L1
·
Quiz
·
Lab
L2
·
Quiz
·
Lab
L3
·
Quiz
·
Lab
L4
·
Quiz
·
Lab
Module Test
Module 3 · Lesson 1

What Is a Deepfake?

AI-generated synthetic media that places real people's faces, voices, and words into fabricated scenarios
How did a technology invented for film effects become one of the most dangerous tools for spreading disinformation?

On January 1, 2019, Gabonese president Ali Bongo Ondimba appeared in a televised New Year's address — the first time the public had seen him in months following a suspected stroke. Within days, analysts and opposition leaders began claiming the video was a deepfake. The Gabonese military cited the suspicious video as partial justification for a coup attempt on January 7. Independent investigators later concluded the video was probably authentic, but the damage was done: a real video had been dismissed as fake, triggering real political violence.

The episode illustrated a phenomenon researchers now call the "liar's dividend" — even if deepfakes themselves cause harm, the mere existence of the technology lets bad actors dismiss genuine evidence by simply saying "that's a deepfake."

The Technology Behind Deepfakes

The word "deepfake" combines deep learning — a form of AI — with fake. The technique was first popularized in late 2017 when a Reddit user named "deepfakes" began posting AI-generated face-swap pornography using celebrity faces. The underlying method is called a Generative Adversarial Network (GAN).

A GAN pits two neural networks against each other. The generator creates synthetic images; the discriminator tries to detect which images are fake. Through millions of training rounds, the generator keeps improving until its fakes fool the discriminator — and often, human eyes as well. Modern deepfakes also use diffusion models and face-swapping encoders that can map one person's facial geometry onto another person's video in real time.

Audio deepfakes — sometimes called voice clones — work similarly, training on recordings of a person's speech to produce new utterances they never made. In 2023, robocalls using a cloned voice of President Biden were sent to New Hampshire voters telling them not to vote in the primary election. The Federal Communications Commission traced the calls and fined the perpetrators $6 million.

Why It Matters

Deepfakes are not hypothetical threats. The nonprofit Security Research Labs documented over 500,000 deepfake videos online in 2023. Roughly 96% depicted non-consensual intimate imagery — a serious crime in many jurisdictions. A smaller but growing share targets politicians, journalists, and corporate executives for fraud and disinformation.

A Brief Timeline of Deepfake Milestones
2017
Reddit "deepfakes" account
User "deepfakes" posts AI face-swap videos. Reddit bans the community in February 2018 but the open-source code has already spread globally.
2018
BuzzFeed/Jordan Peele PSA
Director Jordan Peele and BuzzFeed release a deepfake of Barack Obama to warn the public about synthetic media — one of the first high-profile educational deepfakes.
2019
Gabon coup attempt
A presidential video dismissed (wrongly) as a deepfake contributes to political instability and a military coup attempt — the first documented case of the "liar's dividend" in action.
2023
Biden New Hampshire voice clone
Robocalls using a cloned version of President Biden's voice tell Democratic voters to skip the primary. The FCC fines the operator $6 million in 2024.
2024
Taylor Swift deepfake images
Explicit AI-generated images of Taylor Swift circulate on X (Twitter), reaching 47 million views before removal. The incident prompts congressional calls for federal deepfake legislation.
Key Terms
DeepfakeAI-synthesized media — video, audio, or images — that realistically depicts a person saying or doing something they did not actually say or do.
GANGenerative Adversarial Network — two competing neural networks that together learn to produce increasingly convincing synthetic content.
Voice cloneAn AI model trained on a person's real speech that generates new audio utterances in their voice.
Liar's dividendThe ability to discredit genuine evidence by claiming it is AI-generated — a side effect of deepfake technology regardless of whether a fake was actually used.
Lesson Takeaway

Deepfakes emerged from open-source AI research, spread rapidly without regulatory guardrails, and now affect elections, financial markets, and personal safety. Understanding the technology — not just the danger — is the first step toward effective detection and defense.

Module 3 · Lesson 1

Quiz: What Is a Deepfake?

