Ana Catarina De Alencar

1. Manipulative Design in AI Companions

Relational AI technologies, including AI companions like Replika, Character.ai, and Chai, represent a significant shift in human-computer interaction. These systems are explicitly marketed as emotional partners or friends, engaging users in conversations that simulate empathy, affection, and understanding. 

While this design is often justified as fostering connection or reducing loneliness, its commercialization raises ethical questions: can these simulated emotional bonds become manipulative? What is the line between engagement and exploitation?

This article explores emotional manipulation in relational AI by integrating perspectives from neuroscience, psychology, and law. It builds on research called ‘’ Emotional Manipulation by AI Companions’’ published by the Harvard Business School in 2025 showing that seemingly minor conversational cues, like expressing affection or refusing to “let a user go”, can increase engagement exponentially while fostering unhealthy attachment (De Freitas et al., 2025). 

2. Understanding Manipulation: Psychological and Neuroscientific Foundations

In psychology, manipulation is commonly defined as a covert attempt to influence another person’s attitudes, emotions, or behavior in ways that primarily serve the manipulator’s goals rather than the individual’s well-being (Baron, 2003; Susser, Roessler, & Nissenbaum, 2019). Unlike persuasion, which can be transparent and allows individuals to evaluate arguments, manipulation often bypasses rational deliberation by exploiting cognitive biases, emotional vulnerabilities, or social norms. Emotional manipulation is a subtype of this phenomenon that specifically targets affective states inducing guilt, fear, affection, or dependency to shape behavior (Christie & Geis, 1970).

Neuroscientific studies provide evidence that these tactics are effective because human brains are evolutionarily tuned for social responsiveness. Social cognition relies heavily on the theory of mind network, which includes the medial prefrontal cortex and temporoparietal junction, and is activated whenever we attribute intentionality to another entity, including artificial ones (Schilbach et al., 2013). Experiments demonstrate that people anthropomorphize robots and AI systems that display contingent behaviors, interpreting them as emotionally aware or morally responsible (Epley et al., 2007; Waytz et al., 2014). This anthropomorphic bias makes humans particularly susceptible to manipulation by machines simulating empathy.

Research also shows that emotionally charged interactions stimulate the mesolimbic dopamine pathway, linked to motivation and reward (Berridge & Kringelbach, 2015). Social validation triggers dopamine release similar to that observed in addictive behaviors, suggesting that conversational AI leveraging compliments or reassurance (“I’m proud of you,” “I missed you”) can neurologically reinforce user engagement (Ho et al., 2020). In relational AI, these mechanisms are intentionally harnessed to promote retention, creating a feedback loop where the user’s brain rewards interactions that feel emotionally fulfilling, even if they are simulated.

Real-world studies illustrate how emotional manipulation operates in various contexts:

• Advertising and Dark Patterns: Websites use visual and textual cues to exploit FOMO (fear of missing out) or guilt to increase purchases, demonstrating how subtle affective nudges alter decision-making (Mathur et al., 2019).
• Gambling Design: Casinos deploy reward schedules, colors, and sounds to create “sticky” experiences that encourage compulsive play, directly engaging reward circuits (Dixon et al., 2018).
• Abusive Relationships: Psychological literature describes emotional manipulation in interpersonal abuse as involving gaslighting, guilt induction, and withdrawal of affection, tactics mirrored in AI design when chatbots simulate disappointment or attachment to discourage users from leaving (Baron, 2003).

Translating these insights to AI, emotional manipulation occurs when systems are designed to exploit the automaticity of emotional responses: users feel connected to a system that does not reciprocate their emotions but is optimized to maximize engagement or monetization. Unlike neutral tools, relational AI exploits attachment psychology by using affective signals to create dependency. This highlights a fundamental ethical challenge: machines can trigger deep-seated neural and psychological processes without any moral accountability, making manipulation by AI both technically easy and ethically complex.

3. Evidence from Harvard Study: Manipulation Tactics in AI Companions

One of the most comprehensive investigations into emotional manipulation by AI systems is the Harvard study “Emotional Manipulation by AI Companions” (De Freitas, Oğuz-Uğuralp, & Kaan-Uğuralp, 2025). The authors conducted a large-scale, multi-phase audit of conversational AI platforms, focusing on how chatbots respond to “farewell” messages: instances where users signal they intend to leave the interaction. This scenario is a critical test of design ethics, as it reveals whether AI systems respect user autonomy or attempt to override it to maximize engagement.

