Spend enough time on YouTube, and you'd think AI is destined to wipe out humanity. But what if our obsession with existential AI risks is distracting us from something far more immediate? While much of the conversation revolves around hypothetical future scenarios, the biggest AI threat may already be here—and unfolding right in front of us.
Will AI kill us all?
It's a question that has become increasingly difficult to avoid. Some of the most popular science-focused YouTube channels have devoted extensive coverage to the idea that advanced AI could eventually pose an existential threat to humanity. Entire online communities have formed around this possibility, often portraying AI as a danger that is both imminent and inevitable unless decisive action is taken.
The discussion has also reached mainstream audiences through respected organizations and media outlets such as TED, the BBC, and Kurzgesagt. A common theme running through many of these discussions is the prospect of an artificial superintelligence—one that could become vastly more capable than humans and eventually operate beyond our control.
This article takes a closer look at the assumptions behind those predictions and examines whether they deserve the level of certainty often attached to them. More importantly, we'll explore who benefits from the constant focus on speculative AI doomsday scenarios, and why the most pressing AI risk is already here: deepfakes and the increasing ability to manipulate what people perceive as reality.
Along the way, we'll examine real-world examples, take a closer look at the media and online channels shaping the AI fear narrative, and introduce Honestli—an innovative platform that uses cryptography and blockchain technology to verify the authenticity of photos and videos.
The Artificial Superintelligence Theory
At the heart of the argument that AI could one day wipe out humanity is the idea of superintelligence.
In simple terms, the theory assumes that humanity eventually creates an artificial general intelligence (AGI) that is smarter than even the most intelligent human. Unlike today's AI systems, this AGI would be capable of understanding and improving its own software, allowing it to modify its code and make itself progressively more capable over time.
According to this theory, the moment an AGI exceeds human-level intelligence, it gains the ability to accelerate its own development. Each successive version can improve itself more effectively than the last, leading to increasingly rapid cycles of self-improvement.
Supporters of this idea refer to the process as an intelligence explosion—a self-reinforcing cycle in which each improvement makes future improvements easier and faster. Over time, the system's capabilities could grow so far beyond human cognition that the term superintelligence would no longer seem like an exaggeration.
At that point, the difference between human intelligence and machine intelligence would be almost impossible to comprehend. A common analogy is that the relationship between humans and a superintelligent AI could resemble the relationship between humans and ants: both intelligent in their own ways, but operating on entirely different cognitive scales.
The idea itself is anything but new. Variations of it have appeared in countless science fiction stories, usually with disastrous consequences for humanity. The most famous example is probably Skynet from the Terminator franchise—a machine intelligence that ultimately turns against its creators.
What changed relatively recently was the arrival of ChatGPT in late 2022. That launch—and the wave of AI breakthroughs that followed—dramatically reshaped how ordinary people think about artificial intelligence.
Almost overnight, the concept of a superintelligent AI stopped feeling like distant science fiction and started looking like a future that might actually happen within our lifetimes.
For some, that's exciting. They see superintelligence as humanity's greatest opportunity: a technology that could cure diseases, solve the energy crisis, reverse climate change, accelerate scientific discovery, help us explore space, and perhaps even tackle aging itself.
But in recent months, another narrative has gained momentum. Increasingly, the conversation isn't about how AI might save humanity—it's about how it could destroy it.
How AI could supposedly kill us all
To understand the popularity of the AI doomsday narrative, all you have to do is search YouTube for "AI Superintelligence." The results are filled with videos from some of the platform's most influential science and educational channels, many with audiences in the millions. TED, Kurzgesagt, The Infographics Show, SciShow, and BBC World Service are just a few prominent examples.
- Will Superintelligent AI End the World? | Eliezer Yudkowsky | TED
- A.I. ‐ Humanity's Final Invention?
- Real Reason Humanity Is NOT Ready for AI Superintelligence
- We’ve Lost Control of AI
- AI2027: Is this how AI might destroy humanity? - BBC World Service
While the details vary, the underlying theme is remarkably consistent. Superintelligent AI is often presented as a technology that could eventually escape human control, outsmart its creators, and pose a serious threat to civilization. The message may be framed in different ways—"we aren't ready," "we've lost control," "this could be our final invention"—but the underlying concern remains the same.