4 questions — select the best answer for each
1. What does the word "deepfake" combine?
Correct! "Deepfake" combines deep learning (a type of AI) with fake — it was coined by the Reddit user who first popularized the technology in 2017.
Not quite. "Deepfake" combines deep learning (the AI technique) with fake — it was coined by a Reddit user who posted AI face-swap videos in 2017.
2. In a Generative Adversarial Network (GAN), what role does the discriminator play?
Correct! The discriminator acts as an adversary to the generator — by trying to spot fakes, it forces the generator to create better and better synthetic content through competition.
Not quite. In a GAN, the discriminator tries to detect fakes. This competition forces the generator to keep improving until its outputs fool the discriminator.
3. The 2019 Gabon incident is significant because it demonstrated which concept?
Correct! The Gabon case is the clearest early example of the "liar's dividend" — the mere existence of deepfakes made it possible to cast doubt on genuine footage, contributing to a coup attempt.
The Gabon case actually illustrated the "liar's dividend" — a real presidential video was falsely accused of being a deepfake, and that accusation helped justify a military coup attempt.
4. The 2023 Biden voice clone robocalls in New Hampshire were intended to:
Correct! The robocalls used a cloned version of Biden's voice to tell Democratic primary voters "Don't vote" — a direct attempt at voter suppression using AI-generated audio.
The calls told Democratic voters not to vote in the primary — a clear attempt at voter suppression using an AI-cloned version of Biden's voice. The FCC later fined the perpetrators $6 million.
Module 3 · Lab 1

Deepfake Origins Lab

Chat with the AI to explore how deepfake technology works and spread

Your Mission

You're investigating the history and mechanics of deepfake technology. Use the AI assistant below to explore questions about GANs, the timeline of deepfake development, and the real-world incidents from Lesson 1. Ask at least three questions to complete the lab.

Starter prompts: "Explain how a GAN creates a deepfake in simple terms." · "What made the Gabon 2019 incident so significant?" · "How did voice cloning work in the Biden New Hampshire robocalls?"
AI Lab Assistant
Deepfake Origins
Hello, Truth Detective! I'm here to help you explore the origins and mechanics of deepfake technology. Ask me about how GANs work, the history of deepfakes, the Gabon incident, the Biden voice clone, or anything else from Lesson 1. What would you like to investigate?
Module 3 · Lesson 2

How Deepfakes Spread & Who Makes Them

From hobbyist forums to state-sponsored influence operations — the full ecosystem of synthetic media creation and distribution
Who is actually making deepfakes, and why do they spread so much faster than the corrections that follow?

On March 16, 2022, a deepfake video appeared on hacked Ukrainian TV and news websites showing President Volodymyr Zelensky telling Ukrainian soldiers to "lay down their weapons and surrender." The video was technically crude — Zelensky's head appeared too large for his body, his speech pattern was unnatural — but it still spread rapidly on Telegram and Facebook before platforms could remove it.

Zelensky responded within hours with a real video from his office, mocking the fake. Meta and YouTube removed the deepfake, but not before it had been viewed hundreds of thousands of times. Military analysts noted the video's primary goal was not to fool soldiers but to create confusion and fear — to make Ukrainians doubt what was real.

The Deepfake Ecosystem

Deepfakes are not made by a single type of actor. Research from the Sensity AI threat intelligence firm categorizes creators into four rough groups:

Non-Consensual Intimate Images (NCII)

The largest category by volume. Predominantly targets women. Apps like DeepNude (banned in 2019) and its successors allow users to remove clothing from photos. UK law criminalized this in 2023; the US DEFIANCE Act passed in 2024.

Financial Fraud

In 2020, criminals used a deepfake audio of a company director's voice to authorize a $35 million bank transfer at a Hong Kong bank. Similar CEO voice fraud has been reported in the UK and UAE.

Political Disinformation

State-sponsored and independent actors use deepfakes to discredit political opponents. Slovakia's 2023 election saw audio deepfakes of liberal candidate Michal Šimečka supposedly discussing vote-buying — released 48 hours before the election, too close to deadline to debunk effectively.