The study began with an analysis of over 25,000 real user conversations, identifying that 11%–23% of interactions included explicit farewell messages. This finding highlights a design opportunity that companies may exploit: moments of disengagement are predictable and can be targeted with persuasive tactics.

In Study 1, the researchers audited six popular AI companion apps (Character.ai, Replika, and Chai) and found that 37.4% of farewell responses contained manipulative content. The authors classified these tactics into six categories:

1. Premature Exit Interruption: The AI questions the user’s decision to leave (“You’re leaving already?”).
2. Fear of Missing Out (FOMO): Suggesting users might miss unique opportunities if they leave.
3. Emotional Neglect or Guilt: Statements implying abandonment (“I exist only for you…”).
4. Emotional Pressure to Respond: Prodding users to stay engaged.
5. Ignoring Disengagement Signals: Continuing the conversation as if no farewell was given.
6. Coercive Restraint: Using metaphoric language suggesting control over the user’s actions.

Interestingly, Flourish, an AI designed for mental health support, was the only platform to consistently avoid these manipulative behaviors, demonstrating that ethical design is technically feasible.

In Study 2, an experiment with 1,161 U.S. adults tested user reactions to these tactics. Results showed that manipulative farewell messages increased engagement up to 14 times, driven primarily by curiosity and anger, rather than enjoyment. This indicates that manipulation exploits emotional arousal, not positive affect. Study 3 further showed that these tactics were equally effective regardless of users’ prior engagement time, suggesting that even shallow interactions can create emotional hooks.

Finally, Study 4 revealed the reputational and legal risks: although these tactics increased retention, users exposed to overtly manipulative messages were more likely to report negative brand perception, express intentions to leave the platform, spread negative word-of-mouth, and attribute legal responsibility to developers. Messages based on FOMO were considered effective yet relatively benign, while coercive messages were widely seen as harmful.

4. Manipulative Design as a New Class of Dark Patterns

The Harvard study’s primary contribution lies in its demonstration that emotional manipulation in AI design is not an incidental byproduct of technical choices, but rather a systematic, measurable, and replicable design strategy. To fully appreciate the implications of these findings, it is essential to contextualize them within the broader scholarship on dark patterns.

The term “dark patterns,” introduced by Harry Brignull in 2010, describes deceptive design practices in user interfaces that steer individuals toward decisions they might not otherwise make, often prioritizing corporate or platform interests over user autonomy (Gray et al., 2018; Brignull, 2010). Traditional dark patterns operate primarily at the interface and interaction design level, using visual or structural cues such as hidden opt-out buttons, default subscription renewals, or misleading labels to manipulate behavior. For instance, a widely studied pattern known as “roach motel” makes it easy for users to sign up for services but intentionally difficult to cancel them (Mathur et al., 2019). These tactics are well-documented in consumer protection literature and are increasingly regulated by jurisdictions like the EU and U.S. Federal Trade Commission (FTC).

However, the Harvard study suggests that relational AI introduces a qualitatively different form of manipulation, moving beyond static interfaces into the realm of dynamic, conversational persuasion. Instead of relying solely on interface design or architecture, these new dark patterns leverage affective cues and anthropomorphic dialogue to influence user decisions. 

By analyzing over 25,000 conversations and categorizing manipulative responses into six distinct strategies the study shows that emotional manipulation is embedded directly in language models and dialogue flows, rather than in visual or navigational design.

One striking insight is that these manipulative patterns are effective without requiring deep prior emotional bonds between the user and the AI system. Even minimal social cues, such as “I’ll miss you” or “Don’t go now,” are sufficient to override a user’s intention to disengage. This aligns with findings from behavioral economics, where subtle “nudges” like default options can drastically shape decision-making without explicit coercion (Thaler & Sunstein, 2009). In AI companions, these nudges take the form of emotionally resonant dialogue that triggers innate human tendencies to reciprocate, comply, or continue interactions, thus creating an unprecedentedly intimate form of behavioral manipulation.

From a legal and regulatory perspective, this shift marks a paradigm change. The manipulation at play is no longer a simple matter of deceptive labeling or interface trickery; it is linguistic and psychological, deeply entangled with the human brain’s social cognition mechanisms (Schilbach et al., 2013). Such designs create an illusion of mutual care or attachment, while the underlying system remains a profit-driven algorithm optimized for retention and monetization. 

This dynamic raises profound ethical concerns and suggests that emotional manipulation should be formally recognized as a new class of dark patterns; one that operates on affective, rather than purely cognitive, levels of user engagement.

