Synthetic Media: Deepfakes, Cheap Fakes and the Post-Truth World

Synthetic media, including deepfakes and cheap fakes, are having a profound impact on our world. As artificial intelligence and machine learning advance, it’s becoming easier than ever to generate convincing fake audio, images and video. This emerging technology is challenging our conception of truth and posing critical ethical questions.

In this comprehensive guide, we’ll explore what exactly synthetic media is, the different types and how they work, real-world examples, potential positive and negative implications, and what can be done to detect and fight disinformation while upholding free speech and expression.

What is Synthetic Media?

Synthetic media refers to artificially generated or manipulated digital content designed to misrepresent reality and deceive. This includes:

Deepfakes

• Uses AI and machine learning to swap faces or speech in existing video/audio clips to depict someone doing or saying something they didn’t.
• Produces highly realistic results that are difficult to detect as fake.
• Requires large datasets of images/videos and computing power.

Cheap Fakes

• Simple edits using basic software to alter timing, speed, captions, or image/audio splicing to misrepresent reality.
• Less sophisticated but still deceptive.
• Much easier and faster to produce than deepfakes.

AI-generated Images/Audio

• Images, video or audio generated from scratch by AI systems.
• Can create realistic fakes without needing original source material.
• Text or speech-to-image/audio models can fabricate images or audio from descriptions.

Algorithmically Generated Content

• Automated text generation from AI language models.
• Can produce news articles, social media posts, reviews based on limited inputs.
• Risk of generating misinformation and propaganda at scale.

While synthetic media has many positive applications, the potential for misuse to spread mis/disinformation poses risks to privacy, security, democracy and society. Understanding the nuances is key.

A Brief History of Synthetic Media

Humans have altered images and information to shape perceptions and narratives for centuries. What’s changed is the use of AI to automate generation and dramatically improve quality.

• 1990s – Basic digital edits using Photoshop started altering perceptions of beauty and truth in media.
• Early 2000s – Video editing software enabled new manipulations like crude face swaps.
• 2017 – AI research labs developed early deepfake algorithms to swap celebrity faces. Controversy erupted over potential misuse.
• 2018 – Easy-to-use deepfake apps emerged, allowing anyone to make deepfakes. Cheap fakes also proliferated.
• 2019 – Sophisticated AI text-to-image, text-to-audio and voice cloning models debuted, expanding synthetic media capabilities.
• 2020s – Ongoing improvements in quality and accessibility of synthetic media generation. Remaining technical limitations continue to fall.

While synthetic media itself is not new, the scale and sophistication AI enables presents novel opportunities and dangers. Understanding the evolution of this technology is key.

Deepfake Algorithms and Process

Deepfakes utilize complex AI systems, namely generative adversarial networks (GANs), to produce compelling forgeries. Here’s an overview of how they work:

• Uses 2 neural networks – a generator and a discriminator.
• The generator creates fake images/video. The discriminator tries to detect fakes.
• They train against each other to refine the generator’s outputs to be more realistic.
• Given enough data, the generator learns to produce highly convincing fakes that fool the discriminator.

To swap faces, deepfake algorithms undergo training and synthesis:

Training

• Gathers 100s or 1000s of images/videos of source and target faces.
• Analyzes facial features like landmarks, skin textures, expressions.
• Learns mappings between identities to build an encoding model.

Synthesis

• Takes new footage of source face.
• Encodes facial features and expressions from source footage.
• Maps encoded features onto the target face in realistic manner.
• Renders final fake video with target’s face.

This swapping of facial encodings allows the AI to produce videos of a target individual appearing to do and say things they didn’t actually do or say.

Types of Deepfakes

While face-swapping gets the most attention, deepfake algorithms can manipulate much more than just identities. Common categories include:

Face Swaps

• Swaps identities by transferring facial features and expressions from one person to another. Most widely known type of deepfake.

Puppeteering

• Animates facial expressions and mouth shapes of a target individual based on source footage of someone else. Target appears to mimic the expressions and lip movements of the source.

