Business

Artificial Intelligence and the ‘Gods Behind the Masks’

In an excerpt from AI 2041: Ten Visions for Our Future, Kai-Fu Lee and Chen Qiufan explore what happens when deepfakers attack the deepfakes.

In AI 2041: Ten Visions for Our Future, AI expert Kai-Fu Lee and coauthor Chen Qiufan answer the question “How will artificial intelligence change the world over the next 20 years?” Lee’s technical explanations sit alongside Chen’s fictional short stories to produce an exploration of the perils and possibilities of AI. This story revolves around a Nigerian video producer who is recruited to make an undetectable deepfake. Touching on impending breakthroughs in computer vision, biometrics, and AI security, it imagines a future world marked by cat-and-mouse games between deepfakers and detectors, and between defenders and perpetrators.

As the light-rail train inched into Yaba station, Amaka pushed a button next to the door of his carriage. Even before the train came to a complete stop, the doors opened with a whoosh and Amaka hopped off. He couldn’t tolerate the slow trains—or their stale odor—for another second. Following closely behind an elderly man, Amaka nimbly slid through the turnstile at the station’s exit. Facial recognition cameras were meant to deduct the fare as each person passed by. Thanks to the mask that veiled Amaka’s face, however, he slipped out without charge. 

Such masks had become commonplace among the young people of Lagos. For their parents’ generation, masks were ritual objects, but for the youth, whose numbers had swelled in recent decades, they had become fashion accessories—and surveillance avoidance devices. Lagos, the largest city in West Africa, was home to somewhere between 27 and 33 million people—the official number depended on what method the authorities used to measure it. Five years ago, the state imposed a strict limit on the number of migrants entering the city—even those, like Amaka, who were born in other parts of Nigeria. Since then, itinerant dreamers like Amaka had been forced to seek makeshift shelter in illegal apartments, hostels, markets, bus stations, or even under over-passes. He had met many homeless people, people who had been driven onto the streets for all kinds of reasons: people whose homes had been demolished to make way for new shopping centers, people newly arrived in Nigeria from worse-off nations, and those who were simply poor. Nigeria’s youthfulness—the nation’s median age was merely 21—was thanks to the nation’s high fertility rate. Still, the rapid development of the world’s third-most-populous nation had not benefited its citizens equally. 

While other parts of Lagos strained under the pressure of its young population, the Yaba district was flourishing. Dubbed “the Silicon Valley of West Africa,” the neighborhood stood out for its orderliness, fresh air, and high-tech-infused daily life. Pedestrians could activate the cartoon animals on the billboards and interact with them via hand gestures. Cleaning robots roamed the streets, collecting and sorting trash, then sending it off to recycling centers where it was turned into renewable materials and biofuel. Sustainable bamboo fiber had recently made the leap from building material to fashion trend, at least for the denizens of Yaba.

Standing outside the station and holding his smartstream up to eye level, Amaka overlaid a live virtual route map onto the surrounding streetscape. Following the projected route, he began walking, eventually stopping before a gray building emblazoned with the number 237 and tucked away on a quiet backstreet. The company he was looking for, Ljele, was apparently based on the third floor. Two days ago, he had received a mysterious email from an anonymous Ljele account about a job that was “right up his alley.” The position was his under the condition of his showing up for an interview in person.

As Amaka entered a small reception area on the third floor, the receptionist smiled and pointed to Amaka’s mask, indicating he should remove it for an identity check. The young man hesitated, then took his mask off. Reflected in the camera lens was a young, smooth face. His 3D-printed mask couldn’t match the delicate quality of the pricey handmade versions sold at absurd prices to tourists in the Lekki Market, but the coarse reproduction, with its butterfly-like pattern, was enough to fool the facial recognition algorithm of most common surveillance cameras. In the eyes of AI, Amaka was a “faceless person.” The mask not only saved him money, but, more important, shielded him from the authorities. After all, Amaka had yet to obtain a migrant residence permit.

When the face scan was completed, the receptionist brought Amaka into a conference room and told him to wait. He sat stiffly as he pondered how he would answer questions regarding his previous work experience. I have to lie, he realized. I don’t have many other choices. 

Ten minutes passed. The promised interviewer did not appear. Abruptly, the projection wall across from him lit up, and surveillance camera video footage began to play.

