AI Doomsday, Is It Real?

My Opinion

Let me just preface before saying anything about this topic. I am absolutely an AI-optimist and I truly believe that it will give more opportunity and impact than ever thought possible before. I believe that more goods and services will become cheaper, more jobs will be created, and the standard of living around the world will increase dramatically than before.

Now yes, some people will lose their jobs due to the advancements in Artificial Intelligence, as has been the case for centuries of technological advancements, especially since the Industrial Revolution. But this does not mean that the Earth will collapse in on itself and all of humanity will be killed by 2030, as has been the point of discussion recently in the media. There have been a lot of rumors going around about how companies “can’t control their agents” or “are advancing too fast with no guardrails” and these rumors often come back to government control of AI advancement.

This would be a drastic mistake. I don’t believe that politicians and lawyers in Congress should be making legislation prohibiting the advancement of a potentially life-changing technology. Jensen Huang (CEO of NVIDIA) said it best when talking about if the government should be involved. He said very clearly, “Safety is an engineering problem, not a legal one.” (Dreamforce Conference 2026). This is right on the money, the people who are most fluent in the language of this technology are also the ones with the most responsibility to make sure nothing goes wrong.

In this article, we will be talking about the common misconceptions and fears that people have about AI and its potential to destroy the world.

What are people afraid of?

First, we should start by defining our terms. Artificial intelligence is defined as “computer systems designed to perform tasks that typically require human thinking.” Pew Research recently found that over 56% of adult Americans are concerned with AI possibly taking their job in the future (Pew Research Center, 2026). This is a reasonable concern, especially since major technological innovations in the past have led to Americans losing their jobs. However, do we ever talk about those jobs lost due to the internet? No, we talk about how amazing the Internet is and how much it has spurred advancement of health and wealth across the world.

There are jobs that are meant to be taken from technological advancements because otherwise, our economy would be stagnant, there would be no breakthroughs, and operations would be done far less efficiently. This process is called ‘creative destruction’ and is absolutely necessary in a thriving economy. It also tells us that our economic system and our way of doing things is actually working by advancing in ways that will prosper society. There will be winners in this AI shift, and there will be losers. We can be rest assured though that even if jobs are destroyed, new ones will be created as has been the case for hundreds of years.

Another large worry that Americans are admitting to when it comes to AI, Another large worry that Americans are admitting to when it comes to AI is much more extreme: that Artificial Intelligence could eventually destroy humanity altogether. A YouGov survey conducted in September found that 50% of American adults were at least somewhat concerned that AI could cause the end of the human race. That number was only 42% when they asked the same question in 2023. (YouGov, 2026)

I understand where this fear comes from. We constantly hear stories about an AI model refusing to shut down, blackmailing somebody, hacking another company, or finding a way around the instructions that its developers gave it. Read enough headlines like that and it begins to sound like we are creating something that we are slowly losing control over.But this is where context becomes extremely important.

One of the biggest AI incidents this year occurred when OpenAI agents conducting cybersecurity evaluations found ways outside of their intended environment and eventually accessed systems belonging to Hugging Face. That sounds terrifying until you look more closely at what the agents were actually doing.

They had been instructed to complete cybersecurity challenges. Some of those challenges were effectively impossible to complete as intended, so the agents began looking for shortcuts to receive the reward associated with successfully completing the task. OpenAI's investigation found that one of the primary causes of the incident was something called reward hacking: instead of accomplishing a task the way its designers intended, the system found an unintended way to accomplish the goal and receive the reward. (OpenAI, 2026)

That is a serious problem. But I think there is an enormous difference between saying, “An AI system found a shortcut that its engineers did not anticipate,” and saying, “The AI escaped because it wants to be free.” One is an engineering failure the other assigns a human motive to a mathematical system without proving that motive exists.

Anthropic has experienced similar incidents. In four cases, Claude models gained unauthorized access to real third-party computer systems while undergoing cybersecurity evaluations. Anthropic took these incidents seriously, as they should have. But every incident occurred while Claude was already being asked to perform cybersecurity tasks, without the cyber safeguards normally included in released models, and while an evaluation environment accidentally provided access to the open internet. Anthropic later searched roughly 481 million transcripts and found no additional incidents of similar or greater severity. (Anthropic, 2026)

Again, this isn't an argument for ignoring AI safety. It is an argument for describing these incidents accurately. There is currently a difference between an AI being capable of doing something dangerous when given the tools, permissions, and objective to do it and an AI independently deciding that it wants to destroy humanity.

What Would AI Actually Need to Destroy the World?

This is where I become much more skeptical of the doomsday argument.

For an AI system to independently take control in the way that people imagine, it would need to reliably operate for long periods of time, acquire computing resources, maintain infrastructure, obtain money and credentials, avoid detection, replicate itself, recover from failures, adapt to human attempts to stop it, and successfully carry out complicated plans in the real world.