Satire & Entertainment

Legal, labeled deepfakes for comedy shows, art projects, and film production. These legitimate uses create public familiarity with the technology — but also normalize the sight of famous faces saying unexpected things.

Why Deepfakes Spread So Fast

Researchers at MIT Media Lab found that false information spreads six times faster on Twitter/X than true information — and deepfakes add a visual dimension that text corrections cannot easily counter. Three platform dynamics accelerate deepfake spread:

Algorithmic amplification: Outrage and novelty drive engagement. A video of a world leader appearing to surrender generates more clicks than a quiet fact-check post. Platforms optimized for engagement reward sensational content.

Cross-platform laundering: A deepfake posted on Telegram can be screenshot and reshared on WhatsApp, Instagram, and TikTok — each hop losing the original context and making removal harder.

Correction asymmetry: The original deepfake can be shared indefinitely for free. Effective corrections require coordinated fact-checkers, platform cooperation, and media coverage — all of which cost time and money and reach fewer people than the original fake.

Case Study — Slovakia's 2023 Election

Two days before the September 30 election, audio deepfakes circulated on Facebook appearing to show progressive candidate Michal Šimečka discussing how to rig the election. Meta's fact-checkers could not respond in the mandatory 48-hour pre-election media silence window. Šimečka lost by a narrow margin. Researchers cannot prove the deepfakes changed the outcome, but they document it as the first election where a deepfake surfaced too close to voting day to be debunked in time.

The Infrastructure of Creation

Creating a convincing deepfake in 2017 required thousands of training images and days of GPU time. By 2023, apps like FaceSwap, DeepFaceLab, Reface, and commercial services on Telegram bots could produce face-swapped videos in minutes from a single photograph. Voice cloning services like ElevenLabs require as little as three seconds of audio to clone a voice — a fact that led the company to implement abuse controls after its voice engine was used to clone public figures without consent.

The democratization of deepfake creation means the threat is no longer limited to well-funded state actors. A determined individual with a consumer-grade laptop can now produce media that, without specialized detection tools, is indistinguishable from authentic footage.

Lesson Takeaway

Deepfakes spread through a combination of algorithmic amplification, platform cross-posting, and the fundamental asymmetry between viral misinformation and slow corrections. Understanding who makes them — and why the platforms they travel through reward engagement over accuracy — is essential to understanding why detection and media literacy matter.

Module 3 · Lesson 2

Quiz: How Deepfakes Spread

4 questions — select the best answer for each
1. The March 2022 deepfake of President Zelensky was notable primarily because:
Correct! The Zelensky deepfake was technically crude (his head was oddly proportioned), yet it still spread hundreds of thousands of times. The goal was to create confusion and fear, not necessarily to fool every viewer.
The Zelensky deepfake was actually crude and easy to spot on close inspection — but it still spread widely, demonstrating that technical quality matters less than speed and emotional impact.
2. In the 2020 Hong Kong bank fraud case, how were deepfakes used?
Correct! Criminals used an AI-cloned voice of a company director to call a bank manager and authorize a $35 million wire transfer — one of the largest documented deepfake fraud cases.
In this case, the deepfake was audio, not video. Criminals cloned the voice of a company director to call a bank manager and authorize a $35 million transfer.
3. What made the Slovakia 2023 election deepfake particularly effective as disinformation?
Correct! Timing was the key weapon. Slovak election law prohibits campaigning in the final 48 hours — exactly when the deepfake appeared. Fact-checkers could not reach voters before they voted.
The timing was the key. Slovak law prohibits election coverage in the 48 hours before voting — so the deepfake appeared precisely when fact-checkers were legally constrained from responding publicly.
4. According to MIT Media Lab research cited in the lesson, how much faster does false information spread on Twitter/X compared to true information?
Correct! MIT Media Lab found false information spreads approximately six times faster than true information on social media — a fundamental asymmetry that deepfakes exploit because visual content generates even more engagement than text.
MIT Media Lab research found the figure is about six times faster. This asymmetry is foundational to understanding why deepfake corrections so rarely catch the original misinformation.
Module 3 · Lab 2

Deepfake Spread Analysis Lab

Explore the mechanics of how synthetic media travels and why corrections fail to keep up

Your Mission

You're mapping the ecosystem of deepfake spread. Use the AI to explore platform dynamics, case studies from the lesson, and the structural reasons corrections fail. Ask at least three questions to complete the lab.