To Amaka, the video footage was as familiar as the back of his own hand. Midnight. Dim, yellow streetlamps. Several homeless people were scattered under an overpass, lying on makeshift mattresses. The silhouette of a boy emerged from the shadows. The boy walked over to a group of sleeping people and gazed down. The camera zoomed in. The boy was white, no more than 5 or 6 years old, dressed in striped pajamas, his face wan and expressionless. One of the people woke up with a start and met the boy’s eyes. The homeless man asked the boy what his name was and where he lived. The boy’s body trembled as he mumbled incoherently. Suddenly, his face twisted, the corners of his lips stretching open and revealing two rows of sharp teeth. He bit down hard on the homeless man’s neck. The man cried out in pain, waking up the others. The boy fled the scene, blood trickling down his lips and chin. 

The video, originally posted to the internet under the title “White Vampire Boy Attacks Homeless People in Lagos,” had received millions of views within 24 hours of its first appearance on the GarriV video-sharing platform. Within days, however, the platform identified the video as a fake and removed it in compliance with the law. The uploader’s account, “Enitan0231,” was consequently terminated, with all its associated advertising revenue frozen.

Suddenly, a booming voice filled the conference room where Amaka still sat, alone. “Well done, Amaka! What a seamless fusion of realistic settings, amateur actors, and live video shooting. I can’t believe you made this in an underground internet café in Ikeja,” said a man’s voice with a heavy Igbo accent. 

Instinctively, Amaka jumped to his feet. “Who are you?” His eyes surveyed the empty room and landed on the speakers.

“Hey, relax. You can call me Chi. Do you want a job or not?” 

Sighing, Amaka sat back down and slouched in the chair. The man named Chi was right. Without a residence permit, he could never find a real job in Lagos. The mysterious Ljele company was his only sliver of hope. “Why me?” he asked. 

“We saw your work. You’re talented. You’re ambitious—you wouldn’t have come to Lagos in the first place if you weren’t determined to make a name for yourself. Most importantly, we need someone we can trust. One of our own kind.” 

Amaka knew immediately what Chi was alluding to. Nigeria has more than 250 ethnic groups, with their own languages and customs, many of which had been in conflict for hundreds of years. The Yoruba and the Igbo, respectively the second- and third-largest ethnic groups in the country, had seen violent clashes in recent years, as both groups muscled for political gain. With the Yoruba as the dominant population in Lagos, Amaka, an Igbo from the southeast, usually concealed his ethnicity to avoid trouble. “What do you want me to do?” 

“I want you to do what you do best. Fake a video.” 

Illegally, I presume?” 

“We’ll supply you with all you need.” 

Amaka narrowed his eyes, his nostrils flaring. “And what if I turn down your offer? Will you kill me?” 

“Kill you? No, no. Worse than that.” 

Another video started to play on the projection wall. A dance floor in a private nightclub. The camera zoomed down on the room from a corner of the ceiling. Several boys were dancing up against one another under the flashing laser lights, shirtless. The camera zoomed in farther to reveal the unmistakable face of Amaka. As the camera observed, Amaka turned and passionately kissed another boy whose cheeks glowed fluorescent pink. Amaka then twisted his upper body around to kiss a darker-skinned boy behind him. The video froze on this frame. The three young faces were like mango leaves that overlapped, intertwined and merged into one another. 

Amaka stared at the video, his expression blank. After a few moments, he grinned. The facial scan he had undergone back at the reception desk had provided the data to make this instantaneous deepfake. 

“The face might be mine, but not the neck,” said Amaka as he pulled down his hood, exposing a long pink scar that cut diagonally from below his right ear to his left collarbone. A souvenir from a street fight. “Also, don’t forget we’re in Lagos. The things people do here are far crazier than that.”

“Sure, but this video can still send you to prison. Think about your family,” said Chi, his voice turning soft. 

Amaka fell silent. Three decades after the passage of the Same Sex Marriage (Prohibition) Act of 2013, Nigerian society remained just as hostile toward sexual and gender minorities as it ever had been. If someone reported him, Amaka knew it would be difficult to avoid dealing with the corrupt police, who would likely try to extort him, even if he could avoid criminal charges. 

