Purpose: You’ll learn what AI actually changes about information manipulation, distinguish genuine AI-enabled threats from hype, and develop frameworks for assessing AI-related claims.
[screen 1]
Starting Point
AI will destroy democracy. AI-generated content is everywhere. We can’t trust anything anymore.
You’ve heard these claims. Some contain truth. Most contain more anxiety than analysis.
AI does change information manipulation. But understanding how requires precision, not panic. The goal here is to give you a framework for thinking about AI threats — not a list of things to fear.
[screen 2]
What AI Actually Does
Generative AI — systems that produce new content — can create:
Text: Articles, comments, emails, social media posts Images: Photorealistic scenes, manipulated photos, diagrams Audio: Voice cloning, synthetic speech, audio fakes Video: Face swaps, synthetic talking heads, full scene generation
This is real. These capabilities exist. The question is: what does this change, practically?
[screen 3]
The “Everything Is Fake” Trap
A common mistake: assuming AI changes everything.
Reality check:
- Most disinformation still doesn’t require AI
- Text-based manipulation works with or without LLMs
- Human-created content remains abundant
- AI-generated content isn’t automatically more effective
The apocalyptic framing (“we can’t trust anything”) is itself a form of information manipulation — paralyzing rather than empowering.
Your job: distinguish what AI actually enables from what it doesn’t change.
[screen 4]
Three Things AI Changes
AI does transform some aspects of information manipulation:
1. Production costs
- Creating text, image, audio content is cheaper
- Variations and translations automated
- Volume possible at lower investment
2. Speed
- Reactive campaigns faster to deploy
- Content can match real-time events
- Iteration and optimization accelerated
3. Personalization potential
- Tailored messaging at scale
- Language and style matching
- Targeted emotional appeals
These are real shifts. But they’re shifts in capability, not magical transformations.
[screen 5]
What AI Doesn’t Change
Despite the hype, fundamentals remain:
Distribution still matters
- AI-generated content needs to reach audiences
- Platform dynamics, seeding, amplification unchanged
- “Making fake content” ≠ “making it spread”
Detection remains possible
- AI-generated content has artifacts
- Behavioral analysis still works
- Metadata and provenance leave traces
Human psychology unchanged
- People still have biases, trust networks, verification habits
- What makes content persuasive isn’t primarily its production method
- Emotional resonance matters more than AI sophistication
[screen 6]
The Asymmetry Problem (Real But Nuanced)
Commonly stated: “AI makes creation faster than detection.”
This is partly true:
- LLMs can generate text faster than humans can read it
- Image generation outpaces manual verification
- Scale is genuinely challenging
But consider:
- Detection can be automated too
- Not all AI content requires individual verification
- Behavioral signals often matter more than content authenticity
- “Winning” isn’t about detecting every fake
The asymmetry is real but manageable with the right approach.
[screen 7]
AI for Text: Capabilities and Limits
Large Language Models (GPT, Claude, Llama, etc.) can:
- Write fluent text in any style
- Generate multiple variations
- Translate between languages
- Adopt personas
What this enables:
- Higher volume operations with less labor
- Consistent voice across many accounts
- Faster translation and localization
What it doesn’t enable:
- Automatic credibility
- Guaranteed virality
- Immunity from behavioral detection
A thousand AI-generated comments still need to come from accounts, follow patterns, exhibit behaviors.