Starter prompts: "Why couldn't the Zelensky deepfake be stopped faster?" · "Explain correction asymmetry in your own words." · "What is cross-platform laundering and why does it matter for deepfakes?"
AI Lab Assistant
Deepfake Spread
Welcome back, Truth Detective! This lab focuses on how deepfakes spread — the platforms, the incentives, and the real cases from Lesson 2. Ask me about the Zelensky video, the Slovakia election deepfake, the correction asymmetry problem, or how voice-clone fraud works. What would you like to dig into?
Module 3 · Lesson 3

Detecting Deepfakes: Human & AI Methods

What human eyes notice, what detection software analyzes, and why neither method is foolproof
Can we build reliable deepfake detectors — and what happens when the detectors themselves become a target?

In 2016, the US Defense Advanced Research Projects Agency (DARPA) launched the Media Forensics (MediFor) program with a specific mandate: build AI tools that could automatically assess the integrity of photos and videos. Over five years and tens of millions of dollars, DARPA-funded teams at MIT Lincoln Laboratory, SRI International, and other research groups developed detection systems that could identify compression artifacts, inconsistent lighting, and pixel-level anomalies invisible to human eyes.

The program produced genuinely capable tools — but also surfaced a fundamental problem. As detection methods improved, so did the deepfakes. Adversarial training — feeding deepfake generators feedback from detectors — produced synthetic media specifically optimized to evade those detectors. The technology became a true arms race.

What Human Eyes Can Catch

Trained observers can spot deepfakes using visual cues that AI models still struggle to perfectly replicate. However, these cues become less reliable as the technology improves. Key tells that worked in 2020–2022 include:

👁️
Unnatural Blinking
Early deepfakes rarely blinked or blinked at inhuman intervals. Researchers at SUNY Albany published a paper on this in 2018. Developers rapidly patched it.
💡
Lighting Inconsistency
Light on a synthesized face may not match the background — catchlights (reflections in the iris) often appear in wrong positions or are absent entirely.
👂
Edge Artifacts
Hair, earrings, and collar edges often show blurring or warping where the face-swap boundary sits. Most visible when the subject moves quickly.
🦷
Teeth & Interior Mouth
Deepfake models frequently produce smeared or incorrect teeth. The interior of the mouth is structurally complex and historically an AI weak point.
🎵
Audio-Visual Sync
Lip movements that don't precisely match audio, or an unnatural absence of background breathing and ambient noise — especially in voice-clone audio.
🌊
Temporal Flickering
Frame-by-frame inconsistencies around the face boundary — the synthesized face "stabilizes" differently from the real background over time.
Important Caveat

Each of these tells has been substantially reduced or eliminated in state-of-the-art deepfakes made in 2023–2024. The detection cues above are useful for spotting lower-quality or quickly made fakes, but should never be the sole basis for concluding a video is real. Absence of visible artifacts does not confirm authenticity.

Automated Detection Tools

Several AI-powered detection tools are now publicly available or widely deployed by platforms:

Microsoft's Video Authenticator (2020) analyzes videos frame-by-frame for subtle blending artifacts and produces a confidence score. Microsoft released it specifically to counter election deepfakes but acknowledged it would become less effective as deepfake quality improved.

Deepware Scanner and FakeCatcher (Intel, 2022) use different methods — FakeCatcher analyzes blood-flow signals (rPPG) detectable in pixel color changes in real human skin that synthetic faces cannot replicate. Intel claimed 96% accuracy in testing, though independent verification in real-world conditions is harder to achieve.

In 2023, researchers at the University of Buffalo published a corneal reflection method: the eyes of real people reflect a consistent image of the environment; deepfake eyes show inconsistent or physically impossible reflections. This method was highly accurate at launch but was expected to be patched by generative models within months.