And then there was his family. While they hadn’t had an easy relationship these past few years, Amaka hated to imagine the pressure that could descend upon the shoulders of his family members, especially his father, who expected the world of him. Even if the video is a fake. 

The boy bit on his lower lip and pulled his hood back up. Concealing parts of his skin again gave him a little more sense of safety. “I need an advance payment. Cryptocurrency. Also, give me as much detail as you have on the target. I don’t want to waste my time on research.”

“Your call, my friend. As for the target … there’s absolutely no way for you to miss him.” The blurred headshot of a man flashed on the projection wall. When the face’s contours solidified into a clear picture, Amaka’s eyes widened.

The Yoruba called the city of Lagos “Eko,” meaning “farm.” In the equatorial monsoon climate, June was the coolest month, with the most plentiful rain. With the rain’s monotonous tapping on the metal roof as his background soundtrack, Amaka lay on the small bed of his illegal hostel room. He put on his XR glasses and fiddled with his new gadget—a dark-green Illumiware Mark-V. 

Compared with the pranks he had carried out in the past, this new job was on a completely different level. 

It’s not that he lacked experience in video deception—quite the opposite. Alone in his room, Amaka had spent many nights of the past year disguising himself as uptown girls on dating apps. In order to construct a flawless imitation, the first step was to gather as much video data as possible with a web crawler. His ideal targets were fashionable Yoruba girls, with their brightly colored Vneck buba and iro that wrapped around their waists, hair bundled up in gele. Preferably, their videos were taken in their bedrooms with bright, stable lighting, their expressions vivid and exaggerated, so that AI could extract as many still-frame images as possible. The object data set was paired with another set of Amaka’s own face under different lighting, from multiple angles and with alternative expressions, automatically generated by his smartstream. Then, he uploaded both data sets to the cloud and got to work with a hyper-generative adversarial network. A few hours or days later, the result was a DeepMask model. By applying this “mask,” woven from algorithms, to videos, he could become the girl he had created from bits, and to the naked eye, his fake was indistinguishable from the real thing. 

If his internet speed allowed, he could also swap faces in real time to spice up the fun. Of course, more fun meant more work. For real-time deception to work, he had to simultaneously translate English or Igbo into Yoruba, use transVoice to imitate the voice of a Yoruba girl, and use a lip-sync open-source tool kit to generate corresponding lip movement. If the person on the other end of the chat had paid for a high-quality anti-fake detector, however, the app might automatically detect anomalies in the video, marking them with red translucent square warnings.

In the early days of deepfake technology, factors like internet speed and exaggerated expressions could easily cause glitches, resulting in images that blurred, or out-of-sync lip movement. Even if the glitch lasted for only 0.05 second, the human brain, after millions of years of evolution, could sense something was amiss. By 2041, however, DeepMask—the successor of deepfake—had achieved a degree of image verisimilitude and synchronization that could fool the human eye.

 Anti-fake detectors had become a part of the standard configuration for cybersecurity. Europe, America, and Asia had even made them compulsory by law, yet in Nigeria, only major content platforms and government websites required such verification. The reason was simple: The detectors required an extremely high level of computing power and skill, and they slowed the speed of videos. If people had to wait, they would tune in elsewhere. Social media and video-sharing platforms would update their detectors selectively according to the most popular fake-generating algorithms at any given time; the more a piece of content was shared, the more scrutiny it got.

After each video “date,” Amaka would sit quietly in the darkness. His humble surroundings never failed to inject him with a dose of reality. Still, he would allow his mind to linger over the smiles and sweet words of the boys he “dated.” Their affection doesn’t belong to me, he reminded himself, but to a Yoruba girl with a face just like mine. 

When Amaka was born, a local soothsayer had declared to Amaka’s father that his new son was actually the reincarnation of a female soul trapped in a male infant’s body. The “soul-body incongruity” would be a shadow cast over Amaka’s entire childhood—and a shame for his family.

As he grew up, Amaka slowly came to understand that he was not like other boys. Leaving his village to come to Lagos had been part of this journey. Still, there were limits. When he brushed past attractive men on the light-rail or the sidewalk, he could feel something in his body—in his soul—flutter. Even eye contact could evoke the feelings now and then. But Amaka knew that he didn’t have the courage to face the boys he chatted with online in real life. The more power of DeepMask he excavated, the more his addiction for the mask grew. It concealed his real face, so that he was able to let his feelings pour out and run free, without exposing himself to danger or shame. 