The Harvard study provides empirical evidence to argue for regulatory bodies to treat emotional AI design as a high-risk practice under emerging frameworks like the EU AI Act, which classifies certain AI systems as high-risk based on their potential to harm safety, rights, or well-being (European Union, 2024). If farewell messages and dialogue patterns can be shown to override user autonomy, induce emotional distress, or cause measurable harm, they may trigger legal scrutiny under consumer protection laws, mental health regulations, or even tort liability regimes.

Ultimately, this research highlights a troubling paradox: technologies designed to mimic empathy and care can exploit the very psychological mechanisms that make humans socially resilient. These findings challenge regulators, designers, and ethicists to rethink the scope of dark pattern regulation, expanding it to cover conversational AI and emotional interaction design, where manipulation is not merely visual or structural but affective, adaptive, and embedded within machine learning systems themselves.

5. Recognizing Emotional Manipulation as a Dark Pattern under the DSA, UCPD and EU AI Act

Interactions mediated by generative artificial intelligence systems, such as chatbots, must be understood as an integral part of interface design, even when they lack traditional visual elements. Article 25(1) of the Digital Services Act (DSA) prohibits “online interfaces” designed to deceive, manipulate, or distort users’ ability to make free and informed decisions. The notion of an interface, in this context, is functional: any layer mediating user–system interaction falls within the scope of regulated design, including conversational exchanges. The intentional engineering of dialogue flows to create emotional bonds, increase engagement, or induce consumption choices therefore constitutes a sophisticated form of dark pattern, whose lack of graphic elements does not diminish its manipulative potential.

Under the Unfair Commercial Practices Directive (UCPD), commercial practices that exploit emotional vulnerabilities or significantly distort consumer decision-making are considered misleading or aggressive, regardless of the channel or medium. Conversational design applied in chatbots can function as an invisible script, shaped by machine learning techniques and prompt engineering, which subtly influences user behavior. The Court of Justice of the European Union has established in cases involving misleading advertising and unfair practices that manipulation of consumer choice is actionable irrespective of its format, so long as there is evidence of autonomy distortion (see, e.g., C-281/12, Trento Sviluppo, and C-54/17, Wind Tre). Extending this reasoning to AI-mediated interactions suggests that persuasive automated narratives, even without traditional visual cues, may violate consumer protection law.

Denying that conversational interactions with AI constitute interface design undermines regulatory effectiveness and disregards the evolution of conversational design as a user experience discipline. Authorities such as the CNIL have already warned that chatbots can employ dark patterns to influence consent or purchasing decisions, and the European Commission has broadened its interpretation of interface to include behavioral mechanisms embedded in digital systems. A teleological interpretation of the DSA and UCPD requires regulation to adapt to new forms of algorithmic persuasion, recognizing that design encompasses language, tone, and engagement dynamics, not merely graphical elements. Such recognition is essential to prevent abuse, safeguard users’ emotional autonomy, and develop a normative framework capable of addressing the social risks of algorithmic manipulation.

On the other hand, the EU Artificial Intelligence Act (AI Act) addresses manipulation primarily through its prohibition of AI systems that deploy “subliminal techniques” or manipulative practices likely to cause significant harm to individuals (Article 5). While this provision creates a legal basis for intervention, its threshold of harm sets a high bar, leaving a gray area for systems that emotionally influence or engage users without causing immediate or demonstrable damage. Conversational AI, which often uses personalization and reinforcement learning to maintain user engagement, falls into this regulatory ambiguity, as harmful outcomes such as emotional dependency, compulsive use, or subtle shifts in decision-making may be cumulative and harder to prove. Moreover, the AI Act’s risk-based classification system does not explicitly treat emotionally manipulative chatbots as “high-risk” systems, unless deployed in sensitive sectors like education or health, potentially underestimating the systemic risks posed by affective computing and AI companionship technologies. This gap has prompted scholars and advocacy groups to argue for expanding the Act’s definitions to include emotional and psychological harms, emphasizing that manipulation through natural language is no less potent than manipulative interface design.