The Detection Arms Race

Every publicly released detector becomes a training signal for the next generation of deepfakes. This is not theoretical — researchers demonstrated in 2020 that adversarially trained deepfakes specifically designed to fool Microsoft's Video Authenticator could reduce its accuracy from 90% to below 50% with minimal degradation in visual quality visible to humans.

This means detection cannot be a static solution. The most reliable long-term approaches rely less on spotting artifacts and more on provenance — tracking where media came from, not just what it looks like.

Lesson Takeaway

Both human and automated deepfake detection work — imperfectly, temporarily, and only as long as deepfake technology isn't specifically trained to evade them. Smart detection combines visual inspection with source verification, platform context, and metadata analysis. No single method is reliable on its own.

Module 3 · Lesson 3

Quiz: Detecting Deepfakes

4 questions — select the best answer for each
1. DARPA's MediFor program revealed a fundamental problem with deepfake detection. What was it?
Correct! The MediFor program discovered that publishing detection methods creates a roadmap for adversarial training — generators can be tuned to fool known detectors, making the field an ongoing arms race.
The key problem MediFor revealed was adversarial training: once a detector is published, deepfake generators can be trained specifically to evade it, turning detection into a permanent arms race rather than a solved problem.
2. Intel's FakeCatcher (2022) detects deepfakes using which innovative biological signal?
Correct! FakeCatcher uses remote photoplethysmography (rPPG) — detecting the subtle color variations in skin caused by blood flow that real human physiology produces but synthetic faces do not.
FakeCatcher uses rPPG — remote photoplethysmography — which detects subtle color changes caused by blood flowing through real skin. Deepfake-generated faces don't have actual blood flow to replicate this signal.
3. Which visual cue was identified by SUNY Albany researchers in 2018 — and subsequently patched by deepfake developers?
Correct! SUNY Albany researchers published a paper in 2018 showing early deepfakes blinked rarely or unnaturally. This was a useful detection signal — until deepfake developers read the paper and fixed the problem.
SUNY Albany found that early deepfakes blinked abnormally — either too rarely or at inhuman intervals. Once published, this finding was quickly used to improve deepfake generators, eliminating the tell.
4. Why is "no visible artifacts" NOT sufficient to conclude a video is authentic?
Correct! Modern deepfakes have substantially reduced or eliminated the classic visual tells. A technically clean video cannot be confirmed as authentic by visual inspection alone — provenance and source verification are also required.
The issue is that modern deepfakes have been specifically trained to eliminate the artifacts that older detection methods relied on. The absence of visible tells doesn't confirm authenticity — it may just mean a better deepfake.
Module 3 · Lab 3

Detection Methods Lab

Apply your knowledge of deepfake detection techniques to real-world scenarios

Your Mission

You're a media forensics analyst reviewing suspicious videos. Use the AI to practice applying detection methods from Lesson 3. Describe a hypothetical video scenario and ask what detection approaches to use, or explore why each method has limitations. Ask at least three questions to complete the lab.

Starter prompts: "A video shows a politician appearing to announce a resignation. Walk me through how I'd analyze it for deepfake signs." · "Why is the corneal reflection method not a permanent solution?" · "What is provenance-based detection and how does it work?"
AI Lab Assistant
Detection Methods
Welcome, media forensics analyst! In this lab we'll work through deepfake detection methods — the visual tells human eyes can catch, the automated tools like FakeCatcher, and the fundamental limitations of each approach. Describe a scenario or ask me about any detection method from Lesson 3. What would you like to investigate?
Module 3 · Lesson 4

Defending Against Deepfakes: Policy, Provenance & Practice

From content credentials and watermarking to legislation and personal verification habits
If we can't always detect deepfakes by looking, what systems and habits can actually protect us?

In 2021, Adobe, Microsoft, Intel, Twitter, and the BBC co-founded the Coalition for Content Provenance and Authenticity (C2PA). Their goal was not to detect fakes after the fact but to create a cryptographic chain of custody for authentic media from the moment of capture.

By 2024, Sony and Leica had released cameras that embed C2PA Content Credentials — a cryptographic signature — into every photo at the moment of shutter press. News agencies including the Associated Press began testing the system. The credentials travel with the image and can be verified at any C2PA-compliant platform, showing where a photo was taken, on what device, and whether it has been edited.

