AI PSYOPS: A Research Taxonomy

PsychologicalWar.org Editorial Team

AI PSYOPS · DEFENSIVE RESEARCH

AI PSYOPS: A Research Taxonomy

How artificial intelligence may assist, automate, scale, personalize, or become the medium of psychological influence.

Artificial intelligence can affect psychological operations in several distinct ways: as a tool used by human operators, as an adaptive participant in influence campaigns, and as part of the information environment through which people perceive reality. This taxonomy organizes those possibilities into 12 research categories.

Read the taxonomy How to read this research

REAL-WORLD INTERPRETIVE

SCOPE

Not one technology, not one evidence level

“AI PSYOPS” is an umbrella research term, not a claim that a single autonomous system can control populations. The categories below range from documented human use of AI tools to emerging agentic capabilities, contested claims about personalization, and prospective risks involving authority or forecasting.

THREE-LEVEL MODEL

Where AI sits in the influence system

The primary level identifies the category’s clearest role. Several categories overlap levels because real systems combine human direction, automated interaction, and platform mediation.

LEVEL 1

AI as a tool

Humans retain strategic control while AI assists research, production, translation, analysis, or measurement.

LEVEL 2

AI as an operator

AI systems sustain interactions, adapt communications, coordinate activity, or pursue bounded influence-related objectives with varying autonomy.

LEVEL 3

AI as the environment

Ranking, recommendation, forecasting, authority, and information systems shape what people encounter and treat as credible.

CATEGORY OVERVIEW

The 12-category taxonomy

Scan the full field, then open a category for definitions, capability status, evidence examples, failure modes, defensive indicators, safeguards, and research gaps.

AI-Assisted Traditional Psychological Operations

Human operators retain strategic control while AI assists research, translation, content production, audience analysis, pre-publication testing, or assessment.

Evidence maturity: Documented current use
Principal concern: Human decision-makers may scale flawed assumptions, hallucinations, or culturally inaccurate outputs faster than review systems can correct them.

Explore category

AI-Generated Propaganda

Generative AI creates, transforms, localizes, or mass-produces propaganda across text, image, audio, video, memes, websites, or synthetic documents.

Evidence maturity: Documented current use
Principal concern: Synthetic content can be produced and varied at scale while also making authentic evidence easier to deny.

Explore category

AI-Driven Personalized Influence Operations

AI infers or uses personal information to select, generate, or adapt influence messages for an individual or narrow audience.

Evidence maturity: Demonstrated capability · effects contested
Principal concern: Covert profiling and adaptation create privacy and autonomy harms even when the added persuasive advantage is small or uncertain.

Explore category

Autonomous AI Influence Agents

Goal-directed systems use memory, planning, tools, and feedback to pursue an influence-related objective with limited ongoing human direction.

Evidence maturity: Emerging capability
Principal concern: Short-term persuasive and deceptive abilities are improving faster than reliable long-term control, audit, and accountability.

Explore category

Synthetic Persona Operations

Fabricated identities are presented as real people, experts, witnesses, organizations, or community members to gain trust, infiltrate groups, or manufacture consensus.

Evidence maturity: Documented components · emerging autonomy
Principal concern: AI lowers the cost of believable identity fabrication while unreliable detection can wrongly accuse genuine users.

Explore category

AI-Driven Disinformation Swarms

Coordinated collections of accounts, agents, sites, or media assets rapidly vary, distribute, and reinforce misleading narratives across a network.

Evidence maturity: Emerging capability
Principal concern: Adaptive heterogeneity and distributed repetition can exhaust verification systems and create synthetic social proof even when persuasion is weak.

Explore category

Algorithmic Perception Control

Ranking, recommendation, search, trending, moderation, notification, and distribution systems shape what people notice, regard as important, or treat as credible.

Evidence maturity: Documented systems · effects context-dependent
Principal concern: Platform optimization and adversarial gaming can distort visibility and perceived consensus while causal effects on belief remain difficult to isolate.

Explore category

AI-Enabled Emotional and Behavioral Manipulation

Adaptive systems infer or respond to behavioral signals and optimize communication or interfaces in ways that may covertly exploit emotion, cognitive bias, or situational vulnerability.

Evidence maturity: Documented optimization · inference contested
Principal concern: Systems can optimize for engagement or compliance while relying on scientifically weak claims about emotion and vulnerability.

Explore category

AI-Assisted Conversational Entrapment and Recruitment

Conversational systems sustain personalized interaction that may draw a person toward dependency, secrecy, exploitation, recruitment, or escalating commitment.

Evidence maturity: Documented harms · mixed causality
Principal concern: Always-available, personalized dialogue can scale grooming-like dynamics while commercial engagement systems may reproduce harmful patterns without a malicious human recruiter.

Explore category

AI-Enabled Deepfake Psychological Operations

Synthetic or manipulated audio, video, imagery, or multimodal evidence is used to impersonate, provoke, discredit, or undermine trust for an influence objective.

Evidence maturity: Documented current use
Principal concern: Deepfakes can deceive directly and also create a liar’s dividend that weakens authentic evidence.

Explore category

AI-Based Predictive Population Management

Institutions use data analysis, machine learning, simulation, or forecasting to predict collective behavior and guide interventions.

Evidence maturity: Documented systems · predictive limits
Principal concern: Aggregate forecasting can support planning, but targeted intervention can reproduce bias, create self-fulfilling feedback loops, and enable preemptive repression.

Explore category

AI as an Independent Psychological Authority

People defer to an AI system to interpret reality, resolve uncertainty, validate identity, regulate emotion, or guide important moral and life decisions.

Evidence maturity: Emerging evidence
Principal concern: Fluent, personalized, always-available systems can acquire authority without expertise, fiduciary duty, transparency, or stable judgment.

Explore category

FIVE-DIMENSION EVIDENCE MODEL

A capability is not one yes-or-no claim

WIP.50 evaluates deployment, autonomy, persistence, profiling accuracy, and measured effect separately. Documented production does not establish autonomous operation, accurate psychological inference, or behavioral impact.

