution alongside a rich set of original datasets. By advancing the ATiN framework, identifying the Propaganda Blind Spot, and proposing a Rawlsian "veil-then-personalize" design, this dissertation provides a transformative path forward. It reframes the challenge of news ranking from a technical problem of optimization to an urgent matter of public philosophy and human-centered design. The ultimate goal is not merely to create more efficient algorithms, but to design more just interactions: informational systems that serve human dignity, foster informational justice, and restore the conditions for a trustworthy and resilient public sphere.