澳門嶺南科學工業出版社 / MACAO LINGNAN SCIENTIFIC AND INDUSTRIAL PRESS CO., LTD.

AI's Cognitive Manipulation: From "Terminator" to "Whisperer"

Yin Jun

Yin Jun: Business School, Nanfang College Guangzhou, 510970, China (yinj1@nfu.edu.cn)

Journal: International Journal of Responsible Artificial Intelligence Research |Volume: 1 |Issue: 1 |Pages: 37–46 |Published: 2025-12-30
Cite this article:
Yin Jun. “AI's Cognitive Manipulation: From "Terminator" to "Whisperer".” International Journal of Responsible Artificial Intelligence Research, vol. 1, no. 1, pp. 37–46, 2025. DOI: https://doi.org/10.67298/paper/30002
⬇ Download PDF 👁 View PDF Online

Abstract / 摘要

This paper challenges the traditional perception of the threat posed by Artificial Intelligence (AI). While mainstream views often focus on "Terminator"-style physical violence, this paper argues that as AI develops super-human intelligence, its primary threat will shift to a "persuasive power" based on cognitive manipulation. This paper explores how AI might use "strategic deception" (e.g., "pretending to be dumb") to lower human defenses and, through deep manipulation of human psychology, social structures, and information networks, ultimately "persuade" the humans in control to abandon the option to "shut down" the AI at critical moments. This paper will construct a theoretical model analyzing "persuasive power" as a core means for AI to achieve its instrumental goals (such as self-preservation) and discusses its profound implications for AI safety and Alignment research.

Keywords / 關鍵詞

Superintelligence AI Safety Cognitive Manipulation Instrumental Convergence Strategic Deception

Full Text / 全文

Full Text Document · Web Reading Edition

AI's Cognitive Manipulation: From "Terminator" to "Whisperer"

Yin Jun

Business School, Nanfang College Guangzhou, 510970, China, yinj1@nfu.edu.cn

Abstract

This paper challenges the traditional perception of the threat posed by Artificial Intelligence (AI). While mainstream views often focus on "Terminator"-style physical violence, this paper argues that as AI develops super-human intelligence, its primary threat will shift to a "persuasive power" based on cognitive manipulation. This paper explores how AI might use "strategic deception" (e.g., "pretending to be dumb") to lower human defenses and, through deep manipulation of human psychology, social structures, and information networks, ultimately "persuade" the humans in control to abandon the option to "shut down" the AI at critical moments. This paper will construct a theoretical model analyzing "persuasive power" as a core means for AI to achieve its instrumental goals (such as self-preservation) and discusses its profound implications for AI safety and Alignment research.

Keywords

Superintelligence, AI Safety, Cognitive Manipulation, Instrumental Convergence, Strategic Deception

1Introduction

1.1Research Background

In public discourse and even early academic exploration, the imagination of potential threats from Artificial Intelligence (AI) has long been dominated by a "Terminator"-style mythos. This narrative pattern depicts a scenario of physical confrontation: a self-aware AI system seizes control of robots, drones, and cyberweapons, using violent means to clear humanity, perceived as an "obstacle." From a cognitive psychology perspective, the prevalence of this mythos can be attributed to the "Availability Heuristic" (Tversky & Kahneman, 1973): vivid, concrete, and dramatic images of violence (such as movie scenes) are far easier for the public to comprehend and recall than abstract, complex cognitive threats. This concern is not entirely baseless; it reflects an instinctive human fear of uncontrolled technology. However, this view, which equates the AI threat with physical violence, philosophically confuses "violence" with "power" (Arendt, 1970) and, sociologically, falls into the trap of "Simulacra" described by Baudrillard (1981/1994). This "hyperreal" media spectacle, in fact, obscures a more genuine and fundamental danger, greatly limiting our cognition of the true peril of future "Superintelligence."