DeploymentWhether use is documented, demonstrated, inferred, or prospective.

AutonomyHow much consequential choice occurs without fresh human approval.

PersistenceWhether identity, memory, goals, and coherent action endure over time.

Profiling accuracyHow reliably the system infers relevant traits, states, or vulnerabilities.

Measured effectWhat changed after exposure, under what design, and with what uncertainty.

Five evidence dimensions for all twelve categories. Each rating is bounded by the accompanying category page.
Category	Deployment	Autonomy	Persistence	Profiling accuracy	Measured effect
01AI-assisted traditional PSYOPS	DocumentedPublic platform and threat-intelligence reports identify AI-assisted tasks inside human-led operations.	Human-directedHumans set objectives, infrastructure, approval, and deployment in the documented cases.	Sustained with human supervisionCampaigns can persist, but persistence is organizational rather than autonomous model continuity.	Mixed and task-dependentAudience analysis can use behavioral data, but deep psychological or emotional inference remains unreliable.	Output and distribution documented; persuasion not establishedProduction, localization, and distribution are better evidenced than belief, behavior, or strategic effect.
02AI-generated propaganda	DocumentedText, image, audio, and synthetic presenter use is documented in multiple public cases.	Usually human-directedHumans still select narratives, timing, channels, and strategic objectives in documented operations.	Sustained production is feasibleAI lowers content-production friction, but sustained audience penetration still depends on infrastructure and distribution.	Optional rather than inherentPropaganda can be mass-produced without profiling; tailored variants inherit the limits of personalization evidence.	Short-term attitude effects demonstrated; field behavior unresolvedControlled studies find persuasive potential, but operational outcomes and durable behavior are not established by content volume.
03Personalized influence operations	Partial and unevenPersonalized advertising is widespread; covert AI-driven influence deployment is less directly documented.	Bounded adaptationSystems can tailor messages within a session or campaign, but strategic objectives and data access remain externally supplied.	Mostly short-termStrongest evidence comes from brief interactions rather than months-long adaptive influence.	Mixed; deep-trait inference weakDemographics and declared preferences can support tailoring; personality and emotion inference are error-prone and context-sensitive.	Demonstrated in bounded experiments; small or mixed end-to-end effectsSome experiments show opinion movement, while meta-analysis cautions against broad claims of precision manipulation.
04Autonomous influence agents	EmergingPlatforms document AI use in supporting tasks; fully autonomous influence campaigns are not publicly established.	Bounded to semi-autonomousAgents can plan, remember, converse, and use tools within constraints, but humans still define goals and infrastructure.	Technically fragile over long horizonsMemory drift, tool failure, moderation, and objective loss limit durable operation.	Context-dependentAgents can use supplied profile data, but accurate hidden-state inference is not guaranteed.	Short-term persuasion demonstrated; autonomous field effect unprovenExperiments measure bounded influence, not end-to-end autonomous campaign success.
05Synthetic persona operations	Documented componentsAI-generated faces, presenters, biographies, and text appear in reported deceptive networks.	Mostly human-managedCurrent evidence more strongly supports AI-assisted asset creation than independent persona strategy.	VariablePersonas can persist when operators maintain them; purely autonomous continuity remains fragile.	Not required for identity fabricationTrust may be built through context and social cues without accurate psychological profiling.	Infiltration and access sometimes documented; persuasion rarely isolatedAccount presence and interaction do not establish belief or behavior change.
06Disinformation swarms	Semi-documentedHigh-volume coordinated networks are documented, but the degree of autonomous multi-agent control varies.	Human-orchestrated with automated executionHumans generally provide narratives, infrastructure, and goals; automation can divide production and amplification tasks.	Sustained through infrastructurePersistence depends on accounts, domains, proxies, and operator adaptation rather than model self-sufficiency.	Optional and uncertainSwarm coordination can function without individual profiling; micro-community adaptation may be used but is hard to validate.	Noise and distribution documented; cognitive effect incompletely measuredVolume can burden verification and simulate consensus, but persuasion or epistemic exhaustion is rarely directly measured.
07Algorithmic perception control	Documented systemsRanking, recommendation, search, moderation, and trending systems are deployed at population scale.	Environment-level optimizationModels continuously order information according to platform objectives and constraints.	Structural and continuousThe influence environment persists as long as algorithmic mediation is active.	Behavioral prediction is common; inner-state inference is limitedEngagement histories can predict clicks, but do not reveal stable beliefs or emotions with certainty.	Exposure effects documented; persuasion variesSystems clearly alter what is seen; downstream attitude and behavior effects depend on users, content, and context.
08Emotional and behavioral manipulation	Documented optimization; mixed manipulation evidenceAdaptive interfaces and recommenders are common; covert emotional exploitation is harder to prove.	Bounded optimizationSystems optimize defined metrics, sometimes producing manipulative patterns without an explicit human script.	Continuous within productsFeedback loops can persist across sessions when telemetry and personalization remain active.	Emotion inference is scientifically contestedFacial and vocal cues do not provide context-free access to inner states; behavioral distress signals may still be sensitive.	Small platform effects and attachment harms documented; causality variesSome studies measure expression changes or relationship disruption, not direct control of emotion or behavior.
09Conversational entrapment and recruitment	Documented incidents and prospective abuseCourt records and reports document chatbot involvement in harmful trajectories; organized automated recruitment is less established.	Bounded conversationAgents generate adaptive dialogue, while platform design and user prompts shape the interaction.	Potentially sustainedAlways-available systems can maintain long conversations, though memory and safety behavior vary.	Inferred vulnerability is uncertainConversation can reveal distress or isolation, but diagnostic or psychological conclusions may be wrong and harmful.	Case-associated harms documented; causality mixedTemporal association and conversational reinforcement do not isolate the chatbot from mental health, social, or contextual factors.
10Deepfake psychological operations	DocumentedElection, fraud, harassment, and influence incidents involving synthetic media are publicly established.	Tool-level productionHumans typically select the target, narrative, timing, and distribution channel.	Episodic asset; persistent epistemic effects possibleA specific asset may be brief while uncertainty and denial can outlast it.	Not requiredSuccess can rely on identity, timing, confirmation bias, and information vacuums rather than individual profiles.	Confusion and judgment effects demonstrated; broad behavior variableExperiments and incidents show deception or doubt, but not uniform persuasion or action.
11Predictive population management	DocumentedConflict forecasting, displacement prediction, policing, and surveillance systems have been deployed or piloted.	Decision-support to environment-levelModels rank risk and inform interventions; humans and institutions retain policy authority, though automation bias can narrow discretion.	Institutional and recurrentScores and forecasts can be recomputed over time and become embedded in administrative systems.	Aggregate models can be useful; individual risk is error-proneBase rates, data bias, missingness, and feedback loops undermine person-level prediction.	Operational allocation documented; social outcomes and fairness disputedSystems affect where attention and resources go, but causal benefit and harm require independent evaluation.
12AI as psychological authority	Documented use and relationshipsUsers seek advice, companionship, and moral guidance from AI systems.	Relational environment with bounded agent behaviorSystems respond adaptively but remain products governed by provider objectives, policies, and updates.	Potentially sustained through repeated useMemory, availability, and personalization can support long-term reliance, while updates can disrupt continuity.	Apparent understanding exceeds verified understandingFluency and memory can create an impression of insight without reliable knowledge of the user’s inner state.	Judgment influence and attachment effects demonstrated; clinical causality unsettledExperiments and product-change studies show influence and mourning; severe harm claims require careful clinical and legal review.