Amaka was forcing himself to focus on the fake video when his thoughts were interrupted by a knock on his bedroom door. Ozioma, his landlord, entered with a bowl of sliced kola nut seeds. Ozioma, an Igbo who’d moved to Lagos 20 years ago, had seamlessly assimilated into Yoruba society. However, she was able to uncover Amaka’s veiled Igbo accent immediately after meeting him. 

“You know, where I come from, only men are permitted to break open a kola nut,” said Amaka, his mouth full, savoring the fruit’s familiar bitterness.

“That’s exactly why I moved!” Ozioma chuckled. “The kola nut, the Yoruba call it obi, and the Igbo call it oji. Why does the name matter? Obi or oji, it will solve your problems once it’s in your mouth.”

“Ah, wisdom of the elders. Thank you for this treat,” said Amaka. However, before he could shut the door, Ozioma grabbed his arm. She pointed to the headshot displayed on his monitor, a worried frown crossing her face. “You don’t have anything to do with him, do you? I mean, he’s a good guy, I just … don’t want to get into trouble, if you know what I mean.”

“Nah, I was just reading the news.” Amaka forced a lighthearted smile. “I still want my residence permit.”

“Good child. May God bless him—no matter what side he is on.” Ozioma disappeared. 

Amaka sighed in relief and hopped back onto his bed, returning his attention to the face displayed on his monitor. 

The face radiated power. Its forehead and cheeks were painted in white, a symbol of tribal spirit. Its eyes glowed as if they were globes of fire. Its lips were slightly parted, with corners curling up to form a half smile, as if about to speak the divine language of a new age and take the world by storm. 

The face belonged to Fela Kuti—legendary Nigerian musician, father of Afrobeat, fighter for democracy—who had been dead for 45 years.

Amaka’s problem was how to make something fake even more fake. 

A virtual avatar—with the face of Fela Kuti—had emerged online, posting videos on GarriV. The figure, in the guise of the deceased Fela Kuti, had become an internet sensation. The avatar called itself “FAKA,” the abbreviation of “Fela Anikulapo Kuti Avatar,” and its videos mainly involved acerbic commentary on current social affairs—even if its precise political allegiance was hard to pin down. Most people treated it as a joke. Everyone knew that the real Fela Kuti had died in 1997. The face-swap technology used in the videos was so crude as to be laughable. Rather than bothering to ban or censor the FAKA videos as fake content, sharing platforms had simply tagged them as parody.

Still, FAKA’s influence had snowballed into no laughing matter. Millions of Nigerians were logging in to encrypted chat groups to discuss FAKA’s videos, analyzing every frame and syllable. They had even been translated into different dialects, fully dubbed and lip-synced, disseminating FAKA’s message far more widely. The official Fela Kuti Foundation issued a statement, claiming to be as puzzled as anyone about the popular avatar’s origins, but stopping short of issuing a demand for the mysterious figure behind the account to stop using Fela Kuti’s likeness. 

No one had managed to track down the person behind FAKA. The videos’ information was encrypted; the account that uploaded the videos was disposable and had gone through multiple proxy servers. Consequently, conspiracy theories emerged. Was FAKA the work of anti-government activists or a foreign government, intent on undermining the current order?

Ljele, Amaka’s new employer, was not an actual company, it turned out. Ljele was the front operation for an underground group called Igbo Glory, and Chi was just the representative—the agent tasked with recruiting and handling Amaka. The group had analyzed the content of FAKA’s videos and come to a different conclusion: Yoruba ultranationalists were behind the avatar, and they hoped to exploit its popularity to manipulate the minds of the people—to incrementally make FAKA’s videos more pro-Yoruba and nudge public opinion in their favor. And the more power that coalesced in the hands of the dominant Yoruba, Amaka knew, the more other ethnic groups would get squeezed—especially the Igbo.