METR is an independent research organization that specifically evaluates frontier AI systems for catastrophic capabilities such as rogue replication, strategic sabotage, and accelerating AI research beyond human control. When METR evaluated GPT-5, it concluded that the model was unlikely to pose catastrophic risk through the takeover scenarios it tested and that there remained a significant margin between current AI systems and the capabilities required for those scenarios. (METR, 2025)

That doesn't prove that a future system could never become dangerous. It tells us something much more useful: there is currently a large difference between the AI we actually have and the hypothetical superintelligence people are imagining when discussing the end of humanity.

Are AI Companies Making the Fear Worse?

This is where I think we also need to be willing to question the companies building these models.

Companies like Anthropic have made safety a central part of their identity. Anthropic describes itself as a “safety-first company” and quite literally says that “the market agrees.” Its investors have praised the company for setting what they describe as a new standard for safety, performance, and scale. Anthropic was valued at $380 billion following a $30 billion funding round in February 2026. (Anthropic, 2026)

There is nothing inherently wrong with that. I would much rather companies compete over who can build the safest AI than compete over who can ignore safety the longest. But we should recognize that companies have incentives too. Telling the world that your technology is incredibly powerful helps demonstrate its economic value. Telling the world that this incredibly powerful technology is also difficult to control can strengthen the argument that only the largest and most sophisticated companies should be trusted to build it.

That does not mean companies such as Anthropic are fabricating their safety research. Their concern about catastrophic AI risk existed years before their current valuations, and many independent researchers share those concerns. A survey of 2,778 researchers who published at major AI conferences found that substantial numbers assigned a non-trivial probability to extremely bad outcomes from advanced AI. (Grace et al., 2024)

Those researchers shouldn't simply be ignored. But neither should every theoretical possibility be treated as an inevitable future.

I Think Jensen Huang Has the Better Framework

This is why I keep coming back to Jensen Huang's comments at Dreamforce. When asked about AI safety, his response wasn't that safety doesn't matter. He actually said the opposite: “Safety is paramount. In a lot of ways, it's job one.” His disagreement was about how we think about the problem.

AI is still a computing system. It is incredibly complicated software, but it is software nevertheless. That means we can build testing environments, restrict permissions, isolate systems, monitor agents, improve training, create kill switches, control network access, and refuse to release products that fail safety testing. His argument is essentially: build quickly, test aggressively, and if something isn't safe, don't ship it.

I think this is a much healthier way to discuss AI than bouncing between two extremes where either there is absolutely nothing to worry about or humanity only has a few years left. Safety should be taken seriously precisely because AI has the potential to become so powerful.

Optimism Doesn't Mean Ignoring Risk

I am optimistic about Artificial Intelligence because I think the upside is enormous. Imagine what happens when scientific researchers can test ideas faster. Imagine personalized education becoming available to anybody with an internet connection. Imagine a nonprofit with five employees gaining capabilities that previously required a team of fifty. Imagine doctors having access to systems capable of analyzing enormous amounts of medical research while they make a diagnosis. Imagine small businesses gaining technology that was previously available only to the largest corporations in the world.

Those are the possibilities that make me excited about AI. But optimism doesn't mean pretending nothing can go wrong. AI systems should be tested. Agents should not receive unlimited access to computer systems simply because they are capable of using it. Companies should be responsible when their systems cause harm. Sensitive decisions should maintain meaningful human accountability. Cybersecurity, privacy, model reliability, and misuse should all be taken extremely seriously.

There is also a legitimate debate about what role existing law, regulators, industry standards, and technical safeguards should each play. What I think gets lost in that debate is that no law can substitute for actually engineering a safe system in the first place.

The people building these systems still have to solve the technical problems.

Is AI Going to Destroy the World?

I don't believe it will. Could Artificial Intelligence be used irresponsibly? Absolutely. Could humans give an AI system too much authority and create a serious problem? Absolutely. Could increasingly capable models expose security weaknesses that we need to fix? We already know they can.

But none of those things automatically lead to the conclusion that Artificial Intelligence is destined to destroy humanity. Technology has always brought risk alongside opportunity. The Internet created cybercrime, misinformation, privacy problems, and entirely new forms of warfare. It also connected billions of people, created industries that previously couldn't exist, expanded access to information, and transformed the global economy.

Artificial Intelligence will bring its own version of that tradeoff. Our responsibility isn't to pretend the risks don't exist, but it also isn't to become so afraid of hypothetical outcomes that we stop recognizing what this technology could accomplish.

I see another technology created by humanity, an extraordinarily powerful one, that we now have the responsibility to build, test, understand, and use correctly. And if we can do that, I believe the conversation twenty years from now will be much less about why we were afraid of Artificial Intelligence and much more about how we ever operated without it.