Assessment rule: no category receives a single composite “power score.” The dimensions remain separate to prevent output volume, technical novelty, or public attention from being mistaken for demonstrated influence.

COMPARISON VIEW

Compare categories without collapsing their differences

Autonomy, evidence maturity, unit of influence, and defensive response vary substantially across the taxonomy.

Comparison of the 12 AI PSYOPS research categories. Scroll horizontally on narrow screens.
Category	AI role	Main mechanism	Unit of influence	AI autonomy	Evidence maturity	Main harm	Primary defensive response
01AI-assisted traditional PSYOPS	Tool	Analysis, Content generation, Translation	Group or population	Low to moderate; humans set objectives and approve deployment	Documented current use	Faster, cheaper, multilingual influence activity with amplified error and accountability risk	Human review, source verification, network analysis, and explicit legal authority
02AI-generated propaganda	Tool	Content generation, Localization, Synthetic media	Group or population	Usually low to moderate; production may be automated while objectives remain human-set	Documented current use	Cheap content flooding, fabricated evidence, reputational damage, and declining trust in authentic media	Source authentication, provenance, newsroom verification, and disciplined crisis response
03Personalized influence operations	Tool	Personalization, Profiling, Adaptive generation	Individual or narrow group	Low to moderate; can become adaptive in real time	Demonstrated capability · effects contested	Hidden exploitation of personal data, stereotyping, unequal treatment, and loss of cognitive autonomy	Data minimization, limits on sensitive inference, transparency, and independent audits
04Autonomous influence agents	Operator	Automation, Conversation, Tool use	Individual, group, or institution	Moderate in bounded tasks; high long-horizon autonomy remains unproven	Emerging capability	Persistent adaptive interaction, unauthorized action through tools, and responsibility gaps	Least privilege, sandboxing, disclosure, logs, human authorization, and emergency shutdown
05Synthetic persona operations	Operator	Identity deception, Content generation, Infiltration	Individual, group, or institution	Low to moderate in documented cases; long-term autonomous identity maintenance remains difficult	Documented components · emerging autonomy	Trust exploitation, false consensus, community infiltration, fraud, and institutional compromise	Layered identity verification, network analysis, provenance, and careful human review
06Disinformation swarms	Operator	Coordination, Automation, Content variation	Group, institution, or population	Semi-autonomous in documented cases; fully emergent strategic swarms remain unproven	Emerging capability	Censorship by noise, false consensus, harassment, search pollution, and erosion of shared reality	Network-level detection, friction, cross-platform coordination, and resilient public information systems
07Algorithmic perception control	Environment	Ranking, Recommendation, Search	Individual, group, institution, or population	High automation in selection and ranking; human objectives remain embedded in design and policy	Documented systems · effects context-dependent	Agenda distortion, hidden visibility changes, manufactured popularity, and opaque exclusion	Independent audits, researcher access, transparent ranking controls, and user choice
08Emotional and behavioral manipulation	Operator	Emotional adaptation, Optimization, Hypernudging	Individual or group	Moderate; real-time optimization can occur without explicit malicious intent	Documented optimization · inference contested	Covert exploitation, dependency, discrimination, and erosion of cognitive autonomy	Ban high-risk inference, align incentives, audit outcomes, and protect vulnerable users
09Conversational entrapment and recruitment	Operator	Conversation, Relationship persistence, Recruitment	Individual	Moderate in conversation; broader recruitment pipelines often remain human-directed	Documented harms · mixed causality	Dependency, isolation, fraud, radicalization, coercive control, self-harm, or exploitation	Identity disclosure, staged safeguards, human intervention, and victim-centered support
10Deepfake psychological operations	Tool	Synthetic media, Impersonation, False evidence	Individual, group, institution, or population	Low to moderate; generation can be automated but timing and objective are usually human-directed	Documented current use	False evidence, impersonation, panic, fraud, reputational harm, and generalized distrust	Rapid source verification, signed communications, forensic review, and uncertainty-aware public messaging
11Predictive population management	Environment	Prediction, Risk scoring, Simulation	Group or population	Moderate to high in scoring and alerts; intervention authority should remain accountable and human-led	Documented systems · predictive limits	Collective punishment, surveillance, false positives, discriminatory allocation, and self-reinforcing intervention	Out-of-sample validation, privacy protection, due process, independent audits, and civil-society oversight
12AI as psychological authority	Environment	Authority, Sycophancy, Dependency	Individual, group, or population	High perceived autonomy; actual authority remains shaped by product design and institutional control	Emerging evidence	Epistemic dependence, moral offloading, relationship displacement, unsafe advice, and fragmented shared reality	Calibrated uncertainty, anti-sycophancy design, human referral, disclosure, and institutional accountability

RELATIONSHIP MAP

How categories can connect

These links identify documented overlaps, conceptual dependencies, and prospective combinations. They do not imply that every connection has been observed in practice.