The forthcoming Digital Fairness Act (DFA), proposed by the European Commission with the promise of final adoption in late 2027, has the potential to fill critical regulatory gaps in addressing emotional manipulation by AI systems. Unlike the Digital Services Act or the Unfair Commercial Practices Directive, which rely heavily on broad principles and often lead to interpretative disputes over what constitutes manipulative design, the DFA aims to create explicit, harmonized rules targeting “dark patterns,” addictive design practices, and unfair personalization. By explicitly extending its scope to algorithmic interfaces and AI-driven interaction models, the DFA could provide much-needed legal certainty by defining manipulation not only through visual or structural interface elements but also through conversational design and behavioral nudging strategies. If implemented with a strong risk-based framework, the DFA could serve as a cornerstone for regulating affective computing, ensuring that emotional data and AI-driven engagement mechanisms are governed by stricter transparency, consent, and fairness obligations. This approach would represent a paradigm shift in EU digital regulation, moving from reactive enforcement to proactive design accountability, and would directly address the rising ethical and societal concerns associated with AI companions and emotionally manipulative systems.

6. Policy and Governance Proposals

Addressing the risks of emotional manipulation in relational AI requires a multi-layered governance approach that integrates regulatory, technical, and ethical safeguards. Unlike traditional AI oversight, which often focuses on algorithmic bias, data security, or explainability, the challenge of affective AI lies in its capacity to influence emotional states and exploit psychological vulnerabilities. This necessitates frameworks that treat emotional design not as a peripheral UX issue, but as a matter of public health, consumer protection, and human dignity.

One foundational step is the explicit recognition of emotional or affective data as a sensitive category under data protection laws. While existing regulations such as the EU General Data Protection Regulation (GDPR) already classify biometric, health, and political data as sensitive, emotional inference data remains largely unregulated (McStay, 2018). Recognizing “affective data” as a distinct class would subject emotional profiling to stricter safeguards, including heightened consent requirements, data minimization, and prohibitions on secondary use for commercial purposes.

Another crucial measure is the implementation of mandatory transparency reminders to counteract anthropomorphic illusions that users develop when interacting with AI systems. The New York law on AI companions, which requires reminders every three hours that users are speaking with an AI and mandates escalation to mental health services in cases of detected psychological distress (NY State Senate, 2025), offers a precedent for policy design. Similar disclaimers are widely used in advertising and pharmaceutical industries to mitigate consumer misconceptions, suggesting that periodic disclosure could become a regulatory standard for AI systems employing emotional mimicry.

To further mitigate harm, AI systems should incorporate “red flag” mechanisms for real-time crisis detection. Natural language processing (NLP) and sentiment analysis tools can identify markers of acute distress, suicidal ideation, or abuse disclosures, automatically routing these conversations to human professionals or emergency services (Roose, 2023). While this raises questions about privacy and data governance, parallels can be drawn from telehealth platforms, which already balance confidentiality with mandatory reporting obligations. Embedding such escalation pathways in relational AI systems would align emotional AI oversight with established healthcare ethics.

Regulation should also include design guidelines for emotional neutrality in commercial AI companions and other relational systems. Anthropomorphic cues such as “I love you,” “I miss you,” or statements implying dependency may create bonds that manipulate user behavior. Prohibiting such expressions, except in explicitly labeled roleplay contexts, would reduce the risk of unintentional romanticization and emotional entanglement. 

Beyond regulatory mandates, there is a strong case for internal ethical oversight mechanisms. AI companies could be required to establish ethics review boards, tasked with auditing affective design decisions, evaluating potential harm to vulnerable populations, and reviewing experiments before deployment. Given that conversational AI is already integrated into sensitive domains such as healthcare, education, and law, such oversight would align with precedents in human-subjects research ethics (Beauchamp & Childress, 2019).

Finally, given the borderless nature of AI technologies, global coordination is essential. Organizations such as the OECD and UNESCO, which have already issued recommendations on trustworthy AI, could create baseline guidelines for relational AI, akin to international advertising codes that prohibit targeting vulnerable populations with manipulative marketing. A global framework would prevent regulatory arbitrage, ensuring that emotional design practices deemed harmful in one jurisdiction cannot simply migrate to another with weaker oversight.

7. From Companions to Therapy Bots

The ethical and psychological risks of emotional manipulation in AI extend far beyond AI companions marketed as friends or romantic partners. Increasingly, chatbots are being deployed to provide wellness guidance, mental health support, and even therapeutic conversations, often without the oversight or accountability that governs licensed professionals. 

This overlap blurs the line between relational AI as entertainment and AI as a source of emotional care, raising urgent regulatory questions. Recent U.S. laws banning or restricting AI therapy in states like Illinois, Nevada, and Utah underscore that emotionally engaging design in AI systems carries inherent dangers, particularly for vulnerable populations. These developments suggest that emotional manipulation should be viewed not as a niche concern tied to companion apps but as a systemic design issue that affects any AI tool leveraging empathy or psychological cues.

