← Back to Insights

Module: Deepfakes and Synthetic Media

By SAUFEX Consortium 23 January 2026

Purpose: You’ll learn to assess synthetic media claims accurately — understanding what’s actually possible, what detection tells you, and how to avoid both panic and dismissal.


[screen 1]

The Setup

A video surfaces. Someone claims it’s a deepfake. Someone else insists it’s authentic. Both cite “evidence.” Neither provides analysis.

You need to make a judgment — or at least an informed assessment of uncertainty. How?

This module provides the framework: what deepfakes actually are, what detection actually works, and how to assess authenticity claims without overclaiming in either direction.


[screen 2]

What “Deepfake” Actually Means

The term gets used loosely. Let’s be precise:

Deepfakes specifically refer to AI-generated synthetic media using deep learning neural networks. They:

  • Learn from large datasets of images/video
  • Generate new content that mimics the source
  • Improve with training data and compute

Not deepfakes (but often called that):

  • Basic photo editing (Photoshop)
  • Simple video cuts and splices
  • Out-of-context genuine footage
  • CGI and special effects (different technology)

Precision matters. “Deepfake” has become a panic term. Most manipulated media isn’t deepfaked.


[screen 3]

Current Capabilities: Realistic Assessment

What’s actually possible now:

Voice cloning:

  • High quality from relatively small samples
  • Can be convincing in phone/audio-only contexts
  • Real-time voice conversion exists but has tells
  • Current state: Impressive and dangerous

Face swapping:

  • Quality varies widely by tool and effort
  • Consumer apps produce obvious fakes
  • Professional tools require skill and time
  • Current state: Variable, often detectable

Lip-sync manipulation:

  • Making someone appear to say different words
  • Improving rapidly
  • Still often has artifacts on close inspection
  • Current state: Moderate threat

Full synthetic video:

  • Creating entirely fake scenes
  • Still limited and often uncanny
  • High-quality requires significant resources
  • Current state: Early stage

[screen 4]

The Evidence Ladder for Synthetic Media

Classify your evidence when assessing authenticity:

Strong signals (authentic):

  • Multiple independent recordings of same event
  • Original metadata intact and verified
  • Corroborating witnesses at event
  • Part of continuous, unedited footage stream
  • Provenance chain clearly documented

Strong signals (synthetic):

  • Confirmed AI artifacts by technical analysis
  • Source file metadata shows generation
  • Creator admitted fabrication
  • Multiple detection tools agree

Medium signals:

  • Single detection tool flags as synthetic
  • Some visual inconsistencies present
  • Missing provenance but no positive indicators
  • Contextual factors suggest manipulation motive

Weak signals:

  • “It looks fake to me”
  • “The technology exists”
  • “Who benefits?” reasoning
  • Claims without supporting analysis

[screen 5]

Detection: What Actually Works

Technical detection approaches:

Artifact analysis:

  • Unnatural blinking patterns (early deepfakes)
  • Lighting inconsistencies on faces
  • Blurred boundaries, especially hairlines
  • Temporal inconsistencies (frame-to-frame)
  • Audio-visual sync issues

Metadata analysis:

  • File creation timestamps
  • Device/software signatures
  • Compression artifacts
  • Chain of custody

Provenance tracking:

  • First appearance of content
  • Who posted and when
  • Platform verification (if available)
  • Cross-referencing with known events

[screen 6]

Detection Limits: What Doesn’t Work

Be honest about what detection can’t do:

No detection is definitive. All tools have false positive and false negative rates. “90% confidence” means 10% wrong.

The arms race is real. Detection improves; generation improves faster. Current artifacts may be absent in future fakes.

Human judgment is unreliable. People overestimate their ability to spot fakes. Experts do only marginally better than laypeople on visual inspection alone.

Absence of evidence isn’t evidence. “No artifacts detected” doesn’t mean “definitely authentic.” It means detection didn’t find anything.


[screen 7]

The Voice Cloning Problem

Audio manipulation deserves special attention:

Why voice is different:

  • Easier to fake convincingly than video
  • Phone/audio contexts reduce quality expectations
  • Less scrutiny than video in most situations
  • Real-time conversion increasingly possible

Real threat scenarios:

  • Impersonation calls for fraud
  • Fake audio “evidence” in disputes
  • Manipulated recordings presented as authentic
  • Voice verification systems compromised

Assessment approach:

  • Context matters more than detection
  • Who recorded? Where? Why?
  • Can the conversation be corroborated?
  • Does it align with known facts?