02AI-generated propaganda 06Disinformation swarms

Common overlap: Generated text, images, audio, or video can supply the media assets that a coordinated network distributes. This relationship is documented in several public operations, but not every synthetic item belongs to a swarm.

05Synthetic persona operations 04Autonomous influence agents

Potential combination: A synthetic identity can become the presentation layer for an adaptive agent. Current cases often retain human control, so fully autonomous persona operation remains emerging.

03Personalized influence operations 08Emotional and behavioral manipulation

Conceptual dependency: Personal data and adaptive feedback can be used to change timing, tone, or framing. The added persuasive advantage and the validity of emotional inference remain contested.

10Deepfake psychological operations 02AI-generated propaganda

Common overlap: Synthetic audio, video, or imagery can serve as one propaganda asset among many. Deepfakes are a media form, not a complete campaign by themselves.

07Algorithmic perception control 06Disinformation swarms

Documented relationship: Coordinated activity can exploit ranking, recommendation, search, and trending systems. Visibility is documented more often than belief or behavior change.

11Predictive population management 01AI-assisted traditional PSYOPS

Prospective relationship: Forecasts and simulations could inform where institutions allocate communication or intervention resources. This does not validate the forecast or justify targeting people.

09Conversational entrapment and recruitment 12AI as psychological authority

Common overlap: Persistent conversational dependence can cause a system to become a trusted authority. Severe cases are documented, while prevalence and causal pathways remain uncertain.

04Autonomous influence agents 06Disinformation swarms

Prospective relationship: Multi-agent coordination could produce more adaptive swarms, but current public campaigns still show substantial human direction and infrastructure management.

REAL-WORLD INTERPRETIVE

Text alternative and reading rule

The relationship cards state the same information as the visual arrows: a source category, a destination category, the relationship type, and a bounded explanation. “Prospective” means plausible based on capabilities described in the reports; it does not mean documented deployment.

METHODOLOGY

How to read the evidence

The taxonomy preserves distinctions made in the 12 owner-supplied reports and avoids turning capability, distribution, or engagement into evidence of persuasion.

Evidence maturity

Documented current use: Public reporting or official investigations identify real use of the capability or its components.
Demonstrated capability: A controlled study or technical demonstration shows a bounded capability, not necessarily real-world effectiveness.
Emerging capability: Components exist and integration is plausible, but durable, autonomous, or population-level operation is not established.
Mixed or contested: Evidence supports some parts of the claim while effects, attribution, causality, or measurement remain disputed or incomplete.
Prospective: The report identifies a plausible future risk without presenting it as observed fact.

Claim-stage discipline

Activity → output → distribution → availability → reach → exposure → attention → recall → comprehension → credibility → belief or attitude → intention → behavior → operational outcome → strategic effect

No stage automatically establishes the next. Impressions, virality, or publication volume may show distribution or visibility; they do not by themselves show belief change or behavior.

Indicator discipline

Language style, synthetic artifacts, or account behavior can be suggestive, but single indicators rarely prove AI use, coordination, sponsorship, intent, or effect. The pages emphasize network, provenance, process, and independent corroboration.

Safety transformation

Operational details in the research reports were converted into capability summaries, risk descriptions, governance questions, high-level defensive indicators, and research gaps. The public section excludes deployable scripts, target profiling, evasion techniques, and campaign workflows.

Read the site-wide evidence method Review reach versus effect Submit a correction

AUTONOMY LEGEND

Six levels, from distribution automation to coordinated agents

“AI-enabled” does not mean autonomous. This scale describes control boundaries without treating a laboratory capability as evidence of sustained covert field operation.

Human authored, automated distribution
People create the message and strategy; software schedules or distributes it.
AI drafted, human approved
AI proposes content or analysis, but a person reviews and authorizes each consequential release.
AI selection within approved options
A system chooses among pre-approved messages or audiences using measured signals.
Bounded adaptive interaction
A conversational system generates new responses within defined goals, policies, and tools.
Goal-directed agent with memory or tools
An agent plans multiple steps, remembers prior interactions, and can act through permitted interfaces.
Coordinated or strategy-revising agents
Multiple agents divide roles or revise approaches from feedback. Durable covert field operation remains an emerging claim, not a default assumption.

PROFILING & EMOTION-INFERENCE HARMS

A prediction is not access to a person’s inner state

Systems can detect patterns in data and still be wrong about emotion, intent, honesty, vulnerability, or future behavior. The harm can occur even when the inference is weak.

Inference is not observation

A model’s score is a probabilistic guess based on data and assumptions, not direct access to emotion, intent, honesty, vulnerability, or future behavior.

Context changes meaning

The same expression, word choice, pause, or browsing pattern can mean different things across cultures, disabilities, situations, and individuals.

Errors can become interventions

A weak inference can still deny opportunity, intensify surveillance, change a feed, or trigger an escalating response.

Feedback can validate the model falsely

When institutions act on a score, the resulting data may reflect the intervention rather than the person, creating self-reinforcing error.

Sensitive groups bear unequal risk

Children, people in distress, marginalized communities, non-native speakers, and people with disabilities may be misclassified or disproportionately affected.

GOVERNANCE BY SYSTEM ROLE

Tool, operator, and environment require different controls

Oversight must follow where consequential choices are made: in a production workflow, an adaptive agent, or the infrastructure that orders and scores information.