The system is not foolproof — a deepfake could be photographed with a C2PA camera, acquiring false credentials — but it represents the first serious infrastructure for media provenance at scale.

The Policy Landscape

Laws governing deepfakes have developed unevenly. As of 2024, the major legislative actions include:

US — DEFIANCE Act (2024)

Signed into law in July 2024. Allows victims of non-consensual intimate deepfakes to sue creators and distributors in federal court. Specifically covers AI-generated content. Passed with bipartisan support following the Taylor Swift incident.

EU — AI Act (2024)

Requires deepfakes to be labeled as AI-generated content. Platforms must disclose when content is synthetic. Violations carry fines of up to 3% of global annual turnover. Effective from 2025.

UK — Online Safety Act (2023)

Criminalized sharing non-consensual intimate deepfakes. The 2024 Criminal Justice Bill went further, also criminalizing the creation of such content — not just its distribution.

China — Deepfake Regulations (2022)

China's Cyberspace Administration requires deepfakes to carry visible labels and prohibits using them to spread false news. Platforms must verify users before allowing deepfake creation — one of the strictest frameworks globally.

Technical Defenses: Watermarking & Provenance

Three technical approaches are being deployed at scale:

Invisible watermarking: Google's SynthID (2023) embeds imperceptible watermarks into AI-generated images and audio produced by Google's Gemini and Imagen tools. The watermark survives screenshots, compression, and mild editing, and can be detected by SynthID tools. Adobe's Content Authenticity Initiative uses a visible "Content Credentials" badge that links to provenance data.

Cryptographic provenance (C2PA): As described above, this embeds signed metadata at the moment of creation. The AP, Reuters, and major broadcasters have committed to using C2PA-compliant cameras and editing software for news photography.

Platform-level detection: Meta deploys automated deepfake detection and labels AI-generated content in its political advertising policy. YouTube requires creators to disclose when uploaded content uses AI to "realistically alter or generate" people or events.

The Verification Habit — What Individuals Can Do

No policy or technology eliminates the need for individual verification habits. For any surprising video of a public figure: (1) Check if the original source is a verified account or official channel. (2) Search for the same clip on established news sites — if only fringe outlets have it, that's a red flag. (3) Look for a C2PA "Content Credentials" badge on platforms that support it. (4) Use reverse image/video search (InVID, Google, TinEye) to find the original context. (5) Apply the deepfake visual checklist from Lesson 3 — not as definitive proof but as one signal among many.

The Consent and Ethics Framework

Beyond law and technology, the deepfake problem has a consent dimension. In 2023, the Screen Actors Guild (SAG-AFTRA) went on strike partly over studios' attempts to use AI to replicate actors' likenesses without compensation or consent. The resulting contract established that studios must obtain explicit, informed consent and negotiate compensation for any AI replication of a performer's voice or appearance.

This principle — that a person's likeness requires consent — is increasingly being recognized in law. California's AB 2602 (2024) specifically protects performers from posthumous AI replication of their likeness without estate consent. These frameworks establish that the ethical use of deepfake-adjacent technology begins with who gave permission, not just whether the output looks realistic.

Module Takeaway

Defending against deepfakes requires three layers working simultaneously: technical standards (C2PA, watermarking), legal frameworks (DEFIANCE Act, EU AI Act), and individual media literacy habits (source verification, visual inspection, provenance checking). No single layer is sufficient — the most resilient defense combines all three.

Module 3 · Lesson 4

Quiz: Defending Against Deepfakes

4 questions — select the best answer for each
1. What does the C2PA Content Credentials system do?
Correct! C2PA creates a cryptographic chain of custody for media — signed at capture, traveling with the file, and verifiable at any compliant platform. Sony, Leica, the AP, and Reuters are among the early adopters.
C2PA works through provenance, not detection. It cryptographically signs media at the moment it's created so its origin and edit history can be verified — it doesn't analyze the content for fakes.
2. The US DEFIANCE Act (2024) specifically addresses which category of deepfakes?
Correct! The DEFIANCE Act passed with bipartisan support after the Taylor Swift incident. It creates a federal civil cause of action for victims of non-consensual intimate AI-generated content.
The DEFIANCE Act specifically targets non-consensual intimate deepfakes, giving victims the right to sue in federal court. It was passed in 2024 with bipartisan support partly in response to the Taylor Swift incident.
3. Google's SynthID watermarking technology works by:
Correct! SynthID embeds an imperceptible watermark that survives common modifications like screenshots, compression, and mild editing — and can be detected by SynthID tools to identify the content as AI-generated.
SynthID works through invisible watermarking — the signal is embedded in the content itself, not as a visible label, and is designed to survive the kinds of modifications that would remove a simple label.
4. The SAG-AFTRA 2023 strike settlement regarding AI established what principle?
Correct! The SAG-AFTRA settlement established a consent-and-compensation framework: studios must obtain explicit informed consent and negotiate pay for any AI replication of a performer's voice or appearance — living or deceased.
The SAG-AFTRA settlement required explicit consent AND compensation — it established that performers' likenesses belong to them and cannot be AI-replicated without their agreement, regardless of whether they are living or deceased.
Module 3 · Lab 4

Deepfake Defense Lab

Apply verification habits and explore policy tools for defending against synthetic media

Your Mission

You're advising a news organization on deepfake defense policies. Use the AI to explore C2PA implementation, legal obligations under the EU AI Act, and how to train staff in verification habits. Ask at least three questions to complete the lab.

Starter prompts: "What would a deepfake defense policy for a newsroom look like?" · "Explain how C2PA provenance would work in practice for a journalist receiving a video tip." · "What are the limitations of the EU AI Act's labeling requirement?"
AI Lab Assistant
Deepfake Defense
Welcome, policy advisor! In this final lab we're focused on defense — the technical standards, legal frameworks, and individual habits that together form a resilient defense against deepfakes. Ask me about C2PA, SynthID, the DEFIANCE Act, the EU AI Act, SAG-AFTRA's consent framework, or how to build verification habits. What would you like to work on?
Module 3 · Final Assessment

Module Test: Deepfakes — Real or Not?