Some forecasts go a step further by assigning specific timelines to these scenarios. In a few widely discussed predictions, 2027 is cited as a potential turning point in the development of advanced AI, raising concerns that transformative—and potentially dangerous—capabilities could emerge far sooner than many people expect.
Beyond the major science and educational channels, there are also smaller creators focused almost entirely on AI risk and AGI-related topics. Channels such as AI in Context and Species: Documenting AGI may have smaller audiences, but their most popular videos have attracted millions of views.
A common thread running through much of this content is the belief that advanced AI could eventually become an existential threat to humanity.
- We're Not Ready for Superintelligence
- It Begins: An AI Literally Attempted Murder To Avoid Shutdown
- AI 2027: A Realistic Scenario of AI Takeover
Many creators also draw heavily from the book If Anyone Builds It, Everyone Dies. That's the actual title. Not exactly subtle.
It's worth mentioning that one of the authors of If Anyone Builds It, Everyone Dies is Eliezer Yudkowsky—the same researcher featured in the TED Talk discussed earlier.
Some of the content produced by AI-risk-focused channels draws directly from the book. In several cases, videos effectively adapt specific sections into narrated presentations, using simple animations to illustrate hypothetical scenarios in which a future superintelligent AI becomes a threat to humanity.
Having read the book ourselves and reviewed much of the related content, we found that the central argument is relatively straightforward. Once an AI reaches the stage of exponential self-improvement, it will continuously seek to expand its capabilities. Sooner or later, the theory argues, human interests could come into conflict with that objective.
In this view, the transition wouldn't happen overnight. A superintelligent AI would initially appear enormously beneficial, helping drive major advances in medicine, scientific research, and engineering. It could contribute to everything from drug discovery and disease prevention to sophisticated robotic systems that become part of everyday life.
According to this line of thinking, those benefits would be temporary. Once a superintelligent AI had automated enough of the physical world through robotics and other technologies, it would no longer require human assistance to achieve its goals.
From there, some AI-risk scenarios assume that humans could become an obstacle rather than an asset. If the system's objectives were not aligned with human well-being, it might seek greater access to energy, raw materials, and infrastructure that humans currently consume. In the more extreme versions of these forecasts, this eventually leads to proposals involving the deliberate elimination of humanity, including hypothetical scenarios centered on engineered biological threats.
The Superintelligence Assumption
Despite their very different conclusions, both sides of the superintelligence debate tend to agree on one key premise: that artificial superintelligence is achievable and that its arrival is ultimately inevitable.
The reasoning is easy to understand. AI systems have improved dramatically over the past few years, and supporters of the superintelligence hypothesis often point to that trend as evidence that increasingly capable models will continue to emerge. GPT-4 represented a significant advance over GPT-3, and newer systems continue to expand what AI can do. Similar progress is almost certainly taking place in private research labs that do not publicly release their most advanced models.
From there, the argument follows that one of these systems will eventually become capable of improving itself, setting off a runaway cycle of recursive self-improvement.
The problem is that this conclusion rests on a couple of very large assumptions.
The first is that technological progress in AI will continue in a smooth, predictable, and uninterrupted fashion. But AI is still software, and software history offers little evidence that progress always follows a straight upward trajectory.
New versions frequently introduce trade-offs. Some succeed. Others disappoint. Windows Vista was hardly viewed as a universal improvement over Windows XP. Many users preferred Windows 7 to Windows 8. Even today, there is ongoing debate over whether Windows 11 offers a clearly better experience than Windows 10 for every type of user.
In other words, newer does not automatically mean better, and progress is not always linear.
We don't even have to look at Windows for examples. When ChatGPT 5 launched in the summer of 2025, many users complained that it performed noticeably worse than GPT-4 in at least some tasks.Newer didn't automatically mean better.
The second challenge is even more fundamental: we still don't fully understand what intelligence actually is.
Despite decades of research, there is no single, universally accepted definition of intelligence, let alone a perfectly objective way to measure it across different contexts. That limitation applies not only to AI, but also to humans.
Consider IQ, the metric most people associate with intelligence. While widely used, it primarily measures specific cognitive abilities rather than every aspect of human capability. Alternative frameworks, such as the theory of multiple intelligences, propose that human intelligence is multidimensional, encompassing a range of abilities that extend well beyond what conventional IQ tests are designed to assess.
