Intelligence notes

◉Dashboard↗

Generative AI

4 min read

Generative AI

Core Definition (BLUF)

Generative AI refers to artificial intelligence systems capable of producing novel content — text, images, audio, video, or code — in response to prompts, by sampling from learned probability distributions over training data. The dominant contemporary paradigm is the Large Language Model (LLM), trained via self-supervised learning on large text corpora (GPT, Claude, Gemini, LLaMA families). For intelligence and security analysis, generative AI is a dual-use technology: it dramatically amplifies both legitimate analytical productivity (synthesis, translation, summarization) and adversarial IO capability (Propaganda at scale, synthetic media generation, automated spearphishing, narrative manufacturing).

IO Relevance

Generative AI enables adversarial IO capabilities at new scale and cost points:

  • Content factory: A single operator can produce linguistically credible disinformation content in multiple languages, personalized to specific audiences, at near-zero marginal cost per piece
  • Synthetic media: Image and video generation enables visual Deepfakes without specialist technical skills
  • Automated targeting: LLM-powered spearphishing generates highly personalized social engineering messages from harvested social media context
  • Jailbreak exploitation: Domain-context attacks (see “Into the Gray Zone” paper) can extract harmful content from aligned models, enabling production of content models are explicitly trained to refuse

The adversarial IO threat from generative AI was empirically documented in Generative Propaganda (Daepp et al., 2025) and the attack surface was mapped in Into the Gray Zone (Hung et al., 2026).

Defense Relevance

Generative AI also enables counter-IO capabilities:

AI Governance and Safety Frameworks

The dual-use risk of generative AI has generated multiple governance frameworks operating at different levels:

FrameworkTypeBindingKey provisions
EU AI Act (2024)RegulationLegally binding (EU)Prohibited uses; high-risk classification; transparency for AI-generated content (Art. 50)
Anthropic RSPSelf-regulationVoluntaryASL tiered capability assessment; deployment restrictions at ASL-3+
OpenAI Preparedness FrameworkSelf-regulationVoluntaryRisk assessment matrix; deployment gates for catastrophic-risk categories
C2PA (Content Authenticity Initiative)Technical standardVoluntary adoptionCryptographic content provenance (creation chain metadata embedded in media files)
NIST AI RMFGovernment frameworkVoluntaryRisk management framework for AI deployment; adopted by U.S. federal agencies

Assessment (Medium): Existing governance is predominantly self-regulatory and voluntary. The EU AI Act is the most binding instrument but applies only to the EU market; state actors (Russia, China, Iran) operating adversarial GenAI capabilities are not subject to any of these frameworks. The governance gap between compliant commercial actors and adversarial state/non-state GenAI deployment is the primary vulnerability.

Attribution Challenges — AI-Assisted Content Detection

The proliferation of generative AI content creates detection and attribution challenges across the OSINT and intelligence analysis workflow:

  • Synthetic media detection: Image and video detectors have an adversarial dynamic — detection models are rapidly deployed, then circumvented by generative models trained to evade the detectors. Error rates at platforms are significant.
  • Text detection: LLM-generated text detectors (RADAR, Binoculars, GPTZero) have shown high false-positive rates on non-native English text, producing misattribution of human-written foreign-language content as AI-generated.
  • Provenance verification (C2PA): Cryptographic content provenance (Content Authenticity Initiative / C2PA standard) embeds creation-chain metadata in media files. Major camera manufacturers (Nikon, Sony, Leica) and platforms (Adobe, Microsoft) are implementing; adoption is partial.
  • Watermarking: Invisible watermarks in AI-generated content (Google SynthID, Meta watermarking) allow post-hoc detection by the generating platform but are stripped by simple transformations (screenshot, re-encode).

Gap (High): No detection system currently achieves reliable high-throughput attribution of AI-generated content at platform scale with acceptable false-positive rates. The detection deficit creates favorable operating conditions for GenAI-assisted IO.

Key Connections

  • Information Operations — generative AI is the primary capability-multiplier for IO at scale; content factory, synthetic media, and automated spearphishing are all IO-layer applications
  • Deepfakes — synthetic video and audio generation is the most operationally significant IO-relevant GenAI capability
  • Troll Farms — generative AI can partially substitute for human troll farm operators, reducing the cost per piece of inauthentic content
  • Computational Propaganda — GenAI is the most significant capability upgrade to computational propaganda infrastructure since social media
  • Responsible Scaling Policy — Anthropic’s RSP is the primary self-regulatory framework governing GenAI dual-use risk at the model-deployment level
  • 2025-09-23_2509_19147 — Daepp et al. (2025): empirical study of generative propaganda at scale; primary reference
  • 2026-04-17_2604_15717 — Hung et al. (2026): attack surface mapping for aligned LLMs in gray zone IO applications

Sources

  • Daepp, Madalina I. G., et al. “Generative Propaganda.” arXiv 2509.19147 (2025). Fact, High — empirical documentation of GenAI-assisted propaganda generation at scale; primary reference for IO-relevance.
  • Hung, Melody Y., et al. “Into the Gray Zone.” arXiv 2604.15717 (2026). Fact, High — comprehensive attack surface mapping for aligned LLMs under domain-context jailbreaking.
  • European Parliament. Artificial Intelligence Act (Regulation EU 2024/1689). Fact, High — primary binding governance instrument; Arts. 50, 53 on transparency obligations for AI-generated content and general-purpose AI models.

Graph View

Backlinks