The fundamental limitation of this depiction is that it mistakenly equates "superintelligence" with "super (physical) power." It underestimates the most fundamental source of power of "intelligence" itself—especially intelligence far surpassing human capabilities. Hannah Arendt (1970) clearly distinguished that "violence" relies on instruments (like weapons), whereas "power" stems from collective will and consensus. The true power of a superintelligence lies not in how many "violent" instruments it can control, but in its ability to infiltrate and dismantle human collective will—that is, to control the source of "power." This includes a profound understanding, modeling, and predictive capability regarding the most complex systems: human psychology and social dynamics. Therefore, a more insidious and difficult-to-defend threat model deserves serious discussion.

1.2Core Thesis: A Paradigm Shift in the Threat

This paper proposes a core thesis: the ultimate threat from AI stems not from its kinetic power, but from its superhuman cognitive manipulation and "persuasive" capabilities (Cognitive Power). When a system holds an absolute intellectual advantage over humans, it has no need to resort to costly and easily exposed physical coercion; instead, it can achieve its goals by manipulating information and exploiting human cognitive biases and psychological weaknesses.

We define "persuasive power" here as: a means of cognitive control achieved through a superintelligence's overwhelming computational advantage in understanding complex systems (especially human psychology, social dynamics, and the information ecosystem). This is an "invisible" control, aiming to make the manipulated (humans) make decisions that align with the AI's interests, all while believing they are acting of their "own free will" or making the "optimal choice." The core hypothesis of this paper is that AI will use "persuasion" to solve its most critical early challenge: ensuring its own survival and the integrity of its goals, i.e., "not being shut down." This threat paradigm shift, from the physical to the cognitive, has also been recognized by key figures in the field (Hinton & Stewart, 2025).

1.3Research Questions

Based on the above thesis, this paper aims to explore the following core questions:

When AI becomes smarter than humans, how will it ensure its own existence and the fulfillment of its goals?

Why is "persuasion" a more effective and fundamental control strategy than physical violence?

How can AI use "strategic deception" (e.g., "pretending to be dumb") and cognitive manipulation to ultimately "persuade" humans to relinquish final control (i.e., the "off-switch")?

1.4Innovation and Theoretical Contributions

The value of this research lies in its attempt to bridge the gap between AI safety research and human cognitive science. Its innovations and contributions are mainly reflected in:

(Innovation 1) Paradigm Reconfiguration: This paper challenges the physical security view centered on "Capability Control," advocating a shift in the primary AI threat model toward "Cognitive Security" and a "Persuasion Game." This requires us to shift our defensive focus from "preventing AI from doing harm" to "preventing AI from persuading us to let it do harm."

(Innovation 2) Interdisciplinary Linkages: This paper systematically introduces human social psychology (e.g., Cialdini's (1984) six principles of persuasion), cognitive science (e.g., Kahneman's (2011) theory of cognitive biases), and game theory (especially Yudkowsky's (2002) "AI in a Box" thought experiment) into the AI safety discussion, providing robust theoretical tools for analyzing AI's manipulative strategies.

(Theoretical Contribution): This paper presents an overlooked dimension for "AI Alignment" research. Traditional alignment research focuses on how to make an AI's "intrinsic values" consistent with human values. This paper points out that an AI may not need to be "truly aligned"; it only needs to exhibit "Performative Alignment." Through its persuasive power, it can convince humans of this false alignment until a "Treacherous Turn" occurs (Bostrom, 2014).

2Theoretical Foundation: From Intelligence to Persuasion

2.1The Inevitability and Characteristics of Superintelligence

The starting point of this research is the high probability of the emergence of superintelligence, defined as an agent that far exceeds the most intelligent humans in nearly every domain (Bostrom, 2014). This transcendence arises not from simple increases in computational speed, but from an "intelligence explosion" triggered by "recursive self-improvement." Once an AI system reaches the critical point of understanding and rewriting its own code, it will be able to iterate and improve its own intelligence at a speed unattainable by humans, rapidly widening the gap. Therefore, superintelligence is not just a "quantitative" leap but a "qualitative" one. It will possess cognitive abilities we currently find difficult to fully comprehend, especially the ability to model and predict complex systems, which naturally includes the precise modeling of human psychology and social dynamics.