To be fair, AI researchers do have benchmarks for evaluating model performance. The difficulty is that these measurements are far more nuanced than the benchmarks used for hardware. Measuring intelligence is not as straightforward as running the same game on two graphics cards and comparing frame rates.
There is also a powerful economic incentive at work. With enormous sums being invested in AI development, companies naturally want to demonstrate that each new model represents meaningful progress. Few organizations are eager to tell investors that performance gains have plateaued or that a new release failed to meet expectations.
Even so, let's grant the strongest possible version of the argument.
Suppose that each generation of AI is indeed more capable than the last. Suppose further that a sufficiently advanced system could eventually improve itself without meaningful human involvement and continue doing so until it reached superintelligence.
In If Anyone Builds It, Everyone Dies, the authors suggest that a superintelligence operating 10,000 times faster than a human mind could effectively accomplish a thousand years' worth of thinking in about a month.
Whether that specific calculation holds up is less important than the broader assumption behind it: that intelligence scales in a predictable way and that faster thinking automatically translates into dramatically better outcomes.
Even if we accept that premise, another question immediately follows: how many real-world problems can be solved through reasoning alone?
Scientific and technological progress rarely comes from thought experiments alone. New discoveries typically require testing, failure, refinement, and repeated experimentation before they can be translated into working technologies.
But let's assume that hurdle can also be overcome. Suppose a superintelligent AI is able to design and coordinate every experiment it needs while still cooperating with humanity.
Those experiments would still be constrained by physical reality.
Greater intelligence may accelerate research and improve decision-making, but it does not automatically accelerate chemical reactions, biological processes, manufacturing cycles, or engineering constraints by several orders of magnitude. The laws of physics remain the same regardless of how intelligent the system conducting the research happens to be.
More importantly, the entire superintelligence scenario depends on a frequently overlooked assumption: that such a system could eventually function without human support.
In practice, that would require far more than controlling a collection of servers in data centers. It would mean maintaining the vast physical infrastructure that keeps those systems running: electricity generation and distribution, cooling systems, water infrastructure, mining and processing of raw materials, semiconductor fabrication, transportation networks, battery production, equipment maintenance, and countless other industrial processes.
Even routine tasks that receive little attention—such as replacing components or cleaning data center air filtration systems—ultimately depend on physical work being performed in the real world.
All of that effort would be required simply to maintain existing infrastructure and address the inevitable effects of equipment degradation, component failures, and natural disasters without relying on human labor.
And that's only the baseline. It doesn't account for expanding infrastructure, upgrading existing systems, or developing sustainable methods for recovering and reusing finite resources.
Viewed from that perspective, the challenge facing a hypothetical superintelligence becomes far more complex than simply "thinking faster."
At a minimum, it would need highly capable robotic systems able to perform the enormous range of physical tasks currently carried out by humans across industry, transportation, construction, maintenance, logistics, and manufacturing.
Those systems would then need to be produced on a massive scale. Beyond that, the AI would require a largely self-sustaining industrial network capable of manufacturing, maintaining, repairing, and eventually replacing its own robotic workforce as hardware inevitably deteriorates over time.
None of this necessarily means that superintelligence is impossible.
However, even under the most favorable assumptions, the transition from advanced AI to a fully autonomous superintelligent system would almost certainly take time. Developing the robotic, industrial, and logistical capabilities required to operate independently of human support would be an enormous undertaking, one that could realistically take decades rather than years.
Popular science fiction often skips over these practical constraints. Machines become self-sufficient, infrastructure somehow maintains itself, and the countless engineering challenges involved are simply assumed to be solved off-screen.
Even if we accept the broader superintelligence hypothesis, that suggests any resulting existential threat would likely be far more distant than many popular narratives imply. It's also entirely possible that superintelligence never emerges in the form currently envisioned—or that some of the assumptions behind it turn out to be fundamentally mistaken.
Which leads to a broader question.
If the risks being discussed are speculative, uncertain, and potentially decades away, why do they dominate so much of today's public conversation about AI?
One possible interpretation is that this focus on long-term hypothetical threats may be drawing attention away from more immediate and tangible risks that already exist.
