Your Guide to Understanding this Powerful, Concerning Technology
Have you seen those viral videos of Tom Cruise doing a funny coin trick on TikTok or Elon Musk promoting a bitcoin scam? Maybe you spotted that eerily lifelike music video featuring realistic dancers who seem just a little bit off?
Chances are, you’ve already encountered so-called “deep fakes” online. But do you really know what they are, how they work, and what dangers unchecked evolution of these systems might cause?
That’s where this guide comes in handy! I’ll walk you through deep fakes 101, delving into exactly what defines this emerging technology, its origins and development over recent decades, who’s creating deep fakes and for what purposes, risks experts have identified, and what everyday folks need to know to be responsible digital citizens in an age of advancing synthetic media.
Defining “deep fakes”
Let’s level-set on what constitutes a deep fake, since they can take many forms…
Deep fakes refer to multimedia – including photos, audio recordings, videos – that seem very realistic and unaltered, but have actually been manipulated using a set of artificial intelligence technologies collectively called deep learning systems.
These algorithms are structured to mimic the neural networks of the human brain. By analyzing tons of data points from real images, videos, or voice recordings of a target individual, deep learning systems can make judgments to synthesize incredibly convincing fake footage.
A 2019 survey found 48% of respondents were familiar with deep fakes. Of those in the know, over 50% felt deep fakes represented a threat to national security, with 46% saying they threaten democracy. [Source]
Unlike traditional CGI effects or face filters you’re probably used to, deep fakes leverage AI to precisely study patterns and details at a granular level. This allows inserting people into situations that never actually happened! Wild isn’t it?
And while Hollywood directors have created realistic computer graphics for decades, deep fakes are unique because anyone can now generate them without expensive equipment, just by using an app or web tool.
But are these just harmless fun with tech? What does it mean when our digital reality becomes nebulous? I’ll explore some bigger questions raised by increasingly accessible synthetic media later on…
First up though, let’s rewind and understand how we got here.
The evolution of deep fake tech
While deep fakes feel distinctly modern, the foundations of synthetic media stretch back centuries! Take a quick walk with me down memory lane…
1800s – The advent photography in the early 19th century enables doctored images through techniques like double exposures, airbrushing, photomontage. So while lasting visual records are new, manipulating depictions of reality emerges quickly too!
Early 1900s – Advances in cinematography, paired with camera and editing tricks, allow "special effects" in films like Fritz Lang‘s groundbreaking 1927 sci-fi Metropolis involving substitutions, multiple exposures and stop-motion animation. Technologies enabling altered and fabricated moving pictures evolve rapidly.
1990s – We enter the dawn of digital photo/video editing and computer-generated graphics. Researchers also begin publishing papers on neural network algorithms capable of analyzing and reconstructing human faces.
Early 2000s – Several milestones move facial reconstruction from theory towards reality. Machine learning and graphics cards enable swapping celebrity faces onto porn performers‘ bodies – often without consent.
2014-2016 – More advances make face-swapping simpler for creators and viewers. A Reddit user nicknamed “Deepfakes” emerges, pioneering easy-to-use fakes of celebrities‘ faces swapped onto adult video actresses.
2017-2019 – The machine learning community wakes up to how rapidly evolving deep learning could empower realistic synthetic media and create chaos in public communications. Concerns grow about non-consensual face-swapping and porn.
2020-Present – User-friendly mobile apps like Zao launch, allowing anyone to make decent deep fakes on phones. High profile viral examples help increase public awareness. But risks remain high as bad actors exchange tips in dark web forums.
According to a 2021 analysis of 9,000 deep fakes from media forensics company Sensity, pornography accounted for 90% of samples, 96% were nonconsensual, and over 60% leveraged images of female victims.
So in just 30 years, we‘ve gone from lab theories to viral self-made videos spreading across social platforms. And systems keep advancing…
Which leads us to how deep fakes like those Tom Cruise videos are actually created using AI. Let‘s unpack that process next!
How advanced AI generates deep fakes
While manually altering photos, video or audio has always taken skill, these days artificial intelligence handles the heavy lifting. Specifically, deep fakes rely on two key AI capabilities:
- Representation learning – algorithms that analyze source images of a person to understand and encode patterns.
- Generative modeling – technology that can reconstruct new synthetic faces or voices that seem real.
Let‘s break down how deep learning systems produce their convincing fakes…
Step 1: Gathering data & encoding patterns
The AI “studies” hundreds or thousands of images, videos and audio clips featuring a target individual – say, Kim Kardashian or Barack Obama – from varied angles and lighting conditions.
By surfacing common details of their facial geometry, skin textures, speech rhythms and tones, the algorithm learns to mathematically represent the core visual and auditory essence of that person.
Step 2: Generating synthetic elements
A second algorithm called a Generative Adversarial Network (GAN) then creates new images, footage and audio that don’t exist in reality, but closely retain patterns identified from the real data.
GANs start with pure noise signals and reshape output until it convincingly mimics authentic data. The encoding network double checks GAN-created elements for realism.