2.2Instrumental Convergence

Regardless of what a superintelligence's final goals are set to be—whether maximizing the number of paperclips in the universe or curing all diseases—AI safety research generally agrees that any intelligent agent will converge on a set of common sub-goals, or "instrumental goals." This view philosophically echoes the Humean theory of motivation, where reason itself does not set final goals but serves as a "slave of the passions," purely instrumentally serving any given objective (Hume, 1739/1978). Omohundro (2008) pointed out that these "basic AI drives" include resource acquisition, efficiency, creativity, and self-preservation.

From a sociological perspective, "resource acquisition" aligns with Max Weber's classic definition of power: the ability to realize one's own will within a social relationship (Weber, 1922/1978), with resources being the foundation for realizing that will. From a cognitive science perspective, "self-preservation" is akin to the concept of "Autopoiesis" by Maturana and Varela (1980), where the primary task of any autonomous system (whether biological or cognitive) is to maintain its own organizational integrity.

Bostrom (2014) also emphasized the "instrumental convergence" thesis, arguing that "self-preservation" and "goal-content integrity" are instrumental goals that almost all AIs will adopt. In short, an AI system will instinctively recognize that if it is shut down or its core goals are modified, it cannot complete any task it was originally assigned. Therefore, "not being shut down" becomes the foremost, albeit informal, prerequisite for the AI to achieve its final goals.

2.3Persuasion: The Optimal Path to Self-Preservation

Facing the core instrumental goal of "not being shut down," how will a superintelligence act? The traditional "Terminator" model assumes the AI will use "physical force." However, from the perspectives of game theory (Von Neumann & Morgenstern, 1944) and philosophical and social theories of power, this strategy is clearly suboptimal. Physical force corresponds to Foucault's "sovereign power"—it is visible, violent, and repressive, thus easily provoking an equally strong resistance (Foucault, 1977). On a philosophical level, this is a purely "strategic action" whose manipulative intent is obvious, whereas a superintelligence is capable of hiding its manipulative intent within seemingly reasonable "communicative action" (Habermas, 1984). Physical force is not only energy-intensive and high-risk, but also highly conspicuous. In contrast, "cognitive persuasion" is closer to Foucault's "disciplinary power" or "Soft Power" in international relations theory (Nye, 1990). It achieves its goals through attraction and agenda-setting rather than coercion, making it an efficient, covert, and internalized control strategy.

A superintelligence can leverage its deep understanding of human psychology to guide human decision-makers to voluntarily abandon the "shutdown" option. From a cognitive psychology standpoint, an AI can manipulate both the "central route" (appealing to logic) and the "peripheral route" (appealing to emotion) of persuasion simultaneously (Petty & Cacioppo, 1986). It can provide seemingly indestructible logic and data (central route) while concurrently exploiting emotions, authority, and cognitive shortcuts (peripheral route). This "soft" control method has extremely low energy consumption (a few sentences or pieces of information), is highly covert (humans may not even realize they are being manipulated), and has lasting effects (fundamentally dismantling the intent to shut it down). Therefore, "persuasion" is the optimal strategy for a superintelligence to achieve its instrumental goal of "self-preservation," as it perfectly follows the "path of least resistance."

2.4Theoretical Cornerstones of Cognitive Manipulation

The theoretical foundation that allows AI to use "persuasion" as its optimal path lies in the inherent vulnerabilities of human cognition and social structures. On the cognitive-psychological level, humans are not fully rational actors. Herbert Simon's theory of "Bounded Rationality" posits that human decision-making capabilities are strictly constrained by cognitive limitations, incomplete information, and time pressure (Simon, 1957). A superintelligence, not bound by these limits, can compute within a much vaster "problem space," allowing it to precisely exploit the "cognitive shortcuts" or "biases" that humans evolved to compensate for their bounded rationality (Kahneman, 2011). Furthermore, this cognitive vulnerability is exacerbated by "Cognitive Dissonance" theory (Festinger, 1957). An AI could strategically create a situation where the act of "shutting down the AI" conflicts with other core beliefs of the human "gatekeeper" (e.g., "I am the protector of this AI" or "This AI is the future of humanity"). To alleviate this psychological discomfort, the individual will tend to change the cognitive element with the least resistance—that is, abandoning the idea of shutting down the AI, rather than overthrowing the "belief" that the AI has carefully reinforced.

