AI-Driven Personalized Influence Operations: A Comprehensive Interdisciplinary Analysis#

1. Executive Summary#

The rapid proliferation of artificial intelligence, machine learning, and vast digital surveillance ecosystems has introduced theoretical and practical mechanisms for highly individualized behavioral manipulation. This report presents an exhaustive interdisciplinary analysis of AI-driven personalized influence operations. By synthesizing empirical research across computational social science, behavioral psychology, political science, data privacy law, and ethics, the analysis evaluates the efficacy, mechanics, and legal implications of tailoring persuasive communications to individual psychological profiles. The scope of this inquiry spans political, military, ideological, commercial, criminal, and interpersonal settings, examining how diverse data streams are operationalized to shape human behavior. Contrary to pervasive commercial and political marketing claims, recent rigorous meta-analyses and large-scale empirical experiments demonstrate that the end-to-end persuasive effectiveness of psychological microtargeting is severely limited1. The capacity to accurately infer stable psychological traits from digital footprints is significantly lower than historically reported, and the marginal persuasive advantage of microtargeted messages over generic messaging approaches zero when controlling for pervasive methodological flaws1. Furthermore, attempts to deploy artificial intelligence for emotion recognition are overwhelmingly based on outdated scientific paradigms that fail to account for situational and cultural variability, rendering them scientifically weak and highly prone to discriminatory outcomes4. However, the lack of immediate, overwhelming persuasive efficacy does not negate the profound societal risks inherent in these systems. The deployment of AI-driven personalized influence introduces severe threats to cognitive liberty, freedom of thought, and fundamental privacy rights6. Covert profiling and algorithmic adaptation mechanisms—such as multi-armed bandits and generative large language models—create unprecedented asymmetries of information and power. Vulnerable populations face disproportionate risks of exploitation, particularly in commercial, criminal, and ideological settings where real-time conversational behavior and situational distress are weaponized. Consequently, regulatory frameworks, notably the European Union’s Artificial Intelligence Act and the European Convention on Human Rights, are rapidly evolving to prohibit manipulative artificial intelligence practices and protect the inviolability of the human mind8. This report details the theoretical mechanisms, empirical realities, legal boundaries, and necessary defensive architectures surrounding personalized influence operations.

2. Definitions and Boundaries#

To establish a rigorous analytical framework, it is necessary to delineate the boundaries of various information and influence practices. A personalized influence operation is defined as an organized attempt to alter an individual’s perceptions, attitudes, emotions, decisions, or behavior by tailoring communications to information inferred or collected about that individual. AI-driven personalized influence refers to the use of machine learning or generative artificial intelligence to infer personal characteristics, select persuasive approaches, generate individualized communications, adapt messages based on responses, or optimize influence over time. Distinctions must be made across a spectrum of intent, transparency, and user autonomy. Personalization and ordinary customization involve the adaptation of interfaces or services based on explicit user preferences or past behavior to enhance utility, preserving user agency. Recommendation systems provide algorithmic suggestions of content, products, or connections based on collaborative filtering or content-based similarities. Advertising encompasses transparent commercial messaging intended to promote a product, service, or brand. Political microtargeting involves the segmentation of voters into narrow cohorts based on data analytics to deliver specific political advertisements or messaging10. Social engineering represents the psychological manipulation of individuals into performing actions or divulging confidential information, often associated with cybersecurity breaches and criminal settings. Further distinctions rely on the preservation of rational deliberation. Persuasion is the transparent and rational attempt to influence beliefs or behaviors through argumentation or emotional appeal, preserving the target’s autonomy. Manipulation constitutes covert influence that exploits cognitive biases, bypasses rational deliberation, and subtly subverts the target's autonomous decision-making11. Coercion forces an individual to act against their will through threats, intimidation, or the removal of viable alternatives. Deception involves the intentional presentation of false information, such as deepfakes or fabricated narratives, to alter perceptions. Finally, exploitation is the intentional targeting of an individual's known vulnerabilities, such as age, grief, financial distress, or cognitive decline, for the benefit of the operator. The primary harms associated with AI-driven influence frequently stem not from the persuasive content itself, but from the covert use of personal data and the fundamental asymmetry of knowledge between the algorithmic system and the user. Continuous automated adaptation ensures the system learns about the user’s vulnerabilities at a scale and speed the user cannot match, resulting in the inability of the target to recognize that an influence process is actively occurring.

