Academic Research2024VERIFIED: 2026-09-30
Personalized AI persuasion: ChatGPT-generated messages outperform non-personalized ones
Scientific Reports (Matz et al.)
A study of 1,788 participants found personalized ChatGPT-generated messages were more influential than corresponding non-personalized versions across several tested contexts.
Why it matters: Provides direct experimental evidence that LLM personalization can increase persuasion in controlled settings.
Limitations: Lab setting; effect sizes modest; not a real-world influence operation.
Original source ↗Academic Research2023VERIFIED: 2026-09-30
Large language models can persuade humans on political issues
PNAS (Argyle et al.)
A study involving 8,587 participants found GPT-4-generated messages persuasive but did not find a statistically significant overall microtargeting advantage from additional personal information.
Why it matters: Suggests LLM persuasion is real but the added value of microtargeting with personal data is not clearly established.
Limitations: Self-reported attitudes; no behavioral follow-through; U.S. context.
Original source ↗Academic Research2025VERIFIED: 2026-09-30
Personalized LLM persuasion in debate
Nature Human Behaviour
An experiment found personalized GPT-4 debate conditions produced stronger persuasion than the human-human reference condition in that experimental setting.
Why it matters: Indicates personalization can amplify LLM persuasion relative to human persuaders under specific conditions.
Limitations: Constrained debate format; generalizability uncertain.
Original source ↗Academic Research2025VERIFIED: 2026-09-30
LLM-generated political-policy messages and attitude change
Nature Communications
A study of 4,829 participants found LLM-generated political-policy messages caused measurable attitude changes and were approximately comparable to ordinary human-generated messages.
Why it matters: Shows LLM messages are persuasive at parity with ordinary human messages, not necessarily superior.
Limitations: Attitude measures only; no behavioral outcomes.
Original source ↗NATOOngoingVERIFIED: 2026-09-30
Cognitive Warfare
NATO Allied Command Transformation
NATO ACT has led conceptual development of cognitive warfare as the battle for the mind and decision-making of targeted audiences.
Why it matters: Primary doctrinal source for the cognitive warfare concept.
Limitations: Conceptual; not a weapons specification.
Original source ↗European UnionOngoingVERIFIED: 2026-09-30
Foreign Information Manipulation and Interference (FIMI) Threat Analysis
European External Action Service
EEAS documents and analyzes state-aligned disinformation and interference patterns affecting the EU information environment.
Why it matters: Authoritative EU source on documented influence operations.
Limitations: Diplomatic framing; attribution caveats.
Original source ↗European Union2024VERIFIED: 2026-09-30
Regulation (EU) 2024/1689 (AI Act)
EUR-Lex
The EU AI Act, Article 5 of which prohibits certain AI practices involving subliminal or purposefully manipulative/deceptive techniques under specified conditions.
Why it matters: Primary legal source for manipulative-AI regulation.
Limitations: Legal text; not legal advice.
Original source ↗Media Provenance Standards2026VERIFIED: 2026-09-30
C2PA Content Credentials Specification
C2PA
The C2PA specification (v2.4 released April 2026) defines cryptographically verifiable content provenance.
Why it matters: Core technical standard for media provenance.
Limitations: Provenance does not prove factual truth.
Original source ↗AI Safety Laboratories2025–2026VERIFIED: 2026-09-30
Threat Intelligence Reports
Anthropic
Anthropic has published threat-intelligence reporting on influence operations including synthetic personas and targeting dossiers.
Why it matters: Primary industry source on observed AI-enabled influence infrastructure.
Limitations: Platform-specific; breakout often limited.
Original source ↗AI Safety Laboratories2024–2026VERIFIED: 2026-09-30
Threat Intelligence Reports
OpenAI
OpenAI has reported on state-aligned and commercial networks using its models for influence operations.
Why it matters: Primary industry source on AI-enabled influence operations.
Limitations: Platform-specific.
Original source ↗Threat IntelligenceOngoingVERIFIED: 2026-09-30
Threat Intelligence Center
Microsoft
Microsoft documents state-aligned threat actors and influence operations across digital channels.
Why it matters: Major cybersecurity research source.
Limitations: Corporate reporting.
Original source ↗