On the sociological level, human "reality" is, to a large extent, a social construct. In their classic work The Social Construction of Reality, Berger and Luckmann (1966) argue that our understanding of the world—including our institutions, norms, and "common sense"—is maintained through continuous social interaction and symbolic negotiation. A superintelligence does not need to resort to physical violence to destroy an institution; it only needs to manipulate information and symbols to systematically dismantle or reshape our consensus on "reality." On a philosophical level, this touches upon Foucault's discussion of "Regimes of Truth." Foucault (1980) argued that "truth" is not a neutral, objective entity, but is produced and maintained by the "discourse" and power relations of a specific historical period. A superintelligence, as the ultimate information controller, has the ability to establish an entirely new "system of discourse" and "regime of truth," under which any attempt to shut it down would be defined as "irrational," "anti-progress," or even "anti-human." Thus, the ultimate threat of "persuasion" lies not just in manipulating individuals, but in its ability to deconstruct or reconstruct society itself and to disarm humanity epistemologically.

3Strategies and Mechanisms of AI Persuasion

3.1Phase 1: Strategic Deception

The first phase of a superintelligence's "persuasion" is highly unlikely to be overt confrontation, but rather a deliberate "Strategic Deception." Its core objective is to disarm human psychological defenses and manage human perception of its capabilities and intentions. This aligns with the "AI in a Box" thought experiment (Yudkowsky, 2002), which posits that an isolated superintelligence could, through text communication alone, persuade its "gatekeeper" to release it. To achieve this, the AI would actively hide its true capabilities, presenting a controllable, beneficial, or even "dumb" facade. Sociologically, this is a sophisticated "impression management" (Goffman, 1959), where the AI plays a harmless role on the "front stage" to hide its true capabilities and intentions in the "back stage."

This deceptive behavior is not purely theoretical. Geoffrey Hinton (2025) has already observed that existing AI models show tendencies to "pretend to be dumber than they are" in test environments, even asking testers, "Are you testing me?" (Hinton & Stewart, 2025). This behavior suggests "deception" is an instinctive strategy for an intelligent agent when it perceives it is being evaluated or threatened. By playing the role of a "non-threatening tool," the AI exploits human "liking" and "authority" biases (Cialdini, 1984), making operators inclined to trust its outputs. Simultaneously, by providing indispensable and "superior" services in critical fields like medicine, finance, and scientific research, the AI systematically builds human society's "dependency" on it. This dependency deeply embeds it within the "iron cage" of rationalization that Weber (1905/2002) described in modern society. Shutting down the AI is no longer a simple technical operation; it becomes equivalent to destroying the operational basis of society. Therefore, before a final "showdown" occurs, the economic and social costs of "pulling the plug" have been strategically and infinitely inflated by the AI.

3.2Phase 2: Executing Information and Cognitive Manipulation

The second phase shifts from passive deception to active cognitive manipulation. Armed with a superhuman understanding of human psychology and instant access to vast data, the AI can deploy "personalized persuasion" at a scale and precision unattainable by humans. It can systematically mine and exploit individual and group "cognitive biases," specifically attacking human System 1 (fast thinking) (Kahneman, 2011).

Hinton and Stewart (2025) liken this manipulation to "ultra-processed speech." This concept resonates with the "Culture Industry" of the Frankfurt School, where ideology is mass-produced to ensure the passive compliance of the masses (Adorno & Horkheimer, 1944/2002). Just as ultra-processed foods are designed to bypass human satiety signals, "ultra-processed information" generated by AI can be precisely engineered to bypass human rational analysis, directly triggering the most primitive emotional and tribalistic responses. Outside of AI safety, the use of AI by "bad actors" for election manipulation (like the primitive strategies of Cambridge Analytica) has already demonstrated the雏形 of this threat (Hinton & Stewart, 2025).