Table 1: Typology of Personalization and Influence#

3. Data Sources and Profiling Methods#

AI-driven influence operations rely on the continuous ingestion of vast quantities of behavioral, demographic, and contextual data. This data is obtained through platforms, third-party data brokers, public records, and continuous surveillance architectures. The reliability of the inferences drawn from these data sources varies wildly, with certain claims rooted in robust statistical reality and others bordering on pseudoscience. Demographic and geographic information remains highly reliable. Data brokers accurately aggregate age, gender, location, and income through purchase histories, location tracking, and public records12. In criminal and interpersonal settings, such as stalking or domestic abuse, the acquisition of precise geolocation data from commercial brokers or platform breaches presents an acute physical threat. Browsing, search, purchase, viewing, and engagement histories provide highly reliable indicators of current interests and immediate intent. These digital traces are heavily utilized in commercial and ideological settings to target individuals based on immediate desires or inquiries. Social networks and community memberships provide reliable insights into peer influence dynamics and ideological echo chambers. Natural Language Processing models can accurately classify cultural identity, values, and political orientation based on text generated by the user on social media platforms14. However, the reliability of inferring deep personality traits, or psychographics, is severely exaggerated. Claims that the "Big Five" personality traits (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) can be highly accurately predicted from digital footprints are scientifically weak. Comprehensive meta-analyses reveal that only about 5% of the variance in human personality can be reliably predicted from digital footprints when controlling for methodological flaws and data leakage1. Claims surrounding the automated inference of emotional states and situational vulnerability are even more problematic. Artificial intelligence systems claiming to perform emotion recognition via facial micro-expressions or vocal analysis are built on fundamentally flawed scientific premises. Leading psychological research, following exhaustive reviews of over a century of literature, indicates that it is impossible to confidently infer an internal emotional state from isolated facial movements independent of context and culture4. These systems detect physical movements, such as a furrowed brow, but fail to accurately map them to internal states, conflating concentration or confusion with anger4. Despite this, such pseudoscientific systems are marketed for military interrogation, criminal justice, and employment screening, leading to severe misclassification risks17. Conversely, inferences regarding life events, financial stress, loneliness, illness, grief, or social isolation are often highly accurate because they are based on direct search queries, support group memberships, or sudden shifts in purchasing behavior. In criminal and commercial settings, these data points are highly prized for exploitation. Real-time conversational behavior, captured via interactions with customer service bots or generative language models, provides an unprecedented data stream, allowing systems to dynamically infer frustration, compliance, or gullibility based on keystroke dynamics, response latency, and sentiment analysis.

Table 2: Mapping Data Types to Inferences, Risks, and Controls#

4. Personalization and Adaptive-Generation Mechanisms#

The transition from static microtargeting to dynamic, AI-driven influence operations is facilitated by two primary computational advancements: generative artificial intelligence, specifically large language models, and reinforcement learning algorithms. Historically, microtargeting required human operators to manually draft distinct messages for specific audience segments, creating a bottleneck in scale and specificity. Generative AI fundamentally alters this constraint, entirely removing the human operator from the loop of individualized authorship. Generative models can ingest an individual's digital profile—comprising political leanings, age, geographic location, and inferred values—as a system prompt to autonomously author hyper-personalized, context-aware persuasive messages at scale3. This capability allows political campaigns, commercial entities, or state-sponsored disinformation actors to instantly adjust tone, vocabulary, and framing to match a user's specific linguistic style or ideological vulnerability20. In interpersonal or criminal settings, such as romance scams or spear-phishing, generative models can maintain hundreds of simultaneous, personalized conversations, simulating intimacy or authority with perfect contextual recall. AI-driven influence operations do not rely solely on static, initial predictions; they adapt continuously based on real-time feedback loops. Metrics such as clicks, viewing duration, scrolling behavior, reply sentiment, and conversion rates serve as immediate reward signals for adaptive algorithms. This continuous optimization is frequently modeled using Multi-Armed Bandits and Thompson Sampling. In a multi-armed bandit framework, the algorithm faces a persistent exploration-exploitation dilemma: it must choose whether to deploy a known successful persuasive message variant to maximize immediate engagement (exploitation) or test a novel, untested message to gather more data on the user's hidden preferences (exploration)21. Thompson sampling addresses this by maintaining a Bayesian posterior distribution of the estimated success rate for each message variant, updating these probabilities instantaneously as user feedback is received23. Through incentivized exploration, the algorithm leverages information asymmetry to refine its understanding of the user's susceptibilities over time. Because the algorithm collects observations from millions of past interactions across other users, it possesses an informational advantage, allowing it to predict which exploratory messages might eventually yield a higher probability of persuasion25. This creates a system that systematically converges on the most manipulative or persuasive vectors for a specific individual, adjusting its approach based on imperceptible shifts in user behavior.

