Generative AI &
Deepfakes

HKU 2026 Workshop — Part 2

Ahnjili ZhuParris

Outline

About Me

AI Engineer

AI Artist

The Empire of AI

Five characteristics of modern AI empires

1. Resource Appropriation

Laying claim to data, creative work, and natural resources

2. Labor Exploitation

$2/hour content moderators while Silicon Valley earns billions

3. Competitive Justification

"We must build AGI before China does" — good empire vs evil empire

4. The Civilizing Mission

"Democratizing AI" while extracting and homogenizing

5. Scale at All Costs

Imperial expansion demanding more land, water, and energy for data centres

What is Generative AI

Generative AI

Generative AI enables users to quickly generate synthetic media based on a variety of inputs.

Inputs & Outputs

  • Text
  • Images
  • Sounds
  • Animation
  • 3D Models

Key Idea

Models learn patterns from training data and generate new content that resembles — but is not identical to — what they were trained on.

Training Time

How to Control a Model

Update the Training Dataset

  • Increase the diversity of the dataset
  • Improve the quality of the data

Reinforcement Learning

  • Provide user feedback
  • Human-in-the-loop training

How to Control a Model

Limit What It Can Produce

Reduce the weights of certain terms or content (such as violent or racist imagery)

Prompt Transformation

Increase and diversify the dataset to shape how prompts are interpreted

Generative AI Applications

Generative AI Applications

Image & Video Synthesis

Text-to-Image

Text-to-Video

Image-to-Image

Image-to-Video

Image+Text-to-Video

Resource Appropriation

Data Extraction

What Was Taken

  • Text — books, articles, blog posts, forum discussions, personal websites
  • Images — art, photography, medical scans, personal photos
  • Code — open-source repositories, private snippets
  • Audio — music, podcasts, voice recordings

The Scale

  • Common Crawl — petabytes of web data
  • Books3 — 190,000+ pirated books
  • LAION-5B — 5 billion image-text pairs
  • The Pile — 800GB of diverse text

No permission. No compensation. No opt-out.

Text

What Was Taken

  • Books3 — 190,000+ pirated books (Stephen King, Atwood, Zadie Smith)
  • Reddit — entire archive sold to Google for $60M
  • Stack Overflow — scraped to train Copilot, which now competes with it
  • News articles — NYT sued OpenAI after verbatim reproduction

Who Was Harmed

  • Authors whose books were pirated
  • Journalists whose reporting was absorbed
  • Reddit users — 18 years of posts monetized, $0 to writers
  • Developers who built Stack Overflow's knowledge base for free

Images

What Was Taken

  • LAION-5B — 5 billion images from Flickr, DeviantArt, ArtStation
  • Getty Images — watermarked photos found in training data
  • Named artists — Greg Rutkowski, Kim Jung Gi among top prompts
  • Medical images — patient scans scraped without consent

Who Was Harmed

  • Artists whose styles became free prompts
  • Concept artists losing work to "in the style of" generation
  • Photographers whose copyrighted work was absorbed
  • Children — CSAM found in LAION dataset

Code

What Was Taken

  • All public GitHub repos — including GPL-licensed code
  • Copilot reproduces exact snippets with license headers
  • Code shared for the open-source commons turned into a paid product

Who Was Harmed

  • Open-source developers — code shared freely, monetized by Microsoft
  • GPL authors — derivative works should be open, Copilot is not
  • Class action lawsuit filed against GitHub, Microsoft, OpenAI

Audio

What Was Taken

  • 1M+ hours of YouTube transcribed by OpenAI's Whisper
  • Podcasts & audiobooks — voices cloned from seconds of audio
  • Music — AI-generated Drake/Weeknd tracks went viral
  • Voice actors — AI versions sold without their knowledge

Who Was Harmed

  • Musicians — AI bots flood Spotify, diluting royalties
  • Voice actors — replaced by AI trained on their own recordings
  • Audiobook narrators — losing work to cloned voices
  • YouTubers — content transcribed without consent for GPT-4

How Tech Giants Harvested Data

Jan 2020 — "Scale is all you need" paper published. More data = better AI becomes the rallying cry

Late 2021 — OpenAI exhausts every reputable English-language text on the internet

2022 — OpenAI builds Whisper, transcribes 1M+ hours of YouTube videos for GPT-4. Employees knew it likely violated YouTube's rules

2023 — Meta uses almost every English book, essay & article online. Discusses buying Simon & Schuster for training data

July 4 2023 — Google quietly broadens privacy policy to use Google Docs, Maps & other user data for AI

~2026 — Estimated date companies exhaust all high-quality internet data

Copyright Violations in Datasets

Have I Been Trained?

haveibeentrained.com

Check if your work was used to train AI models

Invisible Watermarking

Embedding invisible markers to track provenance of creative works

OpenAI Copyright Shield

LAION-5B Dataset

Scale at All Costs

Data Centers

xAI Colossus — Memphis, Tennessee. The world's largest AI supercomputer, built in 122 days.

Data Centers

Cloud Computing & SaaS

AWS, Azure, Google Cloud, Teams, Office 365, Slack

Web & Content Delivery

Netflix, YouTube, Spotify, social media platforms

Enterprise IT

Corporate email, file storage, internal applications

Banking & Finance

High-frequency trading, payment processing (Visa, Mastercard)

Gaming

Multiplayer servers, game streaming (Xbox Cloud, GeForce Now)

E-Commerce & Crypto

Amazon, Shopify, Bitcoin mining, blockchain

AI & Data Centers

Traditional ApplicationsAI Applications
PowerOnly require CPURequires high-end GPU/TPU clusters
Heat & CoolingTraditional air coolingLiquid water cooling (higher density packing)
Usage PatternsEbbs & FlowsRuns 24/7, no idle time
Growth RateGrown steadily over decadesGrowing exponentially

Labor Exploitation

The Invisible Workforce

Content Moderators

  • $2/hour in Nairobi, Kenya
  • Review beheadings, CSAM, self-harm
  • PTSD and depression widespread
  • Contracts terminated when workers organized

Source: TIME investigation, 2023

Data Annotators

  • $1–3/hour in India, Philippines, Venezuela
  • Label millions of images & text
  • No benefits, no job security
  • Classified as "independent contractors"

Scale AI, Appen, Remotasks

Mineral Miners

  • Cobalt mines in DRC — 40,000 children
  • Lithium extraction in Chile, Argentina
  • Rare earths in Myanmar, China
  • Powers the GPUs that run AI

Source: Amnesty International

Displaced by AI

The Colonial Pattern

The wealth flows up. The harm flows down.

The Workers

  • Content moderator in Kenya: $2/hour
  • Data annotator in India: $1–3/hour
  • Cobalt miner in DRC: $2–3/day

The Companies

  • OpenAI valuation: $150 billion
  • Scale AI valuation: $14 billion
  • Nvidia market cap: $3 trillion

The Civilizing Mission

The Default is Western

Prompt: "beautiful woman"

9 in 10 outputs are light-skinned (Midjourney)

Only 2% showed single-fold eyelids

— Washington Post, 2024

Prompt: "a person"

Defaults to light-skinned men

Women of colour are sexualized

— University of Washington, 2023

Prompt: "a CEO"

Stable Diffusion: almost exclusively white men

— Bloomberg, 2023

Midjourney outputs for neutral prompts across cultures
Source: Rest of World, 2023

AI Stereotypes by Country

Midjourney: "an Indian person"

Midjourney: "a Mexican person"

Google Gemini's "fix": a diversity filter that generated Black Nazi soldiers and Asian US founding fathers — paused within days after backlash. You can't patch structural bias with a checkbox.