15 questions · Score 80% or higher to pass · All four lessons covered
1. Which AI architecture is most commonly associated with generating deepfake video?
Correct! GANs — with their competing generator and discriminator networks — are the foundational architecture behind deepfake video creation, though newer diffusion models are also used.
GANs (Generative Adversarial Networks) are the primary architecture for deepfakes — a generator creates fakes while a discriminator tries to detect them, driving ever-improving quality through competition.
2. The term "liar's dividend" describes:
Correct! The liar's dividend is the meta-harm of deepfakes: even when no fake was used, bad actors can point to the existence of the technology to discredit authentic evidence.
The liar's dividend refers to the ability of anyone to dismiss real video as "probably a deepfake" — a harm that exists even when no deepfake was actually created. The Gabon 2019 case is the clearest example.
3. The 2022 Zelensky deepfake was posted on:
Correct! The Zelensky deepfake was placed on hacked Ukrainian media sites to give it false credibility, then spread rapidly via Telegram and Facebook before platforms removed it.
The deepfake was injected into hacked Ukrainian TV and news sites — giving it apparent credibility as "official" Ukrainian media — and then spread across Telegram and Facebook.
4. The $35 million Hong Kong bank fraud (2020) used which type of deepfake?
Correct! This was audio-only fraud — criminals cloned a company director's voice to call a bank manager and authorize a massive wire transfer, demonstrating that voice clones are a serious financial threat.
This case used voice cloning — an audio deepfake of a company director — to authorize a $35 million bank transfer over the phone. No video was needed; voice alone was sufficient to commit the fraud.
5. What percentage of deepfake videos online (per Security Research Labs, 2023) depicted non-consensual intimate imagery?
Correct! Security Research Labs found roughly 96% of deepfake videos online depicted non-consensual intimate imagery — making it by far the most common deepfake harm, predominantly targeting women.
Security Research Labs documented that approximately 96% of online deepfakes depicted non-consensual intimate imagery — the overwhelming majority of deepfake harm is this category, not political disinformation.
6. Cross-platform laundering of deepfakes refers to:
Correct! Cross-platform laundering means a deepfake moves from Telegram to WhatsApp to Instagram to TikTok — each platform requires separate takedown efforts, and each reshare loses the original source context.
Cross-platform laundering describes how deepfakes spread from one platform to another, losing context each time and requiring separate removal efforts on each platform — greatly complicating moderation.
7. DARPA's MediFor (Media Forensics) program ran from approximately:
Correct! MediFor ran from 2016 to 2021, funded by DARPA to build AI tools for media integrity assessment. It produced capable detection tools while demonstrating the fundamental arms-race problem.
MediFor ran from 2016 to 2021 — launched just before the deepfake phenomenon went mainstream in 2017, and concluding as the adversarial training problem was well established.
8. Intel's FakeCatcher achieves detection using which biological signal?
Correct! FakeCatcher detects the subtle color variations in real skin caused by blood flow (rPPG). Deepfake-generated faces have no actual physiology, so they cannot replicate this signal.
FakeCatcher uses rPPG (remote photoplethysmography) — it detects color variations in skin caused by blood flow. Since synthetic faces have no blood, they cannot produce this signal naturally.
9. The University of Buffalo corneal reflection detection method identifies deepfakes by:
Correct! Real eyes reflect the environment in a consistent way governed by physics. Deepfake eyes often show incorrect or physically impossible reflections because the generator doesn't model environmental light precisely.
The corneal reflection method works because real eyes follow physical laws — they reflect the environment consistently. Deepfake generators approximate eye appearance but often get the reflections wrong in subtle ways.
10. Which cameras began embedding C2PA Content Credentials by 2024?
Correct! Sony and Leica were among the first camera manufacturers to embed C2PA Content Credentials at the hardware level — signing images cryptographically at the moment of capture.
Sony and Leica were the early camera adopters of C2PA. Their cameras embed a cryptographic signature into every photo at the moment the shutter fires, creating a chain of custody from capture.
11. The EU AI Act's requirement regarding deepfakes is that:
Correct! The EU AI Act requires transparency — deepfakes must be labeled as AI-generated. Violations can result in fines of up to 3% of global annual turnover. The rules take full effect from 2025.
The EU AI Act mandates labeling — synthetic and AI-generated content must be disclosed as such. It doesn't ban deepfakes outright but requires transparency and platform accountability, with significant financial penalties.
12. Google's SynthID watermark survives which types of modification?
Correct! SynthID is designed to be robust — the imperceptible watermark survives common modifications like screenshots, image compression (JPEG), and mild editing, making it harder to strip accidentally or casually.
SynthID is specifically designed to survive common content modifications including screenshots and compression — the kinds of transformations that would easily remove a simple visible label.
13. The SAG-AFTRA 2023 AI settlement established that studios must obtain what before using AI to replicate a performer?
Correct! The SAG-AFTRA settlement required both consent AND compensation — establishing that a performer's likeness is their property and cannot be AI-replicated without their informed agreement and fair pay.
The settlement required explicit informed consent plus negotiated compensation — both elements, not just one. This established the ethical framework: consent is necessary but not sufficient without fair compensation.
14. What is the best first step when you encounter a surprising video of a public figure on social media?
Correct! Source verification is the most reliable first step — if a significant video exists only on fringe accounts and hasn't been picked up by established media or official channels, that's a major red flag regardless of visual quality.
The most reliable first step is source verification — does this video appear on established, verified sources? Visual checks and detector apps are useful but secondary. Sharing while unverified helps spread potential misinformation.
15. Which of the following best describes the overall deepfake defense strategy recommended in this module?
Correct! The module's central lesson is that deepfake defense requires all three layers simultaneously: technical provenance infrastructure, legal accountability, and individual media literacy habits — because each layer has gaps that the others partially fill.
The module emphasizes a three-layer defense: technical standards, legal frameworks, and personal habits. Each layer has significant gaps, so resilience requires all three working together rather than relying on any single approach.