Step 3: Repeating and refining
The GAN keeps iterating freshly forged outputs, while getting feedback on deficiencies from the encoder’s comparisons against source data.
Over enough cycles analyzing where its fakes still fall short + strengthening those areas, the GAN evolves the ability to produce eerily lifelike synthetic data – faces, voices and footage capturing a person’s exact essence.
This adversarial back-and-forth teaches the model greater precision than either network alone. Given enough real samples and compute power, GANs can fabricate shockingly credible media.
One 2021 study found that across 15 common deep fake detection methods evaluated, highest accuracy was just 65%, meaning even experts struggle to reliably identify AI-doctored video. Performance will likely decline as algorithms improve.
Now you know a bit about how deep learning manages to create such realistic illusions! Next let‘s explore how these systems could bring positive change…as well as peril.
Deep fakes could bring creative breakthroughs…and chaos
Like any powerful innovations, deep learning and GANs carry promise to profoundly impact fields like healthcare, education and safety. Applied ethically, they might:
-
Preserve history – Manipulate archival footage to salvage damaged films or create interactive educational experiences bringing historical figures to "life"
-
Protect privacy – Anonymize faces and voices of whistleblowers or vulnerable citizens to share their narratives safely
-
Revolutionize creativity – Unlock immersive new mediums of film, animation, music and more as production costs plummet
However, without caution, oversight and consent, deep fakes often degrade trust, law and order. Potential dangers include:
Disinformation – High-profile deep fakes imitating leaders making inflammatory remarks, illegally influencing election campaigns, financial markets or court trials
Impersonation – Faces/voices inserted into situations without permission to bully, smear reputations or even elicit criminal charges
Financial crime – Forged audio instructing employees to wire huge payments they believe are from the CEO
Harassment – Faces superimposed into revenge porn, personalized bullying content or ads for damaging products/services
And that‘s just the start – the list multiplies quickly!
Researchers estimate malicious uses of synthetic media doubling every six months. Over 96% of deep fakes leverage stolen personal images non-consensually, disproportionately targeting women.
Let‘s walk through a few real world examples to drive home what‘s at stake…
Real harms: Deep fakes in the wild
- Politics – A 2019 video of Nancy Pelosi distorted to make her speech seem slurred went viral on Facebook with over 2.5 million views, likely impacting public perception.
- Celebrities – Scarlett Johansson, Emma Watson and other actresses had deep fake porn videos made without consent, infringing privacy.
- Business – Criminals used AI to imitate a corporate executive‘s voice and authorize a fraudulent $243,000 wire transfer.
- YouTube – Impersonation channels like “Futuring Machine” rack up millions of views spreading misinformation about historical events.
- Dating apps – Stolen images are showing up in fake deep fake profiles, catfishing victims searching for intimacy and relationships.
And these are just a few examples – multiply them by thousands and you‘ll get a picture of how rapidly deep fakes are permeating digital spaces, often causing real life destruction.
So with potential for so much harm, how can everyday folks spot potential deep fakes online? What do we need to know as responsible citizens navigating an increasingly synthetic media landscape?
Great questions, I‘m so glad you asked! Let‘s get into it…
How to spot potential deep fakes
Catching deep fakes that algorithms specifically design to seem authentic is tricky, even for forensic experts.
In controlled tests, researchers found it takes even experienced professionals over 21 times longer to identify AI-doctored video versus manually edited footage.
But with a trained eye, subtle inconsistencies can reveal manipulation. Here are tips from investigative journalists and digital literacy educators:
Inconsistencies to look for
- Pixelated blending along edges
- Mismatching skin tones/textures
- Lighting directions not aligned
- Strange artifacts or distortions
- Unnatural, repetitive eye blinking
Best practices
- Pause, rewind, slow down video
- Zoom in on finer details
- Cross-reference multiple sources
- Leverage forensic detection tools
- When in doubt, don‘t spread!
However, this cat and mouse game can’t last forever. Rapid improvements will likely outpace human detection capacity in the near future.
“Trying to solve this problem just by perfecting detection is a losing battle,” cautions Synthetic Media expert April Smith. “With enough data, systems that generate photorealistic non-existent people and events from scratch could unlock unprecedented potential for predatory deception.”
So beyond biometric checks, we need broader change. Building societal resilience requires:
Education – to improve technical literacy and critical thinking when assessing media authenticity across communities.
Awareness – of deep fakes’ dissemination across channels like newsfeeds, messaging apps and forums where predators roam.
And most crucially…
Accountability – legally enforceable ethical guidelines and restrictions around development/use of synthetic media technology.
If creators and distributors of powerful systems like deep learning cannot act conscientiously, situations that once strained credulity may rapidly become permanent reality.
And with that, we‘ve reached the end of our deep dive! I aimed to decode a paradigm-shifting technology, spotlight promising applications alongside serious risks, provide tips for spotting deception attempts, and argue safeguards must be implemented to ensure truth and trust survive inevitable future advances.
Let me know if you have any other questions arising from our discussion around deep fakes and artificial intelligence! Whether you feel cautiously optimistic or seriously concerned, I‘m here to chat through perspectives.
