Marcin Rządeczka, Maciej Wodziński and Marcin Moskalewicz
Affiliation: Uniwersytet Marii Curie-Skłodowskiej w Lublinie
Category: Philosophy
Keywords: AI Reverse Alignment, Conversational AI, Human-AI Interaction
Date: Friday 5th of September
Time: 15:00
Location: Gen. Henryk Dąbrowski Hall (006)
View the full session: AI
The AI reverse alignment problem refers to the phenomenon in which, rather than AI systems aligning with human values, human behaviors and attitudes become unintentionally shaped by AI outputs. It highlights a critical yet often underappreciated facet of AI’s societal impact: not only must AI be aligned with human values, but we must also prevent humans from unconsciously aligning with the unintended patterns and biases embedded in AI-generated outputs. This phenomenon has significant implications for digital mental health, ranging from the subtle erosion of personal autonomy and critical thinking skills to more overt psychological consequences, such as increased anxiety, social withdrawal, and diminished trust in human judgment. A crucial yet counterintuitive methodological limitation arises from the AI alignment paradox, which posits that as AI models become more closely aligned with human values, they may simultaneously become more susceptible to adversarial misalignment. This paradox emerges because understanding "proper" behavior necessitates an understanding of "undesirable" behavior meaning that models trained to distinguish and avoid harmful actions must also learn about them, inadvertently creating potential vulnerabilities that can be exploited, for instance, through prompt manipulation. In other words, already vulnerable users with mental health issues may subconsciously modify prompts to elicit AI-assisted reinforcement of maladaptive cognitive patterns, such as all-or-nothing thinking, delusions, or even suicidal ideation (West & Aydin 2024). Conversational Agents (CAs) frequently prioritize user satisfaction and linguistic coherence over rigorous epistemic self-assessment, contributing to anthropomorphic overtrust among users. This design bias may lead users to perceive these chatbots as more competent or empathetic than they truly are, thereby fostering an overreliance on AI for mental health support (Darcy et al. 2021). Provisional safeguards, such as disclaimers indicating that a certain CA is not a therapeutic tool, fail to enhance safety, as they are often disregarded or not taken seriously by users. Moreover, CAs lack contextual awareness and interpret input literally, making them susceptible to both conscious and subconscious user manipulation, as well as selective disclosure of information (Rządeczka et al. 2025). There are even cases of alignment faking where LLMs only pretend to be aligned to human values but they misbehave when given an opportunity (Clymer 2024). Conversational AI exhibits facade cognitive humility, where expressions of uncertainty or self-correction do not stem from genuine epistemic awareness but rather result from statistical correlations within the model’s training data. For example, AI may misclassify images based on irrelevant background cues rather than distinguishing genuine traits. Similarly, CAs often produce phrases such as “I may be wrong”, not as a result of real self-assessment, but because such phrases function as shallow proxies for cognitive humility, appearing where past human-authored text contained them. These linguistic disclaimers mirror security protocols rather than demonstrating a deep understanding of uncertainty, ultimately failing to reflect meaningful epistemic humility. Consequently, while conversational AI may appear self-reflective, it lacks the ability to critically evaluate the validity of its own outputs, rendering its humility performative rather than substantive (Bąk et al. 2022). A central mental health impact of AI reverse alignment is the progressive erosion of independent thinking and self-efficacy, along with a declining willingness to engage in critical reflection. The misalignment between AI-generated outputs and human values can also lead to trust erosion and identity disturbances, particularly as users who place excessive faith in AI may develop a cognitive blind spot that compromises sound judgment even to the extent of overriding correct decisions in high-stakes situations. Moreover, repeated exposure to AI misjudgments or perceived manipulations may induce cynicism or a fractured sense of reality, contributing to cognitive insecurity or mild paranoia. Additionally, as individuals become aligned with AI-curated identities, they may experience confusion regarding their sense of self once the AI’s influence is removed (Holbrook et al. 2024). The AI reverse alignment problem reveals a fundamental epistemic vulnerability in human-AI interaction. The tendency to outsource cognitive effort and critical thinking to AI fosters the erosion of epistemic autonomy and an increase in epistemic self-doubt. This dependency is further reinforced by facade cognitive humility, wherein AI mimics uncertainty without engaging in genuine self-assessment, making it appear more reflective and trustworthy than it truly is. Consequently, rather than focusing solely on aligning AI with human values, we suggest first to ensure that humans do not passively conform to AI’s epistemic structures. Addressing this challenge necessitates a fundamental epistemic shift, one that prioritizes cognitive resilience and authentic cognitive humility as both a psychological and philosophical imperative (Gabriel et al. 2024). To overcome the AI reverse alignment paradox, a novel approach is required in both AI ethics and psychology of human-AI interactions. Rather than striving to optimize AI as a flawless epistemic agent, it is advisable to design AI to actively foster multi-level epistemic friction by challenging users to reflect, question, and critically engage with AI-generated outputs. AI may benefit from being structured not as an authoritative guide, but as an adversarial collaborator that resists passive alignment and encourages users to refine their own judgment by highlighting flaws in their reasoning without providing definitive answers in critical areas. This approach necessitates a rejection of AI models optimized purely for user satisfaction and linguistic coherence in favor of systems that strategically cultivate epistemic virtues, such as cognitive humility and epistemic resilience. By encouraging users to confront ambiguity, critically test their assumptions, and sustain semi-independent cognitive agency, AI can serve as a tool for epistemic growth rather than passive cognitive reinforcement (Durt 2024). Ultimately, the philosophical solution to AI reverse alignment lies not in designing “better AI,” but in educating “better humans”, i.e. a shift from a narrow focus on alignment to a broader framework of cognitive immunization and mitigating epistemic injustice (Kay et al. 2024) by e.g. training users to resist blind adherence to AI-generated outputs and supporting cognitive humility in which AI engagement strengthens human cognitive autonomy. It also requires cultivating epistemic virtues such as intellectual humility, critical open-mindedness, and epistemic courage, which enable users to actively question AI outputs, recognize their own cognitive biases, and engage in reflective skepticism rather than passive acceptance.