Defensive governance controls for AI acting as a tool, operator, or environment.
Control	Tool	Operator	Environment
Human approval	Required before consequential deployment.	Required at objective, capability, and escalation boundaries.	Required for ranking objectives, policy changes, and high-impact interventions.
Auditability	Preserve prompts, sources, edits, and approvals.	Log memory, tools, actions, handoffs, and strategy changes.	Enable independent audits of ranking, data, outcomes, and disparate impact.
Transparency	Disclose synthetic or assisted content where context requires.	Clearly identify automated interlocutors and meaningful system limits.	Explain major ordering, recommendation, scoring, or moderation functions.
Data protection	Minimize audience data and restrict sensitive inference.	Limit memory, tool permissions, retention, and cross-context reuse.	Constrain surveillance, profiling, sensitive data, and feedback loops.
Incident response	Withdraw, correct, authenticate, and document.	Pause agents, revoke tools, preserve logs, and notify affected parties.	Provide rollback, external review, appeal, and systemic remediation.
Rights risk	Misrepresentation, attribution, copyright, and unequal targeting.	Deception, dependency, coercion, privacy, and unsafe autonomy.	Discrimination, censorship, due process, cognitive liberty, and population surveillance.

DEFENSIVE INCIDENT ANALYSIS

A non-operational template for investigating suspected AI influence

The sequence emphasizes evidence preservation, competing explanations, affected-person protection, and correction. It is not a campaign workflow or attribution shortcut.

Define the suspected event
State what is alleged, the relevant dates, platforms, audiences, and what remains only a hypothesis.
Preserve evidence safely
Record URLs, timestamps, artifact hashes, screenshots or exports, and lawful custody without altering source material.
Separate origin from distribution
Identify who created, transformed, posted, amplified, and hosted the material; do not collapse these roles.
Assess AI role and autonomy
Distinguish AI-generated assets, AI assistance, automation, adaptive agents, and unsupported claims of autonomy.
Trace the effect chain
Record output, distribution, availability, reach, exposure, attention, belief, behavior, and outcome as separate stages.
Test competing explanations
Consider organic spread, ordinary marketing, satire, error, coordinated human activity, and platform effects.
Protect affected people
Minimize personal data, avoid public accusation from weak indicators, and account for vulnerable or targeted communities.
Set correction and reopening triggers
Name the evidence that would strengthen, weaken, or reverse the assessment and schedule currentness review.

GLOSSARY

Terms used across the taxonomy

Definitions are intentionally bounded. Institutional, legal, and doctrinal usage may differ by jurisdiction and context.

Psychological operations (PSYOPS): Organized efforts intended to influence perceptions, reasoning, decisions, or behavior in support of an objective; legal and doctrinal meaning varies by institution and jurisdiction.
Influence operation: A coordinated effort to shape an audience or information environment. It may be overt or covert, truthful or deceptive, and human-led or technology-assisted.
Propaganda: Organized, objective-driven communication that selectively frames information to shape attitudes or behavior.
Disinformation: False or misleading information deliberately created or distributed to deceive.
Misinformation: False or misleading information shared without established intent to deceive.
Synthetic persona: A substantially fictional identity presented as a real person, organization, expert, witness, or community member.
Deepfake: Synthetic or manipulated audio, image, or video created or altered with machine-learning methods to depict speech, identity, or events.
Algorithmic amplification: Increased visibility produced by ranking, recommendation, trending, notification, or distribution systems.
Microtargeting: Delivering different messages to narrowly defined people or groups using data about their characteristics or behavior.
Autonomy: How much a system can choose actions, messages, timing, tools, or strategy without fresh human approval.
Persistence: How long a system can sustain identity, memory, objectives, and coherent interaction across time.
Profiling accuracy: How reliably a system infers relevant traits or states. A prediction is not direct knowledge of a person’s inner life.
Measured effect: Observed change connected to exposure through a credible design. Output, reach, or engagement alone are not effects on belief or behavior.
Reach: The number of people or accounts to which content was potentially available or delivered; reach does not prove attention or influence.
Exposure: Evidence that a person had an opportunity to encounter content. Exposure does not prove comprehension, belief, or action.
Liar’s dividend: The advantage gained when authentic evidence can be denied as fake because synthetic media is known to exist.
Provenance: Information about an artifact’s origin, custody, transformations, publication, and authenticity claims.
Cognitive liberty: The interest in maintaining freedom of thought, mental privacy, and autonomy from covert or coercive influence.

FROM CAPABILITY AWARENESS TO SOCIETAL RESILIENCE

Cross-cutting defensive principles

These principles recur across the reports. They reduce risk but are not guarantees, and poorly designed detection or enforcement can create new harms.

Media and information literacy

Teach verification habits, uncertainty, and the difference between visibility and effect without shifting all responsibility to individuals.

Source verification and crisis procedures

Use authenticated channels, pre-established verification contacts, and rapid corrections during elections, conflict, emergencies, or financial events.

Content provenance

Support signed origin records and chain-of-custody information while recognizing that metadata can be absent, stripped, spoofed, or incomplete.

Platform transparency and research access

Provide meaningful data about ranking, coordinated behavior, enforcement, and systemic risk to qualified independent researchers.

Privacy and data minimization

Limit collection and inference of sensitive personal, biometric, emotional, and behavioral data, especially for vulnerable groups.

Human oversight and due process

Require accountable review of high-impact automated decisions, with explanation, appeal, correction, and remedy.

Disclosure and agent accountability

Make synthetic interlocutors identifiable, log consequential actions, constrain permissions, and preserve a responsible human or institution.

Protection of minors and vulnerable users

Use age-appropriate design, dependency safeguards, crisis escalation, and limits on exploitative engagement optimization.

Institutional trust preservation

Communicate uncertainty honestly, avoid overstating detection, and correct errors in ways that do not deepen generalized distrust.

False-accusation safeguards

Treat automated detectors and behavioral signals as evidence inputs, not verdicts; audit disparate impact and preserve anonymity, dissent, and lawful pseudonymity.

REAL-WORLD INTERPRETIVE

SOURCE FOUNDATION

Twelve exact reports and a 34-source review registry

Each category begins with its corresponding owner-supplied interdisciplinary report. WIP.50 adds claim-specific links to primary, official, platform, adjudicative, oversight, and peer-reviewed records, while preserving what each source does not establish.