[screen 8]
AI for Images: Beyond “Photoshop”
Image generation tools can:
- Create photorealistic scenes from text descriptions
- Edit photos with natural language instructions
- Generate faces of nonexistent people
- Modify existing images seamlessly
Real threat:
- Fabricated “evidence” (photos of events that didn’t happen)
- Impersonation through generated headshots
- Manipulated context
Practical limits:
- Artifacts often detectable (for now)
- Provenance and context still matter
- Reverse image search still works on authentic components
- Metadata often reveals generation
[screen 9]
AI for Audio and Video: The “Deepfake” Question
Current capabilities:
- Voice cloning from small samples (impressive)
- Face swapping in video (variable quality)
- Lip-sync manipulation (improving)
- Full synthetic video (still limited)
The media narrative:
- “Anyone can create fake video of anyone”
- “You can’t trust video evidence”
The reality:
- High-quality video deepfakes still require effort
- Real-time deepfakes exist but have tells
- Most video manipulation is simpler (selective editing, false context)
- Audio cloning is more advanced than video synthesis
[screen 10]
Assessing AI Claims: The Evidence Ladder
When you encounter claims about AI-generated content, classify the evidence:
Strong signals:
- Confirmed AI tool artifacts by technical analysis
- Provenance data showing generation source
- Creator admission or platform attribution
- Multiple independent technical assessments
Medium signals:
- Stylistic indicators consistent with AI
- Image artifacts visible on close inspection
- Account behavior patterns suggesting automation
- Content volume inconsistent with human capacity
Weak signals:
- “It looks AI-generated”
- “The style seems off”
- “This is too polished to be real”
- Absence of provenance (could be anything)
[screen 11]
Common Analytical Errors
Error 1: Everything suspicious is AI
- “This campaign seems sophisticated, must be AI”
- Reality: Sophisticated operations existed before AI
Error 2: AI detection tools are definitive
- “The detector said 90% AI-generated”
- Reality: Detection tools have significant false positive/negative rates
Error 3: AI content is automatically more dangerous
- “AI-generated, therefore high priority”
- Reality: Impact depends on reach and resonance, not production method
Error 4: Absence of AI artifacts proves authenticity
- “No AI detected, therefore genuine”
- Reality: AI detection is imperfect; human-created content can also be false
[screen 12]
Practical Scenario
Situation: A video surfaces of a political candidate making controversial statements. Critics claim it’s a deepfake. Supporters claim it’s genuine leaked footage. You have 3 hours before your outlet needs to decide whether to report.
Your task:
- Assess authenticity (what evidence would you seek?)
- Rate confidence (with available evidence)
- Recommend action (report, investigate more, or pass)
What you don’t do: Claim “deepfake” or “authentic” without supporting evidence.
[screen 13]
Working the Scenario
Step 1: Source assessment
- Where did video first appear? (Original source = medium signal)
- Who is promoting it? (Cui bono isn’t proof, but context matters)
- What’s the provenance chain? (Each step degrades reliability)
Step 2: Technical analysis (if time permits)
- Facial movement consistency
- Audio-visual sync
- Lighting and shadow consistency
- Background element coherence
- Known AI artifacts
Step 3: Context verification
- Was the candidate at claimed location/time?
- Do statements align with known positions?
- Any corroborating or contradicting evidence?
Step 4: Confidence rating
- Multiple strong signals → higher confidence either way
- Medium signals only → acknowledge uncertainty
- Weak/conflicting signals → more investigation or don’t report
[screen 14]
Sample Assessment
“A 47-second video shows Candidate X stating ‘I support policy Y.’ Analysis:
Source: First appeared on anonymous Telegram channel, then amplified on X/Twitter by partisan accounts. No original source identified.
Technical: Lip-sync appears consistent. No obvious deepfake artifacts on initial review. Professional analysis would require time we don’t have.
Context: Statement contradicts Candidate X’s documented positions. No corroborating appearance at claimed venue.
Assessment: Authenticity uncertain. Technical indicators inconclusive. Contextual factors raise questions.
Confidence: Low. Neither authentic nor deepfake confirmed.
Recommendation: Do not report as fact. If covering, present as ‘unverified video’ with authentication caveats. Continue investigation.”
[screen 15]
The “Liar’s Dividend” Problem
When deepfakes become possible, authentic footage becomes deniable.
The dynamic:
- Any embarrassing video can be dismissed as “deepfake”
- “Plausible deniability” expands for genuine recordings
- This happens whether or not deepfakes are actually used
This is already happening:
- Politicians have claimed authentic recordings are fake
- “It’s AI” becomes universal excuse
Implication:
- Deepfake capability harms truth even when not deployed
- Authentication infrastructure becomes more important
- Context and corroboration matter more than technical analysis
[screen 16]
AI-Enabled Amplification
Beyond content creation, AI changes distribution:
Automated engagement:
- AI-powered bots with more human-like behavior
- Conversational agents that respond contextually
- Pattern variation to evade simple detection
Targeting optimization:
- AI analyzes which content performs
- Automatic A/B testing at scale
- Rapid iteration toward effective messaging
But remember:
- Sophisticated bots still exhibit behavioral patterns
- Account-level signals remain relevant
- Platform detection also uses AI
The arms race is real, but detection isn’t defeated.