[screen 8]

The Liar’s Dividend

When deepfakes become possible, authentic content becomes deniable.

The dynamic:

  • Real footage exists of embarrassing/incriminating behavior
  • Subject claims “it’s a deepfake”
  • Without definitive verification, doubt persists
  • Authentic evidence loses impact

This already happens:

  • Politicians dismissing real recordings as fake
  • “That video is manipulated” as reflexive defense
  • Audiences pre-disposed to believe denials

Key insight: Deepfake capability creates damage even when no deepfake is used. The possibility enables denial of reality.


[screen 9]

Practical Scenario

Situation: A video appears showing a political candidate making a controversial statement. Within hours:

  • Campaign claims it’s a deepfake
  • Opponents claim it’s authentic
  • Media is inquiring whether to report

You have: 4 hours before editorial deadline

Your task: Assess authenticity with confidence level and recommendation


[screen 10]

Working the Scenario

Step 1: Provenance check (30 min)

  • Where did video first appear?
  • Who uploaded it?
  • Is there a chain of custody?
  • Any original metadata available?

Step 2: Context verification (30 min)

  • Was candidate at claimed location/date?
  • Any corroborating footage/witnesses?
  • Does statement align with or contradict known positions?

Step 3: Technical analysis (1 hour)

  • Visual inspection for obvious artifacts
  • Run through available detection tools
  • Check audio-visual sync
  • Analyze compression and encoding

Step 4: Assessment synthesis (1 hour)

  • Weigh evidence by strength
  • Identify what’s known vs. unknown
  • Determine confidence level
  • Write recommendation with caveats

[screen 11]

Sample Assessment

“A 23-second video shows Candidate X stating ‘I would support [controversial policy].’

Provenance: First appeared on anonymous account, shared to platform at 14:32. No original metadata. Account created 3 days ago.

Context: Candidate was in the city shown on the claimed date (confirmed via public schedule). Statement contradicts 6 months of documented positions. No other footage of this appearance has surfaced.

Technical: No obvious deepfake artifacts on visual inspection. Single detection tool flags as ‘possible manipulation’ (67% confidence). Audio-visual sync appears consistent. Video compression is unusual (non-standard encoding).

Assessment: Authenticity uncertain. No definitive evidence of manipulation, but provenance is weak, context raises questions, and one technical indicator is anomalous.

Confidence: Low. Cannot confirm as authentic or fake with available evidence.

Recommendation: Do not report as fact. If covering, note ‘unverified video’ with authentication caveats. Continue investigation. Request campaign provide evidence of deepfake claim.”


[screen 12]

Real-World Deepfake Incidents

Context helps calibration:

Zelenskyy surrender video (2022):

  • Crude deepfake urging Ukrainian surrender
  • Quickly identified and debunked
  • Low quality, minimal impact
  • Lesson: Most FIMI deepfakes have been poor quality

CEO voice fraud (various):

  • Multiple cases of voice cloning for wire fraud
  • Often successful initially
  • Detected through unusual requests, not voice analysis
  • Lesson: Audio deepfakes are the immediate commercial threat

Political non-events:

  • Many viral claims of “deepfake” are actually authentic or simple edits
  • “Deepfake” has become generic accusation
  • Lesson: Most “deepfake” claims are wrong

[screen 13]

The Quality-Impact Tradeoff

Counterintuitive finding:

High-quality deepfakes are rare in disinformation campaigns.

Why?

  • Creating convincing deepfakes requires time and skill
  • Most disinformation doesn’t need them
  • Simple manipulation (selective editing, false context) works fine
  • Effort vs. return often doesn’t justify

When deepfakes matter:

  • Targeted fraud (worth individual investment)
  • High-value targets (worth the effort)
  • Specific “evidence” fabrication
  • Harassment campaigns against individuals

For most information manipulation, traditional techniques suffice.


[screen 14]

Non-Consensual Intimate Imagery

The most common actual use of deepfakes:

Reality:

  • Most deepfakes (by volume) are NCII
  • Primarily targets women
  • Used for harassment, extortion, abuse
  • Technology makes creation trivially easy

Why this matters for FIMI:

  • Shows what’s technically possible
  • Demonstrates real threat to individuals
  • Often overlooked in political discourse
  • Detection and response lessons apply

This isn’t abstract future risk. It’s current, prevalent harm.