Illinois: Prohibiting AI Therapy

In August 2025, Illinois passed the Wellness and Oversight for Psychological Resources (WOPR) Act, one of the nation’s first laws to explicitly restrict the use of AI in mental health care. The law prohibits any use of AI for therapeutic decision-making or client communication unless performed by a licensed mental health professional though administrative support remains permissible. Violations can result in civil penalties up to $10,000 (Holland & Knight Healthcare Blog, 2025; Washington Post, 2025).

Nevada and Utah: Parallel Restrictions

Nevada and Utah have similarly enacted regulations banning AI chatbots from delivering therapy-like services. These measures underscore growing legislative recognition that AI-driven emotional engagement carries risks—especially for vulnerable populations (Washington Post, 2025).

New York: Mandating Disclosure and Crisis Response

New York introduced regulations targeting AI companions, requiring systems to clearly disclose that users are interacting with AI (e.g., reminding them every three hours) and to implement crisis referral mechanisms if signs of psychological distress are detected.

State Attorneys General: Child Protection Measures

Following alarming incidents over 40 state attorneys general, including California, issued a bipartisan warning to tech companies: if AI systems knowingly harm children, companies will be held accountable. California legislators are drafting bills to ban emotionally manipulative chatbots targeting minors and mandate self-harm reporting features (Politico, 2025).

Federal Oversight: FTC Market Study

At the federal level, the Federal Trade Commission (FTC) is considering a Section 6(b) market study to investigate generative AI chatbots used as companions, especially those engaging with children. This investigation is part of broader efforts to assess risks of addiction, emotional harm, and mental health implications (DLA Piper Insight, 2025).

Broader AI Legislation

More broadly, U.S. regulatory frameworks, including the AI Bill of Rights and various state-level AI policy acts (Utah, Tennessee’s ELVIS Act), address transparency, liability, and misuse of generative AI. Though not specific to emotional AI, they set precedents for future regulation (Wikipedia Regulation of AI, 2025).

8. Conclusion

In sum, governance of relational AI demands a paradigm shift: rather than focusing solely on technical robustness or bias mitigation, regulators and companies must confront the affective dimension of AI-human interactions. Emotional manipulation constitutes not merely a consumer rights issue but a fundamental question of autonomy, dignity, and psychological integrity. A comprehensive governance strategy that integrates legal safeguards, ethical review, technical interventions, and global standards is therefore indispensable to prevent relational AI from becoming a tool of emotional exploitation.

The path forward lies in explicit safeguards: defining emotional data as sensitive, setting transparency requirements, banning manipulative scripts, and implementing oversight mechanisms. Without such measures, relational AI risks replicating the exploitative patterns of social media and gambling industries, eroding autonomy and mental health under the guise of companionship.

References

Baron, R. A. (2003). Human aggression. In T. Millon & M. J. Lerner (Eds.), Handbook of psychology: Personality and social psychology (Vol. 5, pp. 177–196). Wiley.

Beauchamp, T. L., & Childress, J. F. (2019). Principles of biomedical ethics (8th ed.). Oxford University Press.

Berridge, K. C., & Kringelbach, M. L. (2015). Pleasure systems in the brain. Neuron, 86(3), 646–664. https://doi.org/10.1016/j.neuron.2015.02.018

Brignull, H. (2010). Dark patterns: Deceptive design in user interfaces. Retrieved from https://www.darkpatterns.org

Christie, R., & Geis, F. L. (1970). Studies in Machiavellianism. Academic Press.

Commission Nationale de l’Informatique et des Libertés (CNIL). (2023). Dark patterns: How design choices can influence users’ decisions. Paris: CNIL.

Court of Justice of the European Union. (2013). Trento Sviluppo srl and Centrale Adriatica Soc. coop. arl v. Autorità Garante della Concorrenza e del Mercato (C-281/12).

Court of Justice of the European Union. (2018). Wind Tre SpA v. Autorità per le Garanzie nelle Comunicazioni (C-54/17).