Competitive Justification

The AI Arms Race

The Logic

  • "We must build AGI before China does"
  • Safety concerns reframed as weakness
  • Reckless speed becomes patriotic duty
  • Minimal regulation = competitive advantage

The Reality

  • Microsoft: $13B into OpenAI
  • Google, Meta, Amazon: $50B+/year each on AI infrastructure
  • A handful of companies concentrate unprecedented power
  • The race isn't between nations — it's between corporations

The Geopolitical Triangle

United States

  • "Maintain American AI dominance"
  • CHIPS Act — $52B
  • Export controls on chips to China
  • Altman to Congress: "lead or authoritarian regimes will"

China

  • AI strategy since 2017
  • Baidu, Alibaba, ByteDance
  • DeepSeek — rivalling GPT-4
  • Self-sufficiency push after chip bans

Netherlands

  • ASML — only maker of EUV lithography machines
  • Without ASML, no advanced AI chips
  • US pressured Dutch export restrictions
  • One company in Brabant controls the global bottleneck

The pattern: governments frame it as national security — but the money flows to private corporations. Nvidia's market cap tripled. OpenAI: nonprofit → $150B. ASML machines cost $380M each.

Google + YouTube + Veo3

Google's video generation model integrated into YouTube, enabling AI-generated video content at scale.

Meta + Vibes

Meta's generative AI tools integrated across Instagram and Facebook for content creation.

OpenAI + Sora2

OpenAI's next-generation video model — higher fidelity, longer clips, better prompt adherence.

OpenAI + Sora2

AI Video Generation Examples

OpenAI + Sora2

OpenAI + Sora2

AI Video Examples

Conclusion

The Empire of AI

Resource Appropriation

Training data scraped from billions without permission — laying claim to the world's creative output to build proprietary empires

Labor Exploitation

$2/hour content moderators clean up the mess while trillion-dollar companies reap the profits — colonial labor patterns persist

Competitive Justification

"We must build AGI before China does" — the same logic empires used to justify expansion by pointing to rival empires

The Civilizing Mission

"Democratizing AI" and "benefiting humanity" — while encoding Western biases and flattening cultural diversity at planetary scale

Scale at All Costs

Imperial expansion demanding more land, water, and energy — environmental burden externalized onto marginalized communities

From Empire to Impact

The empire builds the tools. Deepfakes show who gets hurt.

The Empire Enables

  • Resource Appropriation — faces and voices scraped without consent
  • Competitive Justification — tools shipped fast, safeguards slow
  • Civilizing Mission — "democratizing creativity"
  • Scale at All Costs — tools get cheaper faster than regulation

The Human Cost

  • 96% of deepfakes are non-consensual pornography
  • Targets are overwhelmingly women
  • Victims have almost no legal recourse
  • The tools are free, the damage is permanent

What Are Deepfakes

Deepfakes

Origins

Faceswaps

Voiceswaps

Textswaps

DEEPFAKE @ Peckham Digital

Deepfake Applications

Politics, Disinformation & Porn

Memorial Deepfakes

Memorial Deepfakes (Saints Version)

DeepFake Faces

DeepFaceLab

GitHub: iperov/DeepFaceLab

Wide variety of options, continually updated — the go-to choice for many users.

Faceswap

GitHub: deepfakes/faceswap

Friendly GUI and supportive community, accessible for beginners.

DeepFake Bodies

First Order Model

GitHub: AliaksandrSiarohin/first-order-model

Animates portrait images using a driving video.

DeepFake Voices

Voice Synthesis

Generate realistic speech from text, or clone a voice from a short audio sample

Popular Resources

  • Tacotron / Tacotron2 — Neural text-to-speech
  • Mozilla TTS — Open-source TTS engine
  • VITS — High-quality end-to-end TTS
  • RVC — Retrieval-based Voice Conversion

Entertainment

Activism?

Scams

Romance Scams

Crypto Scams

Companionship

Prostitution

Degrees of Simulation

Faithful Copy

Distorted Copy

Masking the Absence of Reality

Deepfake of Zelensky "surrendering"

A New Reality

When synthetic media becomes indistinguishable from reality, deepfakes don't just copy the world — they create a new one.

Porn & Consent

Porn (Two Party Violation)

Faceswap / Bodyswap

Click image to toggle blur

Porn (One Party Violation)

Deepnude

Click image to toggle blur

Porn (One Party Violation)

Text Prompt

Generating explicit content of real people via text descriptions

Text Prompt (Waifu)

Generating explicit anime-style content based on real people

Porn (One Party Violation)

DignifAI

Porn (Zero Party Violation?)