[screen 15]

Response Approaches

How to respond depends on the situation:

If deepfake suspected:

  • Investigate before claiming
  • Don’t amplify by over-covering
  • Note uncertainty if reporting
  • Preserve original for analysis

If deepfake claimed about authentic:

  • Demand evidence for the claim
  • “It’s a deepfake” is an assertion requiring proof
  • Context and corroboration matter
  • Don’t let denial stand without scrutiny

If authenticity genuinely uncertain:

  • State uncertainty clearly
  • Don’t default to either position
  • Continue investigation
  • Avoid contributing to noise

[screen 16]

DIM Application

When deepfakes are the concern:

Gen 3 (Prebunking):

  • Inoculate audiences that deepfakes exist
  • Teach verification habits before incidents
  • Set expectation that video requires verification

Gen 4 (Platform intervention):

  • Synthetic media policies
  • Labeling requirements
  • Detection integration

Gen 5 (Structural):

  • Build authentication infrastructure
  • Develop provenance standards
  • Create trusted verification institutions

Gen 2 (Debunking):

  • Useful for specific fake instances
  • Risk: amplification of the fake
  • Most effective when fake is already widespread

[screen 17]

Stop Rules

Know when to stop investigating:

Stop when:

  • Evidence clearly points one direction (authentic or synthetic)
  • You’ve reached your timebox
  • Additional analysis unlikely to change confidence level
  • Stakes don’t justify continued investigation

Document uncertainty:

  • “Could not verify authenticity with available evidence”
  • “No definitive indicators of manipulation detected; authenticity not confirmed”
  • “Claim of deepfake is unsubstantiated”

Saying “I don’t know” with documentation is better than false certainty.


[screen 18]

Common Mistakes

Mistake 1: “It’s a deepfake” as explanation Don’t use “deepfake” as catch-all for suspicious video. Most manipulation isn’t AI-generated.

Mistake 2: Detection tools as arbiters No tool is definitive. They provide input, not answers.

Mistake 3: “The technology exists, therefore…” Capability ≠ deployment. Most scenarios don’t involve deepfakes even though they could.

Mistake 4: Visual intuition as evidence “It looks fake to me” is weak evidence. Document specific indicators.

Mistake 5: Assuming all denials are false Sometimes authentic footage is wrongly called deepfake. Sometimes it actually is.


[screen 19]

Building Long-Term Resilience

Beyond individual assessment:

Institutional:

  • Provenance documentation standards
  • Authentication infrastructure
  • Verification capacity building

Personal:

  • Healthy skepticism as default
  • Verification habits before sharing
  • Tolerance for uncertainty
  • Resistance to both panic and dismissal

Societal:

  • Media literacy education
  • Trust in verification institutions
  • Legal frameworks for synthetic media
  • Platform accountability

The deepfake challenge is partly technical, mostly social.


[screen 20]

Module Assessment

Scenario: Three pieces of content surface during an election period:

Content A: Audio recording of a candidate making controversial remarks. Campaign says authentic, opposition says AI-generated. No video, originally shared on encrypted platform.

Content B: Video of candidate at a rally saying something incendiary. Low resolution, shared as screenshot-video of someone’s phone. Campaign denies event happened.

Content C: High-quality video interview with candidate making policy statements that contradict previous positions. Published by established media outlet.

Task (15 minutes):

  1. For each content piece, list the strongest signal type available
  2. Which would you prioritize for investigation and why?
  3. For Content A (audio), what specific verification steps would you take?
  4. What would you NOT claim about any of these without additional evidence?

Scoring:

  • Credit distinguishing signal strength across media types
  • Reward appropriate prioritization
  • Penalize treating all as equal threat
  • Credit acknowledging audio vs. video detection differences

[screen 21]

Key Takeaways

  • “Deepfake” is a specific technology, not a general term for fake media
  • Current capabilities: voice cloning is advanced; video varies; full synthetic is limited
  • Detection tools provide input, not definitive answers
  • The liar’s dividend means capability creates harm even without deployment
  • Most manipulated media isn’t AI-generated — don’t overclaim
  • Provenance and context often matter more than technical analysis
  • Voice/audio manipulation is the more immediate practical threat
  • NCII is the most common actual deepfake use — don’t ignore it
  • Stop rules and uncertainty documentation are essential
  • Match verification effort to stakes; “I don’t know” is acceptable

Next Module

Continue to: AI-Powered Targeting and Personalization — How AI enables individualized manipulation, what’s actually deployed vs. theoretical, and detection approaches.