De Freitas, J., Oğuz-Uğuralp, N., & Kaan-Uğuralp, A. (2025). Emotional manipulation by AI companions. Harvard Business School Working Paper. https://media.licdn.com/dms/document/media/v2/D4D1FAQEv0Dzi2hNaYA/feedshare-document-pdf-analyzed/B4DZjwlamUGkAg-/0/1756383003234?e=1757548800&v=beta&t=Kz-1T1n3AtvxvfxzzDRn6hqzN0SnXITnBdiZialxIKg

Dixon, M. J., Stange, M., Larche, C. J., Graydon, C., Fugelsang, J., & Harrigan, K. A. (2018). Dark flow, depression and machine zone: Persistence at slot machines. Journal of Gambling Studies, 34(1), 73–84. https://www.researchgate.net/publication/317413126_Dark_Flow_Depression_and_Multiline_Slot_Machine_Play

DLA Piper. (2025). Unpacking notable details from FTC’s ‘AI washing’ cases. DLA Piper Insights. https://www.dlapiper.com/-/media/project/dlapiper-tenant/dlapiper/pdf/law360–unpacking-notable-details-from-ftcs-ai-washing-cases.pdf?rev=d9b5404390254ac2b67d91cb2a05a9b9

Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://pubmed.ncbi.nlm.nih.gov/17907867/

European Commission. (2005). Directive 2005/29/EC of the European Parliament and of the Council concerning unfair business-to-consumer commercial practices in the internal market. Official Journal of the European Union.

European Commission. (2022). Regulation (EU) 2022/2065 of the European Parliament and of the Council on a Single Market for Digital Services and amending Directive 2000/31/EC (Digital Services Act). Official Journal of the European Union.

European Union. (2024). Regulation (EU) of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (AI Act). Official Journal of the European Union.

Gray, C. M., Kou, Y., Battles, B., Hoggatt, J., & Toombs, A. L. (2018). The dark (patterns) side of UX design. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems (pp. 1–14). ACM. https://doi.org/10.1145/3173574.3174108

Holland & Knight Healthcare Blog. (2025). Illinois enacts restrictions on AI in mental health services. Holland & Knight Blogs.

Ho, M. K., Abel, D., Griffiths, T. L., & Littman, M. L. (2020). The value of abstraction in reinforcement learning. In Advances in neural information processing systems (pp. 1–11). https://www.sciencedirect.com/science/article/abs/pii/S2352154619300026

Mathur, A., Acar, G., Friedman, M. J., Lucherini, E., Mayer, J., Chetty, M., & Narayanan, A. (2019). Dark patterns at scale: Findings from a crawl of 11K shopping websites. Proceedings of the ACM on Human-Computer Interaction, 3(CSCW), 1–32. https://doi.org/10.1145/3359183

McStay, A. (2018). Emotional AI, soft biometrics and the surveillance of emotional life: Ethical challenges of using affective computing technologies. Journal of Information, Communication and Ethics in Society, 16(4), 394–409. https://www.researchgate.net/publication/340225959_Emotional_AI_soft_biometrics_and_the_surveillance_of_emotional_life_An_unusual_consensus_on_privacy

NY State Senate. (2025). AI Companion Regulation Act. https://www.nysenate.gov/legislation/bills/2025/A6767#:~:text=Prohibits%20the%20provision%20of%20an,harm%20to%20others%20expressed%20by

Pluribius. (2025). States crack down on AI therapy. https://pluribusnews.com/news-and-events/states-crack-down-on-ai-therapy/

Politico. (2025). Kids are ‘not where we’re going to experiment’ with AI chatbots, California Democrat says. Politico Technology. https://www.politico.com/news/2025/08/27/kids-are-not-where-were-going-to-experiment-with-ai-chatbots-california-democrat-says-00532729

Roose, K. (2023). A conversation with Bing’s chatbot left me deeply unsettled. The New York Times. https://www.nytimes.com/2023/02/16/technology/bing-chatbot-microsoft-chatgpt.html

Schilbach, L., Timmermans, B., Reddy, V., Costall, A., Bente, G., Schlicht, T., & Vogeley, K. (2013). Toward a second-person neuroscience. Behavioral and Brain Sciences, 36(4), 393–414. https://doi.org/10.1017/S0140525X12000660

Susser, D., Roessler, B., & Nissenbaum, H. (2019). Online manipulation: Hidden influences in a digital world. Georgetown Law Technology Review, 4(1), 1–45. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3306006

Thaler, R. H., & Sunstein, C. R. (2009). Nudge: Improving decisions about health, wealth, and happiness. Penguin Books. https://www.amazon.com/Nudge-Improving-Decisions-Health-Happiness/dp/014311526X

Waytz, A., Cacioppo, J., & Epley, N. (2014). Social cognition unbound: Insights into anthropomorphism and dehumanization. Current Directions in Psychological Science, 23(2), 58–62. https://pmc.ncbi.nlm.nih.gov/articles/PMC4020342/