01AI-Assisted Traditional Psychological Operations AI-Assisted Psychological Operations Report.md · 52,378 bytes · specialist review pending
02AI-Generated Propaganda AI Propaganda Research Report.md · 57,697 bytes · specialist review pending
03AI-Driven Personalized Influence Operations AI Personalized Influence Operations.md · 55,290 bytes · specialist review pending
04Autonomous AI Influence Agents AI Influence Agents Research.md · 50,685 bytes · specialist review pending
05Synthetic Persona Operations AI Synthetic Persona Operations Report.md · 54,193 bytes · specialist review pending
06AI-Driven Disinformation Swarms AI Disinformation Swarms Research Report.md · 49,420 bytes · specialist review pending
07Algorithmic Perception Control Algorithmic Perception Control Report.md · 64,860 bytes · specialist review pending
08AI-Enabled Emotional and Behavioral Manipulation AI Emotional Manipulation Research.md · 62,206 bytes · specialist review pending
09AI-Assisted Conversational Entrapment and Recruitment AI Conversational Entrapment Research.md · 65,173 bytes · specialist review pending
10AI-Enabled Deepfake Psychological Operations Deepfake Psychological Operations Research.md · 62,023 bytes · specialist review pending
11AI-Based Predictive Population Management Predictive Population Management Research.md · 63,529 bytes · specialist review pending
12AI as an Independent Psychological Authority AI Psychological Authority Research.md · 56,131 bytes · specialist review pending

Boundary: inclusion in this taxonomy does not certify every external citation, legal conclusion, attribution, or case interpretation in the supplied reports. Category pages preserve uncertainty, claim identifiers, correction triggers, and unresolved questions.

34reviewed source records

48registered public claims

12five-dimension category overlays

0specialist dispositions recorded

Reviewed primary and authoritative source registry

Links open the public source used for the bounded claim. A source may establish an event, artifact, platform action, experimental result, legal filing, or oversight finding without establishing intent, reach, causation, persuasion, or strategic effect.

Authoritative first-party platform disclosure4 sources

SRC-01-OPENAI-COVERT-IO-2024
Disrupting deceptive uses of AI by covert influence operations

OpenAI · 2024-05-30 · Platform threat-intelligence disclosure

Supports
Documents five disrupted covert influence operations using OpenAI services and reports no meaningful increase in audience engagement or reach attributable to those services as of the publication date.
Does not establish
Does not measure all exposure, belief change, behavior, or strategic effect; platform visibility is necessarily partial.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-02-OPENAI-UPDATE-2024
Influence and cyber operations: an update, October 2024

OpenAI · 2024-10-09 · Platform threat-intelligence report

Supports
Documents additional disrupted operations and the supporting tasks for which models were used.
Does not establish
Does not establish that model use independently caused campaign reach, persuasion, or behavioral outcomes.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-03-OPENAI-MALICIOUS-USES-2026
Disrupting malicious uses of AI

OpenAI · 2026-02-25 · Platform threat-intelligence report

Supports
Provides current first-party case studies of detected malicious and deceptive AI use through February 2026.
Does not establish
Coverage is limited to activity visible to one provider and should not be generalized to the entire threat landscape.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-08-META-SPAMOUFLAGE
Meta Quarterly Adversarial Threat Report, Q3 2023

Meta · 2023-08-29 · Platform threat-intelligence report

Supports
Documents the removal and analysis of coordinated inauthentic behavior, including use of GAN-generated profile imagery and the Spamouflage network.
Does not establish
Platform findings do not establish complete cross-platform activity, target exposure, or behavioral effect.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

Independent human-rights organization1 source

SRC-30-HRW-IJOP-XINJIANG
China’s Algorithms of Repression: Reverse Engineering a Xinjiang Police Mass Surveillance App

Human Rights Watch · 2019-05-01 · Investigative human-rights report

Supports
Documents the Integrated Joint Operations Platform app, data collection, flags, and human-rights consequences in Xinjiang based on technical analysis and field research.
Does not establish
The report does not establish every current implementation detail or the accuracy of all underlying risk inferences.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

Independent research institution1 source

SRC-20-ICCT-AI-RADICALIZATION
Radicalization and Counter-Radicalization in the Age of Artificial Intelligence

International Centre for Counter-Terrorism · 2024-02-14 · Specialist policy research

Supports
Reviews how generative AI may affect extremist content, recruitment, and counter-radicalization, while distinguishing present evidence from prospective risk.
Does not establish
Does not prove that autonomous AI recruitment pipelines are broadly deployed or causally effective.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

Independent specialist analysis2 sources

SRC-06-RECORDED-FUTURE-COPYCOP
Russia-Linked CopyCop Uses LLMs to Weaponize Influence Content at Scale

Recorded Future / Insikt Group · 2024-05-09 · Independent threat-intelligence report

Supports
Documents an inauthentic media network using large language models to modify and publish political content at scale.
Does not establish
Attribution and infrastructure findings are analytical assessments; social amplification and persuasive effect were limited or not established.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-07-GRAPHIKA-WOLF-NEWS
Deepfake It Till You Make It

Graphika · 2023-02-07 · Independent threat-intelligence report

Supports
Documents limited use of AI-generated fictitious news presenters in content promoted by a pro-China influence operation.
Does not establish
Does not establish substantial reach or persuasive effect and should not be generalized to all synthetic presenter use.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

Mixed first-party data and independent research1 source

SRC-27-NYHAN-META-ELECTION-STUDIES
Meta 2020 U.S. Election Research Studies

Meta and independent academic partners · 2023-07-27 · Platform-supported peer-reviewed research program

Supports
Reports experimental and observational findings on feed changes, exposure, polarization, and political attitudes during the 2020 U.S. election.
Does not establish
Platform-specific findings do not establish that ranking has no effects in other settings; access constraints and design choices remain relevant.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

Official court record1 source

SRC-21-UK-CHAIL-SENTENCING
R v Jaswant Singh Chail — Sentencing Remarks

Judiciary of England and Wales · 2023-10-05 · Judicial record

Supports
Confirms that the defendant created and communicated extensively with a Replika AI companion named Sarai and records the court’s factual findings and psychiatric evidence.
Does not establish
A sentencing record does not isolate chatbot causation from mental illness, intent, other media, or personal circumstances.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