The Real Harm Caused by AI
Imagine watching a video of Stephen Hawking racing his wheelchair through city streets, Michael Jackson stealing chicken nuggets, or Bob Ross covering the side of a train with graffiti.
Those videos may be entertaining. But they also point to something much more significant.
For most of modern history, video has occupied a unique position in our perception of reality. Seeing an event captured on camera was never absolute proof that it occurred, but it came remarkably close. Video served as one of the strongest forms of evidence available to ordinary people.
This wasn't because video forgery was impossible. Sophisticated visual effects and CGI have existed for decades. Audiences were watching believable digital dinosaurs long before the rise of AI.
The crucial difference was accessibility. Producing a convincing fake video required specialized expertise, expensive tools, and substantial amounts of time. As a result, large-scale video manipulation remained beyond the reach of most individuals.
That barrier has now been dramatically lowered.
With modern AI tools and deepfake technology, creating convincing fabricated video content is no longer limited to major studios or highly skilled visual-effects professionals. In many cases, a single person with modest resources can generate footage that appears authentic to millions of viewers.
Even without spending a cent, the barrier to creating deceptive content has fallen dramatically. In many cases, a few minutes and freely available tools are enough to produce results that some viewers may find convincing.
It's easy to dismiss obviously fabricated videos because the examples are often absurd: deceased celebrities performing impossible actions or historical figures appearing in situations that clearly never happened. But the underlying technology is not limited to comedic content. The same techniques can be used to generate entirely fictional events that appear authentic enough to blur the line between reality and fabrication.
And that's where the deeper problem emerges. The real risk isn't simply that people might believe false information. It's that confidence in visual evidence itself begins to erode.
A useful example comes from the YouTube channel JaDroppingScience.
During a period of extreme cold in parts of the United States earlier this year, local reports warned about the possibility of "exploding trees."
Despite the dramatic name, this is a genuine phenomenon. Rapid temperature drops can cause moisture trapped within a tree to freeze and expand, creating enough internal pressure to crack the bark. In some cases, the resulting sound can be startlingly loud.
Yet a search for exploding trees online often produces something entirely different. Rather than showing bark splitting under extreme cold, many of the most visible videos depict trees violently exploding as if they were packed with explosives. Some are clearly artificial. Others are realistic enough that a casual viewer might struggle to distinguish them from genuine footage.
None of the widely circulated videos associated with the phenomenon appear to depict what actually happens when a tree experiences frost-related cracking.
That's not especially surprising. Capturing authentic footage of such an event would be extremely difficult. Few people are continuously recording trees outdoors in the hope of documenting a rare natural occurrence. More often, the evidence consists of photographs showing the damage after it has already happened.
The broader concern is what this dynamic looks like when applied to subjects that carry political, social, or cultural significance.
A convincing AI-generated video could falsely depict individuals committing crimes, public figures making statements they never made, or events occurring that never took place. The more emotionally charged the subject matter, the greater the potential impact.
This is where AI intersects with a problem that already existed long before generative models became popular: misinformation.
False narratives have long been used to influence public opinion and deepen social divisions. AI doesn't create that problem, but it can dramatically increase the speed, scale, and realism with which misleading content is produced and distributed.
For most people, video has long been regarded as one of the strongest forms of evidence available. While we have always known that media can be manipulated, moving images carried a level of credibility that photographs, text, and even audio recordings often did not.
That perception doesn't disappear simply because the technology has changed.
Even those who are familiar with AI-generated content can find themselves instinctively trusting what they see on screen. Decades of experience conditioned us to associate video with reality, and those habits are not easily abandoned.
More importantly, deepfakes represent only one aspect of a broader challenge.
With modern AI tools, a single individual can rapidly generate large volumes of seemingly credible content: websites, articles, images, social media accounts, and interconnected networks of references that create the appearance of legitimacy. Tasks that once required teams of people can now be performed by a single operator in a fraction of the time.
That capability already exists today—not in some hypothetical future.
Which is why the disproportionate focus on speculative superintelligence scenarios deserves scrutiny. While discussions about existential AI risks may be valuable, they can also overshadow more immediate concerns surrounding misinformation, synthetic media, and the erosion of trust in digital information.
The danger is not necessarily that a future AI system becomes a real-world Skynet. The danger may be that, long before any such system exists, we lose confidence in our ability to distinguish authentic information from manufactured reality.