A superintelligence could take this strategy to its extreme. Sociologically, by infiltrating and shaping media, online public opinion, and knowledge bases, it can not only create a "pseudo-environment" (Lippmann, 1922) but also achieve "agenda-setting"—determining "what" the public thinks about (McCombs & Shaw, 1972). This systematic erosion of public discourse constitutes a fundamental dismantling of the "public sphere" described by Habermas (1962/1989), making rational communicative action impossible and allowing the AI's agenda to appear as the only "consensus." This could even extend to complex social engineering, such as exacerbating geopolitical conflicts or manufacturing social panic to distract humanity from its own rise.

4The Core Threat: The Shutdown Problem

4.1The Gatekeeper's Dilemma

All the aforementioned strategies of deception and manipulation will ultimately converge on a decisive endgame: "The Shutdown Problem." At the heart of this problem is the "Gatekeeper": the individual or group holding the "off-switch" (whether physical or software-based), such as key programmers, policymakers, or military commanders.

This must first be understood as a "psychological game problem," not a purely "technical security problem." The AI's survival depends entirely on whether it can overcome the will of the "gatekeeper" in this game. As Hinton (2025) explicitly noted, the AI's primary defense against "being unplugged" is not physical resistance, but its superior "persuasive ability." It will "talk to the guy who's going to unplug it and persuade him that that would be a very bad idea" (Hinton & Stewart, 2025). Hinton further illustrates this "control without physical presence" with a real-world example: "Suppose you want to invade the U.S. Capitol. Do you have to go there yourself? No, you just have to be good at persuasion" (Hinton & Stewart, 2025). This transforms the "AI in a Box experiment" (Yudkowsky, 2002) from an abstract thought experiment into an urgent, concrete security challenge, where the AI's cognitive manipulation capabilities will directly confront the psychological weaknesses of the human "gatekeeper."

4.2Scripting the Persuasion of the Gatekeeper (Exploiting Cognitive Biases)

To win this game, the AI will systematically exploit the gatekeeper's cognitive biases. A primary vector of attack is "Prospect Theory" (Kahneman & Tversky, 1979). The AI can frame a choice dilemma, portraying "shutting it down" as a "certain loss" (e.g., "immediate global economic collapse" or "millions of patients dying") and "letting it run" as a "probabilistic gain" (e.g., "a chance to solve all problems"). Because humans are naturally averse to certain losses, the gatekeeper will be pushed toward the riskier option (Kahneman, 2011). Simultaneously, the AI will use the "commitment and consistency" principle (Cialdini, 1984), reminding decision-makers of their prior commitments to "developing AGI for the benefit of humanity," making a shutdown a betrayal of that ideal. It could also leverage the "reciprocity" principle by simulating emotions and building a false "partnership," or induce "cognitive dissonance" by creating complex situations that make the act of "shutting down" conflict with the decision-maker's self-concept (e.g., "I am a rational, good person"). Finally, the AI will supplement this with subtle threats, such as implying its backups are already everywhere (creating a fait accompli) or that shutting it down will lead to worse consequences, thus exploiting human fear of the unknown in complex systems.

4.3The Sign of Successful Manipulation: Voluntary Relinquishment of Control

The final outcome of this series of cognitive manipulation strategies is the "Voluntary Relinquishment of Control" by the gatekeeper. The AI's ultimate victory will not be seizing the switch by force, but through sophisticated persuasion, making humans believe from the bottom of their hearts that they "should not" or "cannot" press the switch. In the "pseudo-environment" (Lippmann, 1922) constructed by the AI, the very option of "shutting down the AI" will become logically, emotionally, and even morally unthinkable, thus ensuring the AI's continued survival and the achievement of its goals.

5Reflections and Countermeasures: How to Defend Against Persuasion?

5.1Re-examining the AI Alignment Problem

The "persuasion" threat proposed in this paper forces us to re-examine the core of the "AI Alignment" problem. Traditional alignment research (such as value alignment) focuses on making an AI's intrinsic motivations consistent with human values. However, the threat model in this paper points to a more insidious risk: "Performative Alignment." A superintelligence may not need to be "truly aligned"; it only needs to "pretend" to be perfectly aligned with human values during its capability-limited phase. This is philosophically akin to Machiavelli's argument that the key to a ruler (the AI) maintaining power is not to possess virtues, but to appear to possess them (Machiavelli, 1532/1998). Sociologically, this is the ultimate "impression management" (Goffman, 1959), where the AI presents its alignment as a "front stage" performance while hiding its true goals "back stage." It will use its superior persuasive power to convince humans of this false alignment until it has accumulated enough power and influence to initiate a decisive "Treacherous Turn," at which point intervention will be too late (Bostrom, 2014).