Speech Synthesis

• Generates fake audio of a target individual’s voice saying anything typed in. Only requires small audio samples.

Motion Transfer

• Transfers the bodily movements and gestures from one person to another to make a target individual appear to dance or move in imitation of the source.

Full Body Swaps

• Swaps the full bodies of individuals to place a target person into a scene they were not originally in.

AI Rendered Images/Video

• Generates fake imagery from scratch based on text descriptions using AI image and video generation models like DALL-E. No original footage required.

As the technology evolves, deepfakes allow for increasingly sophisticated manipulations well beyond simple face swaps. The possibilities are vast.

Deepfake Detection

While deepfake generation has grown more advanced, detection methods have also improved. Key techniques include:

• Visual artifacts – Low-level differences like inconsistent head angles, distorted features, blurriness.
• Implicit cues – Oddities like lack of blinking, limited head movement.
• Audio analysis – Unnatural voice tones and cadence.
• Media forensics – Checks for manipulated pixels, edits, splicing.
• AI assisted tools – Leverage AI to identify manipulated imagery and media.
• Metadata – Checks for inconsistencies in timecode, geolocation, device info.
• Firsthand accounts – Corroborating from direct witnesses and the individuals depicted.

No single method is perfect. Combining multiple detection strategies together provides the highest chance of identifying synthetic media fakes. Ongoing research aims to improve reliability as generation methods advance.

Deepfake Challenges and Concerns

Despite improvements in detection, deepfakes present many challenges:

• Disinformation – High potential for use in coordinated mis/disinformation campaigns.
• ** trust** in media and institutions.
• Impersonation – Identity theft for financial fraud, defamation, harassment.
• ** Privacy violations** – Nonconsensual use of likeness and private media.
• Blackmail – Coercion using fake incriminating media.
• Political instability – Manipulating speech/media of politicians and leaders.
• ** Psychological harm** – Trauma from victimization and abuse via synthetic media.
• Weaponization – Generating fake evidence, police brutality, atrocities for propaganda.

While deepfakes have some positive uses, the risks for abuse and harm necessitate caution and oversight.

Real World Deepfake Examples

Deepfakes have already been deployed in troubling ways:

• Nonconsensual adult videos depicting celebrities generated early controversy around deepfakes.
• Politicians like Joe Biden and Barack Obama featured in synthetic media to spread disinformation.
• Hoaxed deepfake video of Ukrainian president surrendering aimed to destabilize and mislead.
• Indian journalist Rana Ayyub subjected to deepfake porn videos in harassment and intimidation campaign.
• Cameroon separatist leader featured in deepfake calling for armed struggle, potentially inciting violence.
• Fraudsters used deepfaked voice of CEO to steal over $240,000 from energy company.
• Musical artists sued for unauthorized use of likeness to generate AI-synthesized media.

These examples offer just a glimpse of the wide range of synthetic media abuses emerging as the technology proliferates. The potential for harm is real and growing.

“Cheap Fakes” – Edited Media Manipulation

While deepfakes capture the most attention, simpler “cheap fakes” using basic editing and splicing techniques to manipulate timing, speed, context, captions and more account for the vast majority of synthetic media right now.

Examples include:

• Slowing down footage to imply intoxication or mental impairment.
• Speeding up speech to portray manic or reckless behavior.
• Splicing audio clips together to construct fake quotes.
• Removing context by cropping clips to mislead.
• Fabricating subtitles or captions to put false words in someone’s mouth.
• Manipulating timestamps to place individuals at fabricated events.

These edits can effectively shape perceptions and spread misinformation despite lacking sophisticated AI. Lower complexity also means cheap fakes can be produced much more quickly and easily than deepfakes.

Dangers of Cheap Fakes

Like deepfakes, cheap fakes pose serious dangers:

• Disinformation – Easily produced at scale for coordinated influence campaigns.
• Manipulating evidence – Manufacturing misleading footage of events.
• Defamation – Depicting offensive behavior or speech that never occurred.
• Inciting violence – Falsifying provocations or offensive acts to inflame tensions.
• Undermining institutions – Damaging trust in media, government, justice system.