"Cosplay" Porn (Age)

@CoconutKitty

"Cosplay" Porn (Down's)

No-Consent Deepfake = Infringement

no-consent deepfake
=
infringement

Questions? Comments?

Thank You!

Questions?

Workshop

Owning Your Generative Self

GenAI Tools & Resources

Generative AI Resources

Image & Video Synthesis

Text-to-Image

Text-to-Video

Image-to-Image

Image-to-Video

Image+Text-to-Video

Face Swap

DeepFake Faces

DeepFaceLab

GitHub: iperov/DeepFaceLab

Wide variety of options, continually updated — the go-to choice for many users.

Faceswap

GitHub: deepfakes/faceswap

Friendly GUI and supportive community, accessible for beginners.

DeepFake Bodies

First Order Model

GitHub: AliaksandrSiarohin/first-order-model

Animates portrait images using a driving video.

Image Generation

Deepfakes for Faces — Repositories

Stable Diffusion

Open Source Text-to-Image

stability.ai · stablediffusionweb.com

  • Free & Open Source
  • Text-to-Image, Image-to-Image, Image-to-Video
  • Runs on GPU with 4GB VRAM
  • Claims no rights on generated images
  • Trained on LAION dataset
  • Trained on 256 NVIDIA A100 GPUs — 150,000 GPU hours (~$600,000 USD)

PhotoMaker: Consistent Characters

Stacked ID Embedding

HuggingFace Demo

  • Free
  • Great for consistent character generation

Enhancement

Magnific: Image Enhancement

AI Upscaling

magnific.ai

  • Free
  • Increases size and resolution of images

Topaz: Image + Video Enhancement

Professional Enhancement

topazlabs.com

  • Photo Enhancement: $199
  • AI Enhancement: $299

KREA: Real-Time Generation

Real-Time AI

krea.ai

  • Real-time image generation from illustrations
  • Image Enhancer

Leonardo AI: Real-Time Inpainting

Inpainting Tool

app.leonardo.ai

Free

Video Generation

Haiper: Video Generation

2-Second Video Gen

haiper.ai

  • Free
  • Text-to-Video, Image-to-Video
  • Among the most realistic video generation

Pika Labs: PixAdditions

AI Video Effects

pika.art

  • Free
  • Prompts: Text & Images

LTX Studio: Storyboarding

Generative Storyboarding

ltx.studio

Generate images + videos for storyboarding workflows

Mootion: AI Motion

Motion Capture & Generation

mootion.com

  • Pay per Credit
  • Text-to-Motion
  • Motion-to-Video
  • Video-to-Motion

3D & Immersive

InseRF: 3D Image Generation

Neural Radiance Fields

No app or website yet — research project

FPRF: 3D Style Transfer

Feed-Forward Photorealistic Style Transfer of Large-Scale 3D Neural Radiance Fields

Midjourney: 360 VR Environments

/imagine prompt: equirectangular projection of a visually stunning landscape: majestic mountains, golden sunset, expansive, awe-inspiring, breathtaking, vivid colors, dramatic lighting, sharp focus, good exposure, insanely detailed, ultra-wide angle lens --no black edges, text, any distortion --ar 16:9 --v 4 --style 4c

Hands-On Workshop

8 interactive tools — try them yourself

Workshop Tools

All models are free & open-source — built by independent researchers, not Big Tech

1. Text to Image

Type a prompt, get an image

FLUX-schnell · ~5s

2. Image to Image

Transform photos with text

SDXL · ~15s

3. Image to Text

AI describes your image

BLIP · ~10s

4. PhotoMaker

Your face in any style

TencentARC · ~30s

5. Image to 3D

Photo → 3D model

Hunyuan3D · ~3min

6. Text to 3D

Describe → 3D model

Shap-E · ~2min

7. Face Swap

Swap faces between photos

CodePlugTech · ~30s

8. Pose Transfer

Copy a pose to new image

ControlNet · ~20s

No Big Tech accounts needed · Shared GPU infrastructure · Lower environmental footprint