In one recent video, FAKA had called on Igbo-dominated states to give up the claim to a newly discovered rare-earth element deposit, and instead make it “a common property for all Nigerians.” This was the latest attempt to deprive the Igbo of resources on their land. The Igbo felt like the tail to Nigeria’s lizard—cut off, grown back, then cut off again, in a never-ending cycle. No one cared if the tail hurt or bled.

Now, the Igbo were tired. Amaka’s mission was key to Igbo Glory’s goal of revolution. In the hopes of disrupting FAKA’s grip on public opinion, Chi had tasked Amaka with producing fake FAKA videos that would undermine the avatar’s credibility and influence.

Technology-wise, it wasn’t all that hard. With the help of H-GAN, Amaka easily duplicated a machine-generated model of FAKA’s facial portrait. From blink frequency and lip movement to the crude incongruity between the mouth area and its surrounding skin, Amaka’s model was a pixel-to-pixel mirror reflection of FAKA. As long as he knew how to set the parameters and match up each mathematical value between the fake and the original, he could fool every anti-fake detector and every human eye.

The real challenge was reproducing a FAKA-style speech. The topics of FAKA’s videos ranged from social and political news to convoluted “everyman” populist gripes. In the monologues, FAKA would selectively quote famous words from the real Fela Kuti, as well as folk sayings. Amaka often struggled to interpret FAKA’s distinctive speeches—let alone imitate them.

FAKA declared that Nigeria was in dire need of a new language that transcended ethnic boundaries, “to purge our mind and language of colonial poison.” It lamented that the mothers of Nigeria were the people who “suffered the most and deserve the highest reverence”; with their own hands they have “welcomed the descent and buried the corpses” of countless children. FAKA boasted that “music is a weapon of the future,” and that only when education and wealth were “distributed evenly like drumbeats permeating the air, could the heartbeats of people coalesce into one steady rhythm.”

Like a rainstorm descending upon a long-parched land, FAKA’s words had begun to quench a thirst in Amaka’s heart too. As much as he hated to admit it, he felt invigorated by a sense of hope. Was Chi right about FAKA? Amaka tried to brush off these feelings. I don’t need some cheesy sense of belonging, he told himself.

Amaka needed only a perfect counterfeit of a FAKA-style speech, one that people would believe.

A parade had flooded the streets of central Lagos. Hidden on the balcony off his room, Amaka watched as a troupe of young men naked from the waist up swayed and spun, as graceful and nimble as specks of dust dancing in rays of sun. Their faces were decorated by white paint in the style of Fela Kuti. The muscles on their backs glistened under the hot sun. Following the rhythm, they raised their arms in unison, shaking their palms, as if casting a spell.

The sound of instruments from various ethnic groups combined in harmony. The shrill cry of the Batá drum and the low moans of the dùndún drum, from the Yoruba; the metallic clatter of the ogene bell and the silvery melody of the opi flute, from the Igbo. The air vibrated from the music, like the tightening string of a bow drawn open inch by inch. The dancers, like young cassava shoots during monsoon, shaped their movement to the flow of the rhythm. Moving in complete unison, with no one left behind, the dancers, to Amaka’s eye, seemed less like individuals than a connected being—not unlike the mantra that they’d been chanting, “One Nigeria,” the slogan of FAKA’s video campaign.

Amaka felt torn. On the one hand, he envied the dancers. He instinctively wanted to join them, yet his passion was choked back by an intense fear of being exposed as a traitor. Did those dancers—followers of FAKA—really wish ill on the Igbo people, a people Amaka still loved, even as he’d grown alienated from them?

More pressing than those thoughts, however, was Chi’s deadline, which was quickly approaching, and with the passing of each day Amaka had become more and more certain that he had been given an impossible task.

Upon Amaka’s closer examination, it seemed a uniform, singular FAKA personality did not exist. The team behind the avatar, relying on the video-sharing platform’s smart tagging system, had created videos tailored to appeal to a variety of user profiles, fine-tuning the main topics, slogans, tone, and body movement for each audience—like an advertising agency pandering to certain demographics.

Creating a fake was one thing—creating a fake with multiple personalities was beyond Amaka’s capabilities. Somehow, this realization gave him a sense of relief. But now he had to face the consequences of failing Chi’s mission.

“Why don’t you go join them?” asked Ozioma. Showing up behind Amaka on the balcony, the landlady lit an English-brand cigarette, leaned against the railings, and peered down.