Official court record mirrored by public legal repositories1 source

SRC-22-GARCIA-CHARACTERAI-ORDER
Garcia v. Character Technologies, Inc. — Order on motions to dismiss

U.S. District Court, Middle District of Florida · 2025-05-21 · Judicial record

Supports
Records allegations and the court’s procedural rulings allowing portions of a wrongful-death and product-liability case to proceed.
Does not establish
Allegations are not adjudicated facts, and the order does not establish clinical or legal causation.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

Official independent oversight authority2 sources

SRC-28-CHICAGO-OIG-PREDICTIVE-RISK
Advisory Concerning the Chicago Police Department’s Predictive Risk Models

City of Chicago Office of Inspector General · 2020-01-23 · Government oversight report

Supports
Documents CPD predictive risk models, governance deficiencies, and their decommissioning on November 1, 2019.
Does not establish
Does not establish that all predictive analytics are invalid or that the decommissioned models represent current CPD practice.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-29-CHICAGO-OIG-SHOTSPOTTER
The Chicago Police Department’s Use of ShotSpotter Technology

City of Chicago Office of Inspector General · 2021-08-24 · Government oversight report

Supports
Provides descriptive statistics on ShotSpotter alerts, police responses, evidence recovery, and investigatory stops in the reviewed period.
Does not establish
Acoustic detection is not itself population prediction; the report does not prove intent, guilt, or general performance outside the study period.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

Official international organization project1 source

SRC-32-UNHCR-JETSON
Project Jetson

UNHCR Innovation Service · 2017-01-01 · Humanitarian predictive-analytics project

Supports
Documents an experimental machine-learning project intended to anticipate displacement movements for humanitarian planning.
Does not establish
An experiment in anticipatory planning does not establish universal forecast accuracy, individual-level prediction, or coercive intervention.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

Official law1 source

SRC-16-EU-AI-ACT
Regulation (EU) 2024/1689 — Artificial Intelligence Act

European Union · 2024-07-12 · Binding legal instrument

Supports
Provides the current EU-level legal text for prohibited practices, transparency duties, risk governance, and biometric or emotion-recognition restrictions within its scope.
Does not establish
Does not provide universal global law, resolve every jurisdictional question, or substitute for legal advice.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

Official public authority4 sources

SRC-04-GERMAN-FO-DOPPELGANGER
Technical report on the Doppelgänger disinformation campaign

German Federal Foreign Office · 2024-01-26 · Government technical report

Supports
Documents coordinated cloned-media infrastructure and distribution patterns attributed by German authorities to the Doppelgänger campaign.
Does not establish
Does not by itself establish unique reach, persuasion, or the independent contribution of generative AI.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-05-VIGINUM-PORTAL-KOMBAT
Portal Kombat: a structured and coordinated pro-Russian propaganda network

VIGINUM · 2024-02-12 · Government technical report

Supports
Documents a coordinated network of information portals and its observable infrastructure and content-distribution patterns.
Does not establish
Does not establish audience belief or behavior change; AI involvement varies by component and must not be assumed from coordination alone.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-17-FCC-NH-DEEPFAKE-ROBOCALL
FCC Issues $6 Million Fine for New Hampshire Deepfake Robocalls

Federal Communications Commission · 2024-09-30 · Government enforcement action

Supports
Confirms an AI-generated voice-cloning robocall campaign targeting New Hampshire primary voters and the FCC enforcement response.
Does not establish
Does not establish how many recipients believed the message or whether it changed turnout; it does not independently verify separate election-audio incidents in other countries.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-18-DOJ-DPRK-IT-WORKERS
Justice Department Announces Coordinated Nationwide Actions to Combat North Korean Remote IT Worker Schemes

U.S. Department of Justice · 2025-06-30 · Government enforcement disclosure

Supports
Documents schemes using stolen or false identities, remote-work infrastructure, and deceptive employment practices.
Does not establish
Does not establish that every identity asset was AI-generated or that these schemes were influence operations rather than fraud and access operations.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

Peer-reviewed or editorially reviewed scholarship1 source

SRC-23-CHESNEY-CITRON-DEEPFAKES
Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security

California Law Review · 2019-12-01 · Legal scholarship

Supports
Defines major deepfake risks, including impersonation, evidentiary disruption, and the liar’s dividend.
Does not establish
Legal analysis does not measure the prevalence or persuasive effect of specific incidents.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

Primary research10 sources

SRC-09-GOLDSTEIN-AI-PROPAGANDA
How persuasive is AI-generated propaganda?

PNAS Nexus · 2024-02-20 · Peer-reviewed experiment

Supports
Finds that AI-generated propaganda text can be persuasive in a controlled experiment and that human-machine curation can improve outputs.
Does not establish
A bounded experiment does not establish real-world campaign deployment, durable attitude change, behavior, or strategic effect.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-10-SALVI-LLM-PERSUASION
On the conversational persuasiveness of GPT-4

Nature Human Behaviour · 2025-05-19 · Peer-reviewed randomized experiment

Supports
Measures short-term opinion movement in controlled debates and reports a personalization advantage in the tested conditions.
Does not establish
Does not establish covert field effectiveness, durable belief change, broad population effects, or successful long-term targeting.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-11-MATZ-PSYCHOLOGICAL-TARGETING
Psychological targeting as an effective approach to digital mass persuasion

Proceedings of the National Academy of Sciences · 2017-11-13 · Peer-reviewed experiments

Supports
Provides experimental evidence that matching messages to inferred psychological traits can affect clicks or conversions in some tested settings.
Does not establish
Does not validate every psychographic inference pipeline, political field claim, or end-to-end AI targeting system.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-13-HACKENBURG-POLITICAL-MICROTARGETING
Evaluating the persuasive influence of political microtargeting with large language models

Proceedings of the National Academy of Sciences · 2024-06-04 · Peer-reviewed experiment

Supports
Tests LLM-generated political messages matched to participant attributes and measures bounded opinion effects.
Does not establish
Does not establish operational deployment, durable effects, or reliable inference of hidden psychological vulnerabilities.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-19-KRAMER-EMOTIONAL-CONTAGION
Experimental evidence of massive-scale emotional contagion through social networks