[screen 17]
The Personalization Threat
Most discussed, least deployed (so far):
The theory:
- AI analyzes your beliefs, fears, preferences
- Generates content specifically designed for you
- Hyper-persuasive individualized manipulation
Current reality:
- Requires extensive data on individuals
- Computational cost scales with personalization
- Still experimental in information operations
- Advertising already does crude versions
Watch for:
- Targeted phishing using personal details
- Tailored narratives for specific communities
- A/B tested messaging for different segments
This is a developing capability, not yet routine in FIMI operations.
[screen 18]
What AI Means for Response
Gen 2 (Debunking):
- AI creates more volume to debunk
- But AI can also assist fact-checking
- Prioritization becomes more important
Gen 3 (Prebunking):
- Inoculating against AI-specific techniques
- Teaching critical evaluation for synthetic media
- “This could be AI-generated” as mental habit
Gen 4 (Moderation):
- AI detection as input for platform decisions
- Behavioral signals often more reliable than content analysis
- Policy decisions about AI content disclosure
Gen 5 (Interaction):
- Building resilience that doesn’t depend on authentication
- Community-based verification
- Trust networks that survive AI capabilities
[screen 19]
Stop Rules for AI Assessment
Know when you’ve done enough analysis:
Stop when:
- Evidence tier is clear (strong/medium/weak)
- Additional analysis unlikely to change confidence
- Time would be better spent on response decisions
- You’ve documented what’s known and unknown
Don’t chase:
- Perfect technical certainty (rarely achievable)
- “Proving” it’s AI when behavior matters more
- Every possible artifact or indicator
The goal is accurate assessment under uncertainty, not forensic proof.
[screen 20]
Avoiding Hype
The AI/disinformation discourse contains significant hype:
Be skeptical of:
- “AI will make truth impossible”
- “Detection is hopeless”
- “Everything online is fake”
- Specific claims without evidence
Maintain perspective:
- AI is a tool, not magic
- Most manipulation doesn’t require AI
- Detection and defense are also advancing
- Human judgment remains essential
Your job: calm analysis, not panic amplification.
[screen 21]
Module Assessment
Scenario: Your organization receives a tip that a widely shared image showing protestors destroying property is AI-generated. The tipster claims “obvious artifacts” but provides no analysis. The image has reached 500,000 views.
Task (12 minutes):
- List what you would check to assess AI generation claims
- Classify potential findings as strong/medium/weak signals
- Write a 3-sentence assessment assuming you find only medium signals
- What action would you recommend?
- Identify one thing you explicitly will NOT claim without strong evidence
Scoring:
- Penalize certainty without evidence
- Credit systematic approach
- Reward acknowledgment of detection limitations
[screen 22]
Key Takeaways
- AI does change information manipulation: production costs, speed, personalization potential
- AI doesn’t change: distribution dynamics, human psychology, fundamental detection approaches
- “AI-generated” is a claim requiring evidence, not an assumption
- Detection tools are imperfect; behavioral signals often more reliable than content analysis
- The “liar’s dividend” means AI capability harms truth even when not deployed
- Avoid both naive dismissal and apocalyptic hype
- Your response should match evidence tier, not worst-case assumptions
AI changes the game. It doesn’t end it.
[screen 23]
Continuing Your Learning
This module provides foundation. Subsequent modules cover:
- Deepfakes and Synthetic Media: Specific techniques and detection approaches
- AI-Powered Targeting: How personalization works and its limits
- Bots and Automated Amplification: Behavioral detection of AI-driven accounts
- AI for Detection and Defense: Using AI tools responsibly
Treat each as building on this framework: assess capabilities realistically, classify evidence properly, respond proportionately.
Next Module
Continue to: Deepfakes and Synthetic Media — Technical details on synthetic video and audio, what’s actually possible, and how to assess authenticity claims.