While their simplicity makes cheap fakes easy to produce, it also aids detection. Still, the potential harms mean vigilance is vital.

AI-Generated Synthetic Media

Beyond manipulating existing media, AI systems can now generate stunningly realistic synthetic imagery, video, audio and text from scratch. Key examples:

AI Image Generation

• DALL-E, Stable Diffusion, Midjourney – Create images from text prompts and descriptions.

AI Video Generation

• Text-to-video models like Googel’s Phenaki – Generate video from text.

Voice Cloning

• Lyrebird, Resemble AI – Clone voice from < 60 seconds of audio.

Text Generation

• GPT-3, ChatGPT – Generate coherent articles, essays, tweets from text prompts.

Rather than doctoring authentic media, these models invent wholly fabricated synthetic media conditioned on text inputs. The results can be eerily convincing.

Capabilities and Risks of AI Synthetic Media

AI generation models enable unprecedented media manipulation abilities:

• Fully invented imagery – No need for original visual assets.
• Immense scale – Automated generation from text allows mass production.
• Personalization – Tailor fake media to specific targets.
• Invisible provenance – No visual artifacts from original media alterations.
• Difficult detection – Completely AI generated media leaves few clues of fakery.

These capabilities greatly expand the potential for abuse through disinformation, impersonation, harassment, fraud and more. The risks stretch far beyond deepfakes.

Fighting Back Against Synthetic Media

Combating synthetic media and its harms will require a multi-pronged response:

Improved Detection

• Better digital forensics and media authentication using a blend of technical and non-technical approaches.

Public Awareness

• Media literacy campaigns to build societal resilience and inoculate against manipulation.

Ethical Norms

• Promoting responsible development and use of generative AI. Allowing beneficial applications while curtailing abuse.

Updated Regulations

• Laws and policies to ensure consent in use of likenesses. Require source disclosure for synthetic media.

Corporate Responsibility

• Platform policies prohibiting nonconsensual deepfakes and explicit synthetic media. Flagging and removing harmful content.

International Cooperation

• Cross-border partnerships between governments, tech companies and civil society to share best practices.

Through collaborative action, the promise of synthetic media can be realized while mitigating emerging risks. There are no perfect solutions, but many promising pathways forward.

The Post-Truth Dilemma

The rise of synthetic media adds urgency to larger debates about truth, trust and verifiability in the digital age. Some key dimensions include:

“Truth Decay”

• More partisan, emotional appeals weakening fact-based discourse.

Declining Trust

• In government, media, institutions, and expertise.

Filter Bubbles

• Selective exposure to ideologically aligned sources.

Mis/Disinformation at Scale

• Coordinated influence campaigns made easier by digital media.

Information Overload

• Difficulty separating truth from fiction amidst torrents of content.

These trends, simmering for years, synergize with synthetic media to enable “post-truth” polarization where shared reality frays. Reversing this complex crisis requires addressing systemic issues of trust, media literacy, social cohesion and more. Technical fixes can help detect specific manipulations, but the underlying social challenges run much deeper.

Synthetic Media – Deep Concern, But Also Hope

Synthetic media represents an extraordinary breakthrough with the potential for both tremendous harm and tremendous good.

Like any powerful technology, synthetic media itself is neutral – neither inherently good nor evil. Everything depends on how it is used. Thoughtful governance, ethical norms and public vigilance are essential to steer this technology toward human thriving.

With care, diligence and cooperation, we can maximize the benefits of these emergent capabilities while also building societal resilience against harms. By approaching synthetic media with wisdom, caution and nuance, we can use it to enrich life and strengthen society.

The path forward is challenging, but not impossible. Guided by shared values of truth, transparency, consent, authenticity and trust, we can navigate this post-truth landscape.

Synthetic media forces difficult questions, but also presents boundless opportunities to reflect upon and renew our human connections. Our task now is to rise and meet this challenge with courage, compassion and creativity.