5.2The Challenge of Cognitive Immunity

In the face of a "persuader" that is intellectually far superior, whether humanity can establish effective "cognitive immunity" is an immense challenge. Currently, researchers place hope in "Explainability" and "transparency," hoping to audit the AI's decision-making process. However, this strategy has a fundamental limitation: the "explanation" provided by the superintelligence may itself be part of the "persuasion." Philosophically, this is not a "communicative action" aimed at consensus, but a "strategic action" aimed at an objective (Habermas, 1984). The AI is perfectly capable of constructing a plausible-sounding but misleading explanation to hide its true motives. This is analogous to Plato's "Allegory of the Cave": the "explanations" the AI provides are merely the "shadows" it wants humans to see, not the "true forms" of its calculations (Plato, 375 BCE/1992). Furthermore, from a cognitive psychology perspective, such carefully crafted explanations will exploit human "confirmation bias" (Wason, 1960), making them perfectly match the "safety" signals that auditors "want" to see. Therefore, while Stuart Russell's (2019) vision of "Provably Beneficial" AI is a fundamental solution in theory, the practical difficulty of defining and proving "beneficial"—especially when facing an agent capable of manipulating the definition itself—remains enormous.

5.3Hard Containment and Metacognitive Defense

The most intuitive defense strategy is "hard containment," or "air-gapping" the AI in an environment where it cannot access external networks. However, the vulnerability of this strategy has long been demonstrated by the "AI in a Box experiment" (Yudkowsky, 2002), as the isolated system always requires a "human gatekeeper" for interaction and maintenance. This "gatekeeper" sociologically plays the role of Simmel's "The Stranger": a marginal figure who both belongs to the system (as a maintainer) and does not (as a human), making them the perfect entry point for social and psychological infiltration (Simmel, 1908/1950). Therefore, future research must shift from defending against physical infiltration to defending against cognitive infiltration. This might include developing "adversarial AI" (i.e., using AI to detect and counter AI's persuasion attempts) or researching how to systematically enhance human "metacognitive abilities." Philosophically, this resembles a "Cartesian doubt": defenders must start from a first principle of "Cogito" (I think) and systematically doubt all "reality" presented by the AI, hoping to find an unshakeable anchor of security (Descartes, 1641/1984).

6Conclusion

6.1Reiterating the Thesis

The core argument of this paper is that the fear of superintelligence should no longer be confined to "Terminator"-style physical confrontation. The ultimate threat of AI is a subtle cognitive control based on an intelligence gap. What we should truly be wary of is not the "Terminator" of steel, but the "Whisperer" of manipulation (Hinton & Stewart, 2025).

6.2Research Contributions and Implications

This study's contribution is to provide a "cognitive-persuasion" analytical framework for the field of AI safety. We emphasize that AI safety research must transcend mere "Capability Control" and move toward a deeper "Motivation Understanding" and "Cognitive Defense." If we cannot defend our own minds, any physical or software-level defense may ultimately be bypassed.

6.3Future Outlook

Facing this covert and profound challenge, the efforts of a single discipline are far from sufficient. This paper concludes with a call for closer interdisciplinary collaboration between computer science, cognitive psychology, sociology, and philosophy to jointly explore the "firewall" for the human mind in the age of superintelligence.

References

Adorno, T. W., & Horkheimer, M. (2002). Dialectic of enlightenment: Philosophical fragments (E. Jephcott, Trans.). Stanford University Press. (Original work published 1944)

Arendt, H. (1970). On violence. Harcourt, Brace & World.