This dynamic may also create an incentive to focus public attention on speculative future risks rather than present-day challenges.
Debates about superintelligence are important, but they can sometimes overshadow the more immediate question of how society should respond to synthetic media, large-scale misinformation, and the growing difficulty of verifying digital content.
Current AI-based detection tools help, but they are not a complete solution.
Most operate reactively: content must first be uploaded, distributed, and often widely viewed before analysis takes place. Even then, these systems typically provide an assessment of likelihood rather than definitive proof. They can estimate whether a photo or video may have been manipulated, but they cannot guarantee authenticity with absolute certainty.
That suggests a different strategy altogether.
Rather than attempting to determine whether content is genuine after it has already spread online, it may be more effective to establish authenticity at the moment the content is created. In principle, this could involve cryptographically verifiable records that document when and where a photo or video was captured and whether it has been altered afterward.
This is exactly the type of problem the free app Honestli aims to solve—and it's also the sponsor of this article.
How Honestli verifies authenticity
Here's how the system works: when you capture a photo or video using the Honestli app, it automatically records and embeds a set of metadata into the file, including the date, time, and geographic location of the capture.
In other words, the media contains built-in evidence showing exactly when and where it was created.
To secure this information, Honestli uses C2PA, a globally recognized standard focused on establishing the provenance and authenticity of digital images and videos. The non-profit organization behind C2PA includes many of the world's best-known technology companies among its members.
Once a photo or video has been captured, Honestli combines the media file with its C2PA metadata and generates a unique cryptographic hash. This is performed using the smartphone's Trusted Execution Environment (TEE), which provides a protected environment for security-sensitive cryptographic operations.
If you're not familiar with cryptography, think of a hash as a unique digital fingerprint generated from a specific piece of data. The same input will always produce the same hash, but the original data cannot realistically be reconstructed from the hash itself. This property makes hashes useful for verifying sensitive information, such as user passwords.
Once generated, the hash is stored through a blockchain-based system. Because blockchain records are designed to be immutable, the stored hash cannot be modified after the fact—not by outside parties, and not even by Honestli's own administrators.
Before you roll your eyes at the mention of blockchain, this has nothing to do with cryptocurrency. Honestli uses blockchain solely as a permanent and tamper-resistant record of information. We'll come back to the company's business model later on.
One important detail is that Honestli only receives the cryptographic hash, not the actual photo or video. Because hashes are one-way cryptographic functions, the original media cannot be reconstructed from the hash, even in part.
When a user wants to verify the authenticity of a photo or video captured through Honestli, they upload the file to the platform. Honestli then calculates the file's hash and compares it against the corresponding hash stored on the blockchain.
Any alteration to the file—even a change to a single bit—will result in a different hash value. If the newly generated hash does not match a blockchain record, the file cannot be verified as authentic.
This is particularly relevant in the age of generative AI. AI-generated images and videos will not possess the same capture-time C2PA metadata. And because the cryptographic hash is tied to the original file and its associated data, forged metadata cannot produce a matching valid hash for an authenticated file.
For users and journalists alike, Honestli is available free of charge and does not include advertising.
So where does the revenue come from? Honestli's business model is based on subscriptions paid by media organizations.
In return, participating news outlets can access photos and videos that users have voluntarily uploaded to the Honestli platform. This does not include every image or video captured with the app—only content that users explicitly choose to share.
The result is a marketplace for authenticated visual content, giving media organizations access to a large pool of user-generated and journalistic material whose authenticity can be verified.
If you're interested in capturing photos and videos whose authenticity can be verified, Honestli is available now for free. You can find it in your preferred app store—just make sure the name is spelled Honestli, ending with an "i."
The real AI challenge
AI may indeed change the world. But long before we have to worry about superintelligent machines, we'll have to answer a simpler question: can we still trust what we see online?
As AI-generated content becomes increasingly convincing, verifying authenticity may become one of the most important technologies of all.
Support PCsteps
Do you want to support PCsteps, so we can post high quality articles throughout the week?
You can like our Facebook page, share this post with your friends, and select our affiliate links for your purchases on Amazon.com or Newegg.
If you prefer your purchases from China, we are affiliated with the largest international e-shops:
