Modules/Module 8/Lesson 4
Lesson 4 of 7 ~10 min read

Deepfakes & Misinformation

Lesson 8.4 — Deepfakes: Detection, Reality, and Media Literacy

Close-up of a face with subtle digital distortion effects

The word "deepfake" entered common usage around 2017, when AI-generated face-swapping technology became accessible to people without specialist skills. Since then, the technology has advanced dramatically — and the range of harms it enables has grown with it. In 2024, AI-generated synthetic media is no longer a niche concern. It is a mainstream challenge that touches electoral politics, personal safety, fraud, and how we evaluate evidence.

This lesson covers what deepfakes are technically, how to detect them, the current reality of voice cloning, available detection tools, and the SIFT media literacy framework adapted for AI content.


What Deepfakes Are

"Deepfake" originally referred specifically to face-swapping technology (deep learning + fake), but the term now broadly covers:

  • Video deepfakes: A person's face replaced with another's in video; a person made to appear to say something they didn't say
  • Audio deepfakes: A synthetic voice cloned from a real person, capable of saying anything
  • Image deepfakes: Generated images of real-looking but non-existent people, or manipulated images of real people
  • Text deepfakes: AI-generated text designed to impersonate a specific person's writing style

The common thread is synthetic media designed to look, sound, or feel authentic.


Detection: Visual Tells

Deepfake detection is a moving arms race. As detection improves, generation improves, and vice versa. However, current deepfake video still has detectable tells in many cases:

TellWhat to look for
Facial edge blurringThe boundary between the swapped face and neck/hair often has a subtle blurring or shimmer
Eye movement anomaliesBlinking patterns may be irregular; eyes may not track naturally
Inconsistent lightingThe face and the environment may have different light sources or shadows
Hair and fine detailIndividual hairs and teeth are notoriously difficult to synthesise convincingly
Unnatural skin textureSkin may look unusually smooth or have a slight waxy quality
Head movement blurFast head movements often reveal glitching or warping at facial edges
Mouth sync errorsEven slight mismatches between mouth movement and audio are perceptible on close review
Reflection inconsistenciesReflections in glasses or eyes may not match the scene correctly
Background anomaliesBackgrounds can morph or blur during movement in some systems
Jewellery and accessoriesEarrings, glasses, and clothing details often behave oddly at frame edges

Important caveat: These tells are decreasing with every generation of technology. The best current deepfakes are not reliably detectable by the human eye, particularly in compressed social media video. Do not rely entirely on visual inspection for high-stakes authentication.


Audio Deepfakes: The Real Situation

Voice cloning has become alarmingly accessible and convincing. Here is the honest picture:

What is currently achievable:

  • A plausible voice clone can be created from as little as 30–60 seconds of audio using consumer tools (ElevenLabs, etc.)
  • Higher quality clones benefit from more audio but are still possible with limited samples
  • The clone can be made to say anything typed
  • Cloned voices are often indistinguishable from real voices in short clips, especially in emotionally charged contexts (urgency, distress)

Documented fraud case (2024): A finance executive at a multinational company received a deepfake video call involving clones of his CEO and CFO, requesting a $25 million transfer. The call appeared to include multiple colleagues simultaneously. The executive was defrauded. Multiple similar cases involving voice cloning of family members ("virtual kidnapping" scams) have been documented in 2023–2024.

What this means practically:

  • A phone call or voice message from a known person is no longer sufficient verification for a significant financial request or sensitive action
  • Establish a "safe word" or verification question with close contacts for unexpected high-stakes requests
  • Any unexpected request — even from a familiar voice — that involves urgency and money should be verified via a separately initiated call

Detection Tools

Several tools attempt to detect AI-generated content:

ToolWhat it detectsAccuracyNotes
Hive ModerationAI-generated images, video, textHigh for AI imagesAPI and web interface; used commercially
Sensity AIDeepfake video and face swapsHighEnterprise-focused; not free
Reality DefenderMulti-modal deepfake detectionHighEnterprise; used by news organisations
AI or NotAI-generated imagesModerateFree consumer tool
Deepware ScannerDeepfake videoModerateFree, consumer-facing
GPTZeroAI-generated textModerate; many false positivesUsed in education; accuracy debated

Honest assessment of detection tools: No tool is 100% accurate. Most detection tools generate both false positives (flagging real content as fake) and false negatives (missing actual deepfakes). They are useful as one signal, not as definitive proof. The detection space is also rapidly evolving — tools that work well today may underperform as generation technology advances.


The SIFT Media Literacy Framework Applied to AI Content

SIFT was developed as a media literacy framework by researcher Mike Caulfield. Originally designed for news and web content, it applies directly to AI-generated and synthetic media.

S — Stop

Before reacting to or sharing content, stop. The moment of first reaction — surprise, outrage, delight — is the moment you are most vulnerable to being misled. Pause before doing anything.

I — Investigate the Source

Who is sharing this content? What do you know about them? Is this from a source you have prior reason to trust? An unfamiliar account, an urgent share from an emotional message, or content from a new platform warrant more scrutiny.

F — Find Better Coverage

Can you find the same claim or event from multiple independent sources? If a shocking video of a public figure exists, would it appear in credible mainstream news outlets? If the content is true and significant, independent sources will typically cover it.

T — Trace Claims, Quotes, and Media to the Original Context

The most important step for deepfakes. Reverse image search (Google Images, TinEye) can find the original context of an image. Video can often be traced to an original upload. Quotes can be traced to their source. A video clip shown out of context — even of a real event — can be profoundly misleading without being technically fake.


Applying SIFT: A Worked Example

You receive a WhatsApp message containing a video clip of a politician appearing to say something shocking. The video is shared by someone you know but the source is unclear.

S — Stop: Resist the impulse to share immediately. The emotional reaction (outrage) is exactly what makes this kind of content spread.

I — Investigate the source: Where did this video come from? What account originally posted it? Is there any visible context (news watermark, date, event)?

F — Find better coverage: Search the politician's name and the supposed content on news sites. If this were real, it would be front-page news. If no credible outlet has covered it, that is significant.

T — Trace the media: Download the video still and reverse image search it. Check when the clip was first uploaded online. Watch the video carefully for the visual tells described above.

Key takeaway: SIFT is not about becoming a forensic investigator for every piece of media. It is about slowing down the moment between seeing and sharing, and building a habit of sourcing before amplifying.


What Platforms and Governments Are Doing

  • Meta, YouTube, and TikTok have policies requiring disclosure of synthetic media, particularly in political advertising
  • The UK Online Safety Act includes provisions targeting non-consensual intimate deepfakes
  • US legislation (No FAKES Act and others) is working through Congress to create legal liability for unauthorised deepfakes of real people
  • C2PA (Coalition for Content Provenance and Authenticity) is developing technical standards for embedding verifiable provenance information into images and videos at creation

These measures are imperfect and enforcement is challenging, but they represent a meaningful response to a documented harm.


Practice Task

Do a deliberate media literacy exercise this week. When you encounter a striking image, video, or quote online — even from sources you generally trust — apply SIFT. Trace the original source of one piece of content. This builds a habit that will serve you well as synthetic media becomes increasingly prevalent.