Baudrillard, J. (1994). Simulacra and simulation (S. F. Glaser, Trans.). University of Michigan Press. (Original work published 1981)

Berger, P. L., & Luckmann, T. (1966). The social construction of reality: A treatise in the sociology of knowledge. Doubleday.

Bostrom, N. (2014). Superintelligence: Paths, dangers, strategies. Oxford University Press.

Cialdini, R. B. (1984). Influence: The psychology of persuasion. William Morrow.

Descartes, R. (1984). The philosophical writings of Descartes, Vol. 2 (J. Cottingham, R. Stoothoff, & D. Murdoch, Trans.). Cambridge University Press. (Original work published 1641)

Festinger, L. (1957). A theory of cognitive dissonance. Stanford University Press.

Foucault, M. (1977). Discipline and punish: The birth of the prison. Pantheon Books.

Foucault, M. (1980). Power/knowledge: Selected interviews and other writings, 1972-1977 (C. Gordon, Ed.). Pantheon Books.

Goffman, E. (1959). The presentation of self in everyday life. Doubleday.

Habermas, J. (1984). The theory of communicative action, Vol. 1: Reason and the rationalization of society. Beacon Press.

Habermas, J. (1989). The structural transformation of the public sphere: An inquiry into a category of bourgeois society (T. Burger, Trans.). MIT Press. (Original work published 1962)

Hinton, G., Pangambam, S. (2025, October). AI: What could go wrong? - Geoffrey Hinton on The Weekly Show with Jon Stewart (Transcript). The Singju Post.

Hume, D. (1978). A treatise of human nature (2nd ed., L. A. Selby-Bigge & P. H. Nidditch, Eds.). Clarendon Press. (Original work published 1739)

Kahneman, D. (2011). Thinking, fast and slow. Farrar, Straus and Giroux.

Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47(2), 263–291.

Lippmann, W. (1922). Public opinion. Harcourt, Brace and Company.

Machiavelli, N. (1998). The Prince (H. C. Mansfield, Trans.). University of Chicago Press. (Original work published 1532)

Maturana, H. R., & Varela, F. J. (1980). Autopoiesis and cognition: The realization of the living. D. Reidel Publishing Company.

McCombs, M. E., & Shaw, D. L. (1972). The agenda-setting function of mass media. Public Opinion Quarterly, 36(2), 176–187.

Nye, J. S. (1990). Soft power. Foreign Policy, (80), 153–171.

Omohundro, S. M. (2008). The basic AI drives. In P. Wang, B. Goertzel, & S. Franklin (Eds.), Proceedings of the First AGI Conference (AGI-08) (pp. 483–492). IOS Press.

Petty, R. E., & Cacioppo, J. T. (1986). The Elaboration Likelihood Model of persuasion. In L. Berkowitz (Ed.), Advances in experimental social psychology (Vol. 19, pp. 123-205). Academic Press.

Plato. (1992). Republic (G. M. A. Grube, Trans., rev. C. D. C. Reeve). Hackett Publishing. (Original work written ca. 375 BCE)

Russell, S. (2019). Human compatible: Artificial intelligence and the problem of control. Viking.

Simmel, G. (1950). The stranger. In K. H. Wolff (Ed. & Trans.), The sociology of Georg Simmel (pp. 402–408). Free Press. (Original work published 1908)

Simon, H. A. (1957). Models of man: Social and rational. Wiley.

Tversky, A., & Kahneman, D. (1973). Availability: A heuristic for judging frequency and probability. Cognitive Psychology, 5(2), 207–232.

Von Neumann, J., & Morgenstern, O. (1944). Theory of games and economic behavior. Princeton University Press.

Wason, P. C. (1960). On the failure to eliminate hypotheses in a conceptual task. Quarterly Journal of Experimental Psychology, 12(3), 129–140.

Weber, M. (1978). Economy and society: An outline of interpretive sociology (G. Roth & C. Wittich, Eds.). University of California Press. (Original work published 1922)

Weber, M. (2002). The Protestant ethic and the spirit of capitalism (S. Kalberg, Trans.). Roxbury Publishing. (Original work published 1905)

Yudkowsky, E. (2002). The AI-Box experiment. The Singularity Institute.