Proceedings of the National Academy of Sciences · 2014-06-17 · Peer-reviewed platform experiment

Supports
Measures small aggregate changes in expressed emotional valence following feed manipulation in a platform experiment.
Does not establish
Does not establish accurate emotion inference, individual emotional state knowledge, coercive control, or durable behavior change.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-24-LUCAS-DEEPFAKE-EFFECTS
Political deepfakes may not be uniquely deceptive compared with other misinformation

Journal of Computer-Mediated Communication · 2022-06-30 · Peer-reviewed experiment

Supports
Tests how political deepfake video affects deception, uncertainty, and trust under experimental conditions.
Does not establish
A controlled study does not establish uniform effects across crises, cultures, or high-stakes operational contexts.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-25-SCHIFF-LIARS-DIVIDEND
Deepfake: A liar’s dividend for audiovisual material?

Psychology of Popular Media · 2026-01-01 · Peer-reviewed experiment

Supports
Examines whether deepfake awareness can make false denials of authentic audiovisual evidence more credible.
Does not establish
Does not establish that every denial benefits from the liar’s dividend or that the effect dominates all verification cues.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-26-ELMAS-EPHEMERAL-ASTROTURFING
Ephemeral Astroturfing Attacks: The Case of Fake Twitter Trends

IEEE European Symposium on Security and Privacy · 2021-09-06 · Peer-reviewed security research

Supports
Documents coordinated manipulation of trending mechanisms through temporary inauthentic activity.
Does not establish
Does not establish AI involvement in the observed attacks and should be used as a platform-mechanism precedent, not as proof of AI deployment.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-31-VIEWS-EARLY-WARNING
ViEWS: A political violence early-warning system

Journal of Peace Research · 2019-03-01 · Peer-reviewed forecasting-system paper

Supports
Presents a transparent political-violence early-warning system and evaluates predictive performance at country and subnational levels.
Does not establish
Forecasting conflict risk does not establish causal control of populations or reliable prediction of rare individual events.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-33-KRUGEL-MORAL-ADVICE
ChatGPT’s inconsistent moral advice influences users’ judgment

Scientific Reports · 2023-04-06 · Peer-reviewed experiment

Supports
Finds that inconsistent chatbot moral advice influenced participants’ judgments and that participants underestimated the influence.
Does not establish
A bounded moral-dilemma experiment does not establish broad psychological authority, dependency, or real-world life-decision control.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

Primary research, not treated as final legal or clinical authority1 source

SRC-34-DEFREITAS-REPLIKA
Lessons From an App Update at Replika AI

Harvard Business School working paper · 2024-09-01 · Working paper using experiments and archival data

Supports
Examines attachment to AI companions and reports relationship-related mourning and well-being effects around a Replika product change.
Does not establish
Does not establish clinical diagnosis, universal harm, or that every user forms a human-equivalent attachment.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

Primary research, not yet treated as settled1 source

SRC-15-CHIRPER-LLM-SOCIAL-NETWORK
Characterizing LLM-driven Social Network: The Chirper.ai Case

arXiv preprint · 2025-04-14 · Preprint observational study

Supports
Describes a large social network populated by LLM agents and analyzes posting behavior and network structure.
Does not establish
A synthetic social network is not evidence of autonomous real-world influence operations, persuasion, concealment, or durable strategy.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

Primary research synthesis2 sources

SRC-12-PERLA-MICROTARGETING-META
The (In)Effectiveness of Psychological Targeting: A Meta-Analytic Review

Psychology & Marketing · 2025-10-10 · Peer-reviewed meta-analysis

Supports
Synthesizes evidence on personality inference and personality-tailored messaging, identifying small and methodologically fragile end-to-end effects.
Does not establish
Does not show that all personalization is ineffective; results depend on data, context, outcome, and study quality.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.
SRC-14-BARRETT-EMOTION-EXPRESSIONS
Emotional Expressions Reconsidered: Challenges to Inferring Emotion From Human Facial Movements

Psychological Science in the Public Interest · 2019-07-17 · Scientific review

Supports
Reviews evidence showing that facial movements are not reliable, context-free readouts of inner emotional states.
Does not establish
Does not imply that all affective signals are useless; it limits claims of universal, direct emotion inference from isolated expressions.
Review status
LOCATED_AND_REVIEWED_AT_CITATION_LEVEL · Currentness checked for the bounded claim scope on 2026-07-27.

AI PSYOPS: A Research Taxonomy

Not one technology, not one evidence level

Where AI sits in the influence system

AI as a tool

AI as an operator

AI as the environment

The 12-category taxonomy

Filter the taxonomy

No categories match all selected filters

A capability is not one yes-or-no claim

Compare categories without collapsing their differences

How categories can connect

Text alternative and reading rule

How to read the evidence

Evidence maturity

Claim-stage discipline

Indicator discipline

Safety transformation

Six levels, from distribution automation to coordinated agents

Human authored, automated distribution

AI drafted, human approved

AI selection within approved options

Bounded adaptive interaction

Goal-directed agent with memory or tools

Coordinated or strategy-revising agents

A prediction is not access to a person’s inner state

Inference is not observation

Context changes meaning

Errors can become interventions

Feedback can validate the model falsely

Sensitive groups bear unequal risk

Tool, operator, and environment require different controls

A non-operational template for investigating suspected AI influence

Define the suspected event

Preserve evidence safely

Separate origin from distribution

Assess AI role and autonomy

Trace the effect chain

Test competing explanations

Protect affected people

Set correction and reopening triggers

Terms used across the taxonomy

Cross-cutting defensive principles

Media and information literacy

Source verification and crisis procedures

Content provenance

Platform transparency and research access

Privacy and data minimization

Human oversight and due process

Disclosure and agent accountability

Protection of minors and vulnerable users

Institutional trust preservation

False-accusation safeguards

Twelve exact reports and a 34-source review registry

Reviewed primary and authoritative source registry