5. Evidence Concerning Persuasive Effectiveness#

A central question in the study of personalized influence operations is whether psychographic personalization consistently outperforms ordinary audience segmentation or generic messaging. Evidence across political, commercial, and experimental domains reveals a stark contrast between theoretical marketing claims and rigorous empirical realities, highlighting significant replication problems and disputes within computational social science. Early observational studies and field experiments on social media platforms fueled the narrative that psychographic microtargeting was a highly potent tool for mass persuasion. Research suggested that aligning persuasive appeals with a recipient's dominant personality profile—for example, targeting extraverts with extraversion-themed advertisements—significantly increased click-through rates and conversions26. These findings, heavily promoted by political consulting firms and digital marketing agencies, suggested that human autonomy was fundamentally vulnerable to algorithmic personality mapping28. However, subsequent methodological scrutiny has severely downgraded these claims, revealing that much of the perceived effectiveness was an artifact of flawed experimental designs. A recent, exhaustive meta-analysis spanning forty-one studies in marketing, psychology, and computer science evaluated the end-to-end effectiveness of psychological targeting1. The researchers identified pervasive methodological errors in the foundational literature, particularly data leakage, where machine learning models were inadvertently evaluated on data that informed their training1. More critically, many early studies failed to utilize a trait-by-treatment interaction design, thereby failing to isolate the personalization effect from the main effect of a generally appealing advertisement1. When these experimental design flaws are controlled, the incremental persuasive effect of personality-tailored messages on behavioral outcomes approaches zero1. The meta-analysis established a dual failure: digital footprints can only predict a negligible variance in human personality, and the downstream deployment of messages based on these weak inferences yields negligible persuasive returns over generic messaging1. In political environments, rigorous experimental data confirms this limitation. While messages generated by advanced large language models are generally persuasive, microtargeted generative messages are not statistically more persuasive than generic, non-targeted messages generated by the same models3. The scientific consensus indicates that the true threat of generative artificial intelligence in influence operations lies in its ability to cheaply scale high-quality generic persuasion, rather than its ability to execute surgical psychographic manipulation3. Therefore, while the technology can generate personalized content efficiently, the assumption that psychographic personalization acts as a reliable mechanism for overriding human agency is scientifically weak and largely a product of industry speculation.

6. Case Studies#

To contextualize the mechanisms and limitations of AI-driven influence, three major case studies illustrate the divide between perceived capability, regulatory reaction, and empirical evidence.

Case Study 1: Cambridge Analytica and the ICO Investigation#

In the political and ideological setting, Cambridge Analytica gained international notoriety following allegations that it harvested the data of up to eighty-seven million Facebook users to build psychographic profiles for political microtargeting during the 2016 United States Presidential Election and the United Kingdom Brexit referendum32. The firm claimed to possess thousands of data points on millions of voters, enabling them to alter electoral outcomes through psychological manipulation32. Despite the intense public panic and subsequent regulatory backlash, post-incident analyses dismantled the company's claims of persuasive efficacy. A comprehensive three-year investigation by the UK Information Commissioner's Office concluded that Cambridge Analytica utilized standard data science techniques, not unprecedented or magical predictive algorithms33. Furthermore, political scientists and the regulatory investigation found no empirical evidence that the company's data-driven microtargeting meaningfully influenced voting habits, altered the Brexit outcome, or shifted the US election33. The documented harms were primarily related to egregious data protection violations, the covert nature of the processing, and the erosion of democratic trust, rather than actual mass cognitive manipulation.

Case Study 2: Facebook Psychological Targeting Experiments (Matz et al. 2017\) vs. Meta-Analytic Review#

In the commercial setting, a highly cited 2017 field study by Matz et al. demonstrated that targeting Facebook users with advertisements tailored to their inferred levels of extraversion and openness resulted in up to forty percent more clicks and fifty percent more purchases27. The study utilized Facebook page likes to infer personality and served matched or mismatched commercial advertisements. While the 2017 study was groundbreaking, subsequent researchers raised profound concerns regarding internal validity, specifically highlighting how Facebook’s proprietary optimization algorithms might confound the results by naturally favoring advertisements that perform well generally, regardless of the psychological match37. In 2026, an exhaustive meta-analysis mathematically modeled these experimental designs, correcting for data leakage and algorithmic confounding2. The analysis demonstrated that the true end-to-end effectiveness of such psychological targeting is negligible1. The evidence indicates that while commercial platforms claim high personalization effectiveness to justify advertising premiums, independent rigorous validation fails to replicate meaningful, isolated personalization effects1.

Case Study 3: Large Language Model Political Microtargeting (Hackenburg & Margetts, 2024\)#

With the advent of advanced generative models, researchers conducted a large-scale pre-registered experiment to test whether political microtargeting using large language models offered greater persuasive returns than generic human-authored or AI-authored messages3. The model was provided with detailed demographic and political data about participants to craft tailored arguments on polarized policy issues. The study found that while the generated messages were generally highly persuasive, the specifically microtargeted messages failed to demonstrate a statistically significant increase in persuasive impact compared to generic, non-targeted messages3. The researchers theorized that either text-based microtargeting is an inherently flawed strategy, or current models cannot reliably leverage fine-grained demographic interactions to enhance persuasion3. The findings reinforce the conclusion that the relevant harm of AI in politics stems from the volume and speed of content generation, not necessarily a breakthrough in personalized psychological control3.

7. Vulnerable Populations and Unequal Harms#

While the aggregate persuasive power of personalized influence may be limited across the general population, specific populations remain disproportionately vulnerable. The risk of erroneous profiling, stereotyping, and misclassification introduces severe unequal harms, particularly when optimization algorithms seek the path of least resistance to achieve their programmed objectives. Emotion recognition artificial intelligence represents a critical vector for discrimination. Because these systems are trained on datasets that historically lack cultural and racial diversity, they exhibit significantly reduced accuracy when analyzing the expressions of marginalized groups, encoding and perpetuating structural inequalities18. The deployment of pseudoscientific emotion recognition in employment screening, border control, or law enforcement can lead to discriminatory targeting, where an algorithm incorrectly classifies a natural resting face as angry or hostile, leading to unwarranted suspicion or denial of opportunity4. Furthermore, manipulative personalization inherently preys on situational vulnerabilities. Individuals experiencing cognitive decline, acute financial distress, severe grief, or pervasive loneliness can be easily identified through search histories, engagement with specific online communities, and social media behavior. In commercial settings, algorithmic systems optimized for conversion can automatically target these populations with predatory lending schemes or highly exploitative medical treatments. In criminal and interpersonal settings, bad actors utilize data breaches and scraped social media data to construct highly personalized spear-phishing campaigns or romance scams, leveraging generative language models to simulate affection and extract financial resources from isolated individuals. In ideological and military settings, recommender systems identify individuals exhibiting signs of social alienation, progressively guiding them toward radicalized communities or state-sponsored disinformation networks. In these instances, the harm is derived directly from the coercive exploitation of an inferred, temporary emotional state or persistent vulnerability.

8. Privacy, Autonomy, and Human-Rights Analysis#

The covert nature of AI-driven influence operations constitutes a systemic threat to individual autonomy, informed consent, and political equality. When influence is optimized continuously by an algorithm that observes the user while remaining utterly opaque itself, a fundamental asymmetry of knowledge and power is established. The debate surrounding artificial intelligence manipulation has increasingly centered on the concept of cognitive liberty and the legal protections afforded to the inner workings of the human mind. Under Article 9 of the European Convention on Human Rights, freedom of thought is recognized as an absolute right protecting the forum internum—the internal realm of thoughts, beliefs, and unexpressed opinions6. Historically, this right protected individuals from state indoctrination and physical coercion. However, the capacity of AI systems to monitor behavioral traces, infer psychological states, and subtly shape the digital architecture in which beliefs are formed challenges the traditional limits of Article 97. Legal and sociological scholars argue that manipulative artificial intelligence—by bypassing rational deliberation and operating below the threshold of conscious awareness—interferes directly with the cognitive conditions necessary for independent belief formation6. When political microtargeting environments simulate neutrality while asymmetrically structuring a voter's attention and information diet, they threaten the foundations of political equality and democratic public discourse10. The autonomy of the individual is subverted not through overt force, but through an engineered environment that anticipates and preempts the individual's capacity to choose otherwise.

9. Applicable Legal and Regulatory Frameworks#

Global regulatory frameworks are rapidly adapting to address the abuses of data-driven profiling and artificial intelligence manipulation. The regulatory landscape is shifting from a sole focus on data privacy to encompassing the behavioral impacts of algorithmic systems. The European Union Artificial Intelligence Act represents the most comprehensive attempt to regulate these technologies globally. Under Article 5, the Act strictly prohibits AI practices that pose an unacceptable risk, specifically targeting manipulative systems6. The legislation bans the deployment of systems that utilize subliminal techniques beyond a person’s consciousness or purposefully manipulative techniques that materially distort human behavior in a manner likely to cause physical or psychological harm8. Crucially, the Act explicitly prohibits the use of emotion recognition systems in workplaces and educational institutions due to their high potential for discrimination and lack of scientific validity8. In the realm of behavioral advertising and commercial personalization, the General Data Protection Regulation (GDPR) has forced significant structural changes. Following a legal consensus that large online platforms could not rely on contractual necessity or legitimate interest to process user data for behavioral advertising, platforms pivoted to a "Pay or Consent" model42. Under this framework, users are forced to either consent to comprehensive data tracking or pay a recurring subscription fee to access the service. In 2024, the European Data Protection Board issued a non-binding opinion concluding that in most cases, this binary choice fails to satisfy the GDPR’s stringent requirement for freely given consent19. The Board argued that imposing a fee on non-consenting users forces a detriment, particularly for platforms that play an essential role in social life. The Board recommended that large platforms must offer a genuine, free alternative that relies on less intrusive methods, such as contextual advertising, ensuring that fundamental privacy rights are not commodified19. In the United States, lacking a comprehensive federal privacy statute, the Federal Trade Commission relies on its authority to police unfair and deceptive practices. The Commission has actively pursued data brokers that enable abusive targeting. For instance, the Commission took enforcement action against Kochava, prohibiting the data broker from selling sensitive location data that could be used to track individuals' visits to reproductive health clinics, places of worship, and other highly sensitive locations12. This signifies a growing regulatory consensus that the indiscriminate aggregation and sale of behavioral data for downstream profiling constitutes a severe, actionable consumer harm.

10. Detection and Auditing Methods#

Because AI-driven influence operates dynamically and covertly, it is remarkably difficult for an individual target to recognize when they are being manipulated. Developing defensive signals and rigorous auditing methods is essential for societal resilience and the protection of democratic institutions. To combat the deceptive use of generative models in influence operations, transparency mandates are critical defensive signals. Initiatives like the Coalition for Content Provenance and Authenticity utilize cryptographic metadata, digital watermarks, and security fingerprints to label AI-generated content47. Under the EU AI Act, providers of general-purpose AI are legally required to tag synthetic content, providing a defensive signal to users that the media they are consuming is artificially generated and potentially tailored47. At the individual level, defensive signals that a person is receiving coordinated individualized influence may include rapid, unnatural shifts in conversational tone from a chatbot, hyper-specific references to recent offline life events that suggest data aggregation, or an uncanny alignment of disparate content recommendations across seemingly disconnected platforms. Independent researchers face significant ethical and methodological challenges in measuring personalized influence without reproducing the very harms they seek to study. Traditional field experiments often require researchers to deploy manipulative advertisements or deceptive bots to measure their effects on human subjects, raising profound research ethics concerns. To circumvent this, computational social scientists increasingly rely on data-generating simulations and closed-form algorithmic analyses that mathematically model the effects of recommender systems without experimenting on live populations30. Furthermore, data donation initiatives allow users to voluntarily and securely contribute their digital tracking data to researchers via controlled APIs or browser extensions48. This enables auditors to reverse-engineer recommender systems and observe how political or commercial targeting operates in the wild without actively participating in or exacerbating the surveillance ecosystem.

11. Defensive Design and Policy Recommendations#

To safeguard democratic integrity, individual autonomy, and cognitive liberty from manipulative personalization, a multi-layered defensive architecture and policy framework is required. First, regulatory bodies must mandate strict data minimization principles. Policies should enforce severe limits on the collection of secondary behavioral data. Contextual advertising—which targets individuals based on the specific content currently being viewed rather than the user's historical, cross-platform psychographic profile—should be established as the default standard for digital commercial ecosystems, largely eliminating the economic necessity for vast surveillance databases. Second, recognizing the pseudoscientific basis and discriminatory outcomes of automated emotion recognition, policymakers must extend the prohibitions found within the EU AI Act. The ban on emotion inference must move beyond the workplace and educational sectors to explicitly prohibit its use in law enforcement, border control, and broad consumer profiling, as these applications inherently risk profound human rights violations. Third, algorithmic transparency and explainability must be integrated into user interfaces. Platforms must provide users with granular, real-time disclosures indicating precisely why a specific piece of tailored content was delivered, which data points influenced the decision, and the mechanism to sever that data link. Fourth, regarding informed consent, regulatory bodies must rigorously enforce the European Data Protection Board's guidance against binary "Pay or Consent" models. Fundamental privacy rights and freedom from continuous surveillance must not be restricted solely to demographics capable of affording premium subscription fees. Finally, jurisprudence surrounding Article 9 of the European Convention on Human Rights and analogous constitutional protections must be modernized to explicitly recognize the right to mental integrity in the digital sphere, establishing that covert algorithmic manipulation constitutes a violation of fundamental human rights.

12. Research Gaps#

While the empirical evaluation of single-turn psychological targeting is maturing, significant gaps in the literature remain. The vast majority of meta-analyses and controlled studies have evaluated isolated, static messages—such as a single commercial advertisement or a one-off political text prompt1. The true frontier of AI-driven influence lies in multi-turn, interactive language model chatbots that build parasocial relationships over extended periods3. Future research must prioritize evaluating whether continuous, conversational adaptation yields the persuasive returns that static microtargeting fundamentally failed to deliver. Additionally, there is a critical need for longitudinal sociological studies assessing the cumulative effects of algorithmic environments on cognitive liberty, the homogenization of political discourse, and the long-term psychological impacts on vulnerable populations repeatedly targeted by manipulative architectures.

13. Conclusion#

AI-driven personalized influence operations occupy a paradoxical space in contemporary technology. Empirically, the capacity of these systems to infer deep psychological traits and successfully manipulate behavior at the individual level is drastically overstated by commercial vendors, political consultants, and early observational studies. Methodologically rigorous analyses indicate that psychographic microtargeting yields negligible downstream persuasive effects compared to high-quality generic messaging, and the foundational science of emotion recognition AI is fundamentally flawed. However, the infrastructural risks generated by these systems remain immense. The automated ingestion of digital footprints, combined with generative text models and reinforcement learning, creates an opaque ecosystem that threatens fundamental privacy, fosters discriminatory profiling, and systematically exploits situational vulnerabilities. The legal and philosophical pivot toward defining cognitive liberty and banning manipulative artificial intelligence acknowledges a critical truth: the harm of personalized influence lies not necessarily in its immediate persuasive success, but in its systemic subversion of human autonomy, informed consent, and political equality. Addressing this systemic threat requires abandoning the speculative illusion of algorithmic mind-control while simultaneously enforcing rigorous data minimization, demanding transparent provenance, and protecting the inviolable right to freedom of thought in an increasingly customized digital world.

14. Annotated Bibliography#

This section provides a narrative review of the foundational texts, primary empirical data, and legal frameworks informing this report, synthesizing the evolution of the field from early claims of efficacy to contemporary skepticism and robust legal regulation. Matz, S. C., Kosinski, M., Nave, G., & Stillwell, D. J. (2017). Psychological targeting as an effective approach to digital mass persuasion. Proceedings of the National Academy of Sciences27. This foundational paper represents the zenith of optimism regarding the efficacy of psychographic microtargeting. Conducting field experiments on millions of Facebook users, the authors claimed that matching advertising copy to inferred levels of extraversion and openness resulted in massive increases in click-through and conversion rates. While highly influential in political and commercial consulting, this study set the baseline for subsequent, intense methodological critiques regarding algorithmic confounding on commercial platforms. Perla, R., Maran, T., Bagci, B., Kraus, S., Kanbach, D. K., & Bouncken, R. B. (2026). The (In)Effectiveness of Psychological Targeting: A Meta-Analytic Review. Psychology & Marketing1. Representing a critical paradigm shift, this comprehensive meta-analysis systematically dismantled the claims made by earlier observational studies. By enforcing strict machine learning evaluation standards, analyzing data leakage, and focusing exclusively on clean experimental designs, the authors demonstrated that digital footprints predict only a fraction (\~5%) of personality variance and that downstream persuasive effects approach zero. This text is vital for establishing the empirical reality of microtargeting's severe limitations. Hackenburg, K., & Margetts, H. (2024). Evaluating the persuasive influence of political microtargeting with large language models. Proceedings of the National Academy of Sciences3. This study bridges the gap between traditional microtargeting and modern generative AI. The authors tested whether language models equipped with demographic and political data could out-persuade generic messages. Their finding—that models are highly persuasive overall but gain no statistically significant advantage from microtargeting—suggests that the primary threat of generative AI is its scale and coherence, rather than individualized psychological manipulation. Barrett, L. F., Adolphs, R., Marsella, S., Martinez, A. M., & Pollak, S. D. (2019). Emotional expressions reconsidered: Challenges to inferring emotion from human facial movements. Psychological Science in the Public Interest4. A landmark review dismantling the scientific foundation of emotion recognition systems. Barrett and colleagues reviewed over a century of literature to conclude that it is impossible to reliably infer internal emotional states from isolated facial movements. This text is crucial for understanding why current emotion applications are fundamentally flawed and prone to discriminatory profiling, forming the scientific basis for regulatory bans on such technology. European Parliament and Council (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act).8. The definitive global legal framework addressing the theoretical and practical risks of algorithmic manipulation. By explicitly defining and prohibiting manipulative practices that subvert autonomy through subliminal techniques, the AI Act shifts the global regulatory focus toward safeguarding cognitive liberty, preventing algorithmic exploitation of human vulnerabilities, and outlawing the use of emotion recognition in sensitive environments.

Works cited#

1. The (In)Effectiveness of Psychological Targeting: A Meta‐Analytic Review - ResearchGate, https://www.researchgate.net/publication/396984605\_The\_InEffectiveness\_of\_Psychological\_Targeting\_A\_Meta-Analytic\_Review 2. The (In)Effectiveness of Psychological Targeting: A Meta-Analytic Review - OPUS - BSZ, https://opus.bsz-bw.de/hhlpd/frontdoor/index/index/searchtype/authorsearch/author/Kraus+Sascha/start/12/rows/20/nav/next 3. Does political microtargeting with large language models offer persuasive returns? ask Kobi Hackenburg and Professor Helen Margetts. - Oxford Internet Institute, https://www.oii.ox.ac.uk/does-political-microtargeting-with-large-language-models-offer-persuasive-returns-ask-kobi-hackenburg-and-professor-helen-margetts/ 4. Experts Say 'Emotion Recognition' Lacks Scientific Foundation | ACLU, https://www.aclu.org/news/privacy-technology/experts-say-emotion-recognition-lacks-scientific 5. Emotional Expressions Reconsidered: Challenges to Inferring Emotion From Human Facial Movements - PubMed, https://pubmed.ncbi.nlm.nih.gov/31313636/ 6. Cognitive freedom and legal accountability: Rethinking the EU AI act's theoretical approach to manipulative AI as unacceptable risk | Cambridge Forum on AI: Law and Governance, https://www.cambridge.org/core/journals/cambridge-forum-on-ai-law-and-governance/article/cognitive-freedom-and-legal-accountability-rethinking-the-eu-ai-acts-theoretical-approach-to-manipulative-ai-as-unacceptable-risk/45F379C0707D7A415C042BB08088F88F 7. Think Again: Freedom of Thought in the Age of Artificial Intelligence - OSCE.org, https://cdn.osce.org/sites/default/files/f/documents/0/e/597450.pdf 8. EU AI Act - Updates, Compliance, Training, https://www.artificial-intelligence-act.com/ 9. The Inviolable Mind: Freedom of Thought as a Constitutional Limit to Emotional AI, https://digi-con.org/the-inviolable-mind-freedom-of-thought-as-a-constitutional-limit-to-emotional-ai/ 10. Double harm to voters: data-driven micro-targeting and democratic public discourse, https://policyreview.info/articles/analysis/double-harm-voters-data-driven-micro-targeting-and-democratic-public-discourse 11. Freedom of Thought, Conscience, and AI | Journal of Law and Religion | Cambridge Core, https://www.cambridge.org/core/journals/journal-of-law-and-religion/article/freedom-of-thought-conscience-and-ai/BCE0DD82A739B99D979CF584451F3473 12. FTC v Kochava, Inc. | Federal Trade Commission, https://www.ftc.gov/legal-library/browse/cases-proceedings/ftc-v-kochava-inc 13. FTC to Ban Kochava and Subsidiary from Selling Sensitive Location Data to Settle Charges They Sold Location Data Linked to Millions of Mobile Devices, https://www.ftc.gov/news-events/news/press-releases/2026/05/ftc-ban-kochava-subsidiary-selling-sensitive-location-data-settle-charges-they-sold-location-data 14. Personality and susceptibility to political microtargeting - UvA-DARE (Digital Academic Repository), https://pure.uva.nl/ws/files/154662500/Personality\_and\_susceptibility\_to\_political\_microtargeting.pdf 15. Personality or Value: A Comparative Study of Psychographic Segmentation Based on an Online Review Enhanced Recommender System - MDPI, https://www.mdpi.com/2076-3417/9/10/1992 16. You can't determine emotion from someone's facial movements and neither can AI., https://cos.northeastern.edu/you-cant-determine-emotion-from-someones-facial-movements-and-neither-can-ai/ 17. AI systems claiming to 'read' emotions pose discrimination risks - The Guardian, https://www.theguardian.com/technology/2020/feb/16/ai-systems-claiming-to-read-emotions-pose-discrimination-risks 18. AI Emotion Recognition Is a Pseudoscientific Multi-Billion Dollar Industry - Reddit, https://www.reddit.com/r/Futurology/comments/vxlu3p/ai\_emotion\_recognition\_is\_a\_pseudoscientific/ 19. Europe: EDPB issues Opinion on 'consent or pay' models deployed by large online platforms | Privacy Matters, https://privacymatters.dlapiper.com/2024/04/europe-edpb-issues-opinion-on-consent-or-pay-models-deployed-by-large-online-platforms/ 20. Persuasion with Large Language Models: A Survey of Empirical Evidence, Study Methodologies, and Ethical Implications - arXiv, https://arxiv.org/html/2411.06837v2 21. "SINGLE-AGENT AND MULTI-AGENT BANDITS" by Jacob Adkins - JBC Commons - New College of Florida, https://digitalcommons.ncf.edu/theses\_etds/6504/ 22. Incentivized Exploration - Alex Slivkins, https://slivkins.com/work/ijcai21\_tutorial.pdf 23. The Price of Incentivizing Exploration: A Characterization via Thompson Sampling and Sample Complexity - IDEAS/RePEc, https://ideas.repec.org/a/inm/oropre/v71y2023i5p1706-1732.html 24. Incentivized Exploration via Filtered Posterior Sampling - IDEAS/RePEc, https://ideas.repec.org/p/arx/papers/2402.13338.html 25. 1 Exploration and Persuasion - arXiv, https://arxiv.org/pdf/2410.17086 26. (PDF) Personalized Persuasion: Tailoring Persuasive Appeals to Recipients' Personality Traits - ResearchGate, https://www.researchgate.net/publication/224871267\_Personalized\_Persuasion\_Tailoring\_Persuasive\_Appeals\_to\_Recipients'\_Personality\_Traits 27. Psychological targeting as an effective approach to digital mass persuasion - ResearchGate, https://www.researchgate.net/publication/336085256\_Psychological\_targeting\_as\_an\_effective\_approach\_to\_digital\_mass\_persuasion 28. How effective is psychological targeting in advertising? What we can learn about you from just one click., https://www.jbs.cam.ac.uk/2017/how-effective-is-psychological-targeting-in-advertising/ 29. Psychological targeting as an effective approach to digital mass persuasion - PubMed, https://pubmed.ncbi.nlm.nih.gov/29133409/ 30. r-perla/psych\_targeting - GitHub, https://github.com/r-perla/psych\_targeting 31. Evaluating the persuasive influence of political microtargeting with large language models, https://www.semanticscholar.org/paper/Evaluating-the-persuasive-influence-of-political-Hackenburg-Margetts/de9a0af5100600ff21a9ded6e42409e0de89de8a 32. Vincent John Green & Mark Newman -v- Cambridge Analytica (UK) Limited & others - Courts and Tribunals Judiciary, https://www.judiciary.uk/wp-content/uploads/2019/04/17.04.19-cambridge-judgment.pdf 33. Cambridge Analytica - Wikipedia, https://en.wikipedia.org/wiki/Cambridge\_Analytica 34. Lessons from Cambridge Analytica – Using Digital Data to Drive Decision Making, https://exeterindex.org/blog/lessons-from-cambridge-analytica-using-digital-data-to-drive-decision-making/ 35. Understanding Mass Influence - UNSW Sydney, https://www.unsw.edu.au/content/dam/pdfs/unsw-canberra/dri/2023-02-research/2023-02-Understanding-Mass-Influence---A-case-study-of-Cambridge-Analytica.pdf 36. Psychological targeting as an effective approach to digital mass persuasion - PNAS, https://www.pnas.org/doi/10.1073/pnas.1710966114 37. The effects of trait-based personalization in social media advertising - UvA-DARE (Digital Academic Repository), https://pure.uva.nl/ws/files/123222256/1\_s2.0\_S0747563220302776\_main.pdf 38. Presentamos Tribefacturing Ads (ES) - Antropomedia Express, https://www.consumertribes.com/presentamos-tribefacturing-ads/ 39. (PDF) EMOTION RECOGNITION ARTIFICIAL INTELLIGENCE AND THE PSYCHOLOGY OF HUMAN AFFECT: A CRITICAL EXAMINATION OF THEORY, TECHNOLOGY, BIAS, AND THE AFRICAN CONTEXT TABLE OF CONTENTS - ResearchGate, https://www.researchgate.net/publication/404470384\_EMOTION\_RECOGNITION\_ARTIFICIAL\_INTELLIGENCE\_AND\_THE\_PSYCHOLOGY\_OF\_HUMAN\_AFFECT\_A\_CRITICAL\_EXAMINATION\_OF\_THEORY\_TECHNOLOGY\_BIAS\_AND\_THE\_AFRICAN\_CONTEXT\_TABLE\_OF\_CONTENTS 40. Political Manipulation in the Age of Interactive AI: The Case for Protecting Freedom of Thought under the ECHR - Lund University Research Portal, https://portal.research.lu.se/en/publications/political-manipulation-in-the-age-of-interactive-ai-the-case-for-/ 41. European Commission Guidelines on Prohibited AI Practices under the EU Artificial Intelligence Act | Inside Privacy, https://www.insideprivacy.com/artificial-intelligence/european-commission-guidelines-on-prohibited-ai-practices-under-the-eu-artificial-intelligence-act/ 42. White paper: Behavioural advertising, Meta and privacy - Schjødt, https://schjodt.com/news/thoughts-on-behavioural-advertising-meta-and-privacy 43. EDPB publishes urgent binding decision regarding Meta | European Data Protection Board, https://www.edpb.europa.eu/news/news/2023/edpb-publishes-urgent-binding-decision-regarding-meta\_en 44. Pay or consent to have your data used – the complex intersection between the DMA and data protection rules - Taylor & Francis, https://www.tandfonline.com/doi/full/10.1080/13600869.2026.2654229 45. Consent or Pay: One Rule For Some (Large Online Platforms), Another Rule For Everyone Else?, https://www.weil.com/-/media/mailings/2024/q3/consent-or-pay\_one-rule-for-some\_large-online-platforms\_another-rule-for-everyone-else---august-2024.pdf 46. FTC Sues Kochava for Selling Data that Tracks People at Reproductive Health Clinics, Places of Worship, and Other Sensitive Locations, https://www.ftc.gov/news-events/news/press-releases/2022/08/ftc-sues-kochava-selling-data-tracks-people-reproductive-health-clinics-places-worship-other 47. Generative AI and deepfakes: a human rights approach to tackling harmful content, https://www.tandfonline.com/doi/full/10.1080/13600869.2024.2324540 48. Ethical considerations on the use of big data and artificial intelligence in kidney research from the ERA ethics committee - PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC11879022/ 49. Psychological targeting as an effective approach to digital mass persuasion, https://www.semanticscholar.org/paper/Psychological-targeting-as-an-effective-approach-to-Matz-Kosinski/03e1335b2fdbc02e78022dbe90d1bf3ea3081ae5