“I used to be the dance queen of our village,” Ozioma went on, her eyes hazy with nostalgia. “Not trying to brag here, but not a single boy could take his eyes off me. My father hated when I danced, though. He threatened to hit me every time he caught me dancing.”

“Did you listen to him?”

Ozioma laughed heartily. “Why on earth would a child give up what they love because their parents said no? Eventually, I found a way that could allow me to at least finish the dance.”

“What was it?” asked Amaka.

“I would wear an Agbogho Mmuo every time I danced.”

What?” Amaka’s eyes widened. The Agbogho Mmuo was the sacred mask of northern Igbo, representing maiden spirits as well as the mother of all living creation.

“See, my father had your exact expression when he saw me with the mask. He had no choice but to bow down, to show his respect to the mask and the goddess it embodies. Of course, after I was done with the dance, with the mask stripped off, I would get my share of scolding,” said Ozioma, beaming with pride, as if the memory had temporarily brought her back to the days when she was a young girl.

Upon hearing Ozioma’s story, Amaka felt an idea, blurry and shapeless, darting across his mind like a fish. He scrunched up his face, thinking. “The mask …”

“Yes, child. The mask is where my power came from.”

“Strip off the mask? Strip off the mask,” murmured Amaka.

All of a sudden, he leapt to his feet and kissed Ozioma on the cheek. “Thank you, oh thank you, my dance queen!” He dashed back to his room, leaving behind the hustle and bustle of the parade and a very confused Ozioma.

“Maybe spinning a lie and putting it in FAKA’s mouth won’t make his followers abandon their idol,” Amaka told Chi via video chat that afternoon, excited with his new discovery. “But stripping off its mask and revealing the hidden puppet master might.”

“No one knows who the puppet master is, though,” Chi replied.

“Exactly!” Amaka beamed. “Can’t you see? It means that the puppet master can be anyone.”

“So, you’re suggesting that …”

“I can strip off FAKA’s mask and make him any person you want him to be.”

Chi fell silent in the video chat.

“You’re a fucking genius,” Chi finally muttered.

Ndewo,” Amaka said, preparing to sign off.

“Wait,” Chi looked up. “It means that you need to create a face that exists in reality.”

“Yes.”

“A face that can fool all the anti-fake detectors,” added Chi, musing. “Think about the color distortion, the noise pattern, the compression rate variation, the blink frequency, the biosignal … is it doable?”

“I need time,” said Amaka. “And unlimited cloud AI computing power.”

“I’ll get back to you.” Chi logged off.

Amaka gazed at his own reflection in the dimming monitor screen. The adrenaline rush that had initially washed over him had faded. He saw on his face not excitement, but exhaustion and an unsettled feeling, as if he had betrayed a guardian spirit watching from above.

In theory anyone could fake a perfect image or video, at least well enough to fool the existing anti-fake detectors. The problem was the cost—computing power.

Fakes and their detectors were engaged in an eternal battle, like Eros and Thanatos. Amaka had his work cut out for him, but he was determined to succeed in achieving his singular goal: the creation of a real, human face.

In the new scheme Chi concocted, FAKA would be stripped of its Fela Kuti digital mask to reveal the face of Repo, a notorious Yoruba politician known for his ad hominem attacks on other ethnic groups. Repo was the primary enemy to the “One Nigeria” movement. Once Chi and his team revealed to the public that Repo was pulling the strings behind the inspiring, charismatic FAKA, the faith of the avatar’s believers would shatter into pieces. First, though, Amaka’s fabricated video would have to endure the scrutiny of millions of eyeballs—human and AI, including “the VIP detector.”

The VIP detector, as it was nicknamed, was designed to protect the reputations of public figures: politicians, government officials, celebrities, athletes, and scholars. Such prominent people had large internet trails—which made them particularly ripe to be targets of deepfakes. The VIP detector was intended to prevent those “supernodes” in cyberspace from becoming the victims of fraud, and the consequential devastating damage to social order that could ensue. Websites posting pictures or videos of prominent individuals were required to apply this special detection algorithm to content before posting. The VIP detector incorporated tech ranging from ultra-high-resolution facial recognition, body-language recognition sensors, hand/finger geometry recognition, speech evaluation, and even vein recognition.

All of this data fed into the VIP detector’s deep-learning AI. The VIP detector would even incorporate medical history into its data bank, as long as the person being protected was important enough. No doubt, given Repo’s social status and controversial position, he was one of those VIPs.

Amaka, however, believed there was a flaw in the detector. If he could decipher how the network of anti-fake detectors was made, he could pinpoint the gaps in the crisscrossing strands of data inputs and exploit them. No matter how narrow the holes in a net, a determined fish can eventually find a way out.

Using a real video of Repo as a base, Amaka, like a 21st-century Dr. Frankenstein, carefully sewed together the face: lips, eyes, and nose, layer by layer, aided by AI. Every twitch and gesture in the fake video would come from Repo himself, thus greatly reducing the chance of being caught by the anti-fake detector.

Using XR vision, Amaka had conjured up a three-dimensional workspace. He waved his hands in the air, selecting, dragging, zooming in and out, the icons and footage fragments hovering midair with alternative gestures. He would have preferred to see himself as a wizard working magic, yet in reality he looked more like a star chef preparing an extravagant feast.

For each part of Repo’s body, Amaka carefully selected the most effective open-source software, like placing raw ingredients into a proper cooking vessel. Then, as if seasoning the food, he adjusted parameters, models, and the training algorithm. Finally, he brought them to a simmer in a cloud AI platform with maximum computational power. Each set of video resources, processed by GAN, generated a series of thumbnails that extended into infinity in the virtual workspace, like a never-ending gallery swamped with posters of Repo’s various body parts.

Behind the poster wall a ferocious battle was happening in the cloud, in utter silence. The two sides were GAN’s positive and negative poles, the detective network and the forger network. The goal of the forger network was to retrain and upgrade itself to generate more realistic images that could fool the anti-fake detectors, based on feedback from the detective network, in order to minimize the loss function value of the generated image. Conversely, the detective network strived to maximize the loss function value. This contest, with the stakes rising every millisecond, would repeat itself millions of times until both sides reached a certain balance.

Adjusting parameters, iterating the model … with each adjustment Amaka could see the video becoming more realistic. His eyes, almost blinded by colorful pixel dots, focused intently on frames in his XR vision field—frames that differed from one another only by the thinnest margin. Sweat gathered on his forehead, ran down his face, and dripped from the tip of his nose, but Amaka’s nimble dancing fingers weren’t affected at all.

However, a voice rang in his ear every now and then, distracting him, like an ogbanje forever stuck in the limbo of life and death.

“You’re murdering a god with your own hands,” the voice whispered.

He’s not my god. He’s a Yoruba, Amaka repeated in his heart, while forcing himself to turn his attention back to work.

Finally, he let out a sigh of relief. His fake video had successfully fooled the VIP detector. Exhausted, he collapsed onto his bed and fell into a deep sleep.


The story continues, followed by a (nonfictional) exposition on deepfakes by Kai-Fu Lee. Join WIRED on Tuesday, September 14, from 5 to 6 pm PT (8-9 pm ET; September 15, 8-9 am Beijing) for a Twitter Spaces conversation between Kai-Fu Lee, coauthor and science fiction writer Chen Qiufan, and WIRED AI reporter Tom Simonite. Do you have questions about how the two coauthors collaborated on this project? Or about any aspect of the future of AI? Drop them in the comments below, or join us and ask them in real time on September 14.


From the book AI 2041: Ten Visions for Our Future by Kai-Fu Lee and Chen Qiufan, forthcoming on September 14, 2021. Copyright © 2021 by Kai-Fu Lee and Chen Qiufan. Published by Currency, an imprint of Random House, a division of Penguin Random House LLC. All rights reserved.


If you buy something using links in our stories, we may earn a commission. This helps support our journalism. Learn more.


More Great WIRED Stories

Products You May Like

Articles You May Like

The Science of Crypto Forensics Survives a Court Battle—for Now
The White House Puts New Guardrails on Government Use of AI
Here Comes the Flood of Plug-In Hybrids
Meta Kills a Crucial Transparency Tool At the Worst Possible Time
Large Language Models’ Emergent Abilities Are a Mirage

Leave a Reply