In 1862, The Lancet published a lengthy analysis titled “The Influence of Railway Travelling on Public Health.” The invention of the railroad was also the invention of railway accidents, which media of that day covered in lurid detail, but that was the least of it. Experts back then warned that as trains got faster, passengers would suffocate, or that looking at unnaturally fast-moving scenery could permanently compromise visual processing. Other experts warned that the human body was fundamentally unsuited to tolerate such unprecedented speeds.
Consider that up until then, no human had ever even moved faster than the speed of a horse (well, except for those falling off cliffs). The invention of technologies that could harness the chaotic energy of combustion, such as rail travel, was shocking in ways hard for us today to appreciate.
Finally, a reason to check your email.
Sign up for our free newsletter today.
In 1881, neurologist George M. Beard—credited as the first to define the condition of neurasthenia, meaning nervous exhaustion and anxiety—published American Nervousness, in which he diagnosed the then-extant, rampant anxieties in America as arising from the rapid pace of contemporaneous changes underway in society. These changes included the deployment of railroads, the proliferation of the press (due to the invention of low-cost newsprint), the emergence of a scientific enterprise, the invention of the telegraph, and the growing “mental activity of women.”
Before the telegraph, information had also never moved faster than a horse. In 1858, the New York Times reported that telegraphy could spread lies “too fast for truth” to catch up. Experts warned that the technology would create a “morbid appetite for startling news and a monomania for extravagant and almost incredible rumors.” In 1861, barely a decade into the telegraph era, the Morning Pennsylvanian observed that “so far as its communications for the public eye are concerned, it is almost an unmitigated curse.” The New York Times aired worries that the electricity in the first-ever proliferation of wires festooning the countryside could endanger animals and crops and quoted experts warning that “girdling the earth with electrical wires” could even risk destabilizing the planet’s rotation.
History is legion with examples of what we view in hindsight as hysterical and even amusing reactions to new technologies. But it seems wired into human nature to ignore the implications because, well, “this time it’s different.” It’s always the case that every new technology is different in the particulars and that otherwise smart people have said exceptionally silly things about new technologies. Reflexive presentism is hard to escape.
Cataloging history’s episodes of techno-hysteria doesn’t mean that new technologies were free of negative consequences and disruptions. But as history also shows, it’s apparently too much to expect that we will react much differently this time than we did last time.
Fast forward to 2026, and we find, as a Washington Post headline recently put it: “Extinction scenarios are taking over the AI debate.” The New York Times’s most recent analysis from its Trust team put it this way: “The questions about artificial intelligence are big, alarming and perhaps alarmist: Will A.I. systems soon move beyond human ability to control them? Could they run amok? Will A.I., in fact, kill us?”
Meantime, consumers and businesses are rushing to embrace AI. On the consumer front, the September release of Muse, Meta’s personal AI agent, set an all-time record for iOS downloads, with nearly 2 million in the U.S. alone in its first 12 days. Why? Because it was well-designed and appears useful. And, as Mark Zuckerberg explained, Meta was slower than others to release its own AI consumer product because it devoted extra time to security and safety features. As former White House AI advisor David Sacks recently noted, liability laws constitute the biggest incentive for ensuring a safe product in all the various interpretations of “safety.”
On the business front, surveys show that firms of every kind and size are testing and deploying AI tools. Measured in financial terms, overall market spending on “generative AI” tools and services has vaulted from near zero in 2022 to an almost $200 billion annualized run rate now. The growth rate in sales for AI is running at triple the velocity seen at the inception of the internet or mobile phones.
Of course, markets can eagerly adopt a new technology while simultaneously acknowledging risks. Going back to the dawn of the steam age, it took a tragic series of accidents and deaths from explosions before boiler standards were enacted. But it was the boilermakers themselves who (finally) stepped up with the epoch-setting approach to creating engineering-standards committees for safety. It took the crash of 1929 to trigger the creation of safety features for the stock market (the Securities and Exchange Commission). Today, AI pessimists argue that we should slow down before there’s a financial Chernobyl, or an actual Chernobyl, because a rogue AI occupies a reactor’s control systems.
In fact, one thing is different today. The modern world has had extensive experience with the rapid introduction of revolutionary technologies and thus the creation of expansive legal and regulatory frameworks to manage or induce good behavior and, when necessary, punish bad actors. The proper question to ask about AI safety is whether existing laws and rules sufficiently encompass risks from the new technology and, collaterally, where a tune-up or clarification may be needed. The answer to that question is rooted in a related question: What are the actual classes of risks from AI?
Real-world risks fall into just three categories: physical safety, cultural dislocations (jobs and social norms), and information security. Thus far, all the things that have happened that have raised alarms relate to AI’s impact on cybersecurity. For example, this past July’s widely publicized incident with Hugging Face saw OpenAI’s research tool break out of the “sandbox” it was supposed be restricted to and merrily “attack” Hugging Face’s open-source collection of other pre-trained AI models. (It bears noting that the Hugging Face incident was caused by programmers failing to do their job properly, not because the AI behaved unpredictably.) Similar rogue AI hacking incidents occurred before and since that one.
Few doubt that cybersecurity is critical in our modern era with so much—indeed nearly all—information exchange, especially financial activities, taking place on the internet. Thus, global spending on cybersecurity already exceeds $300 billion annually. Similarly, few doubt that AI tools make hacking easier.
Fortunately, AI also offers a means for designing cybersecurity to counter hacking threats. And even as AI makes it easier for stock scamsters and boiler-room hustlers, it also enables far more powerful scam-detection tools for citizens.
When it comes to physical risks, there is a case to be made for worrying about AI-enhanced hacking of physical, internet-connected systems, from power and water plants to “smart” appliances and factories. But this is not a new challenge in the internet era. And many, arguably most, truly critical systems are air-gapped, i.e., not connected to the public internet at all. In fact, one benefit of AI hacking threats will be to slow down the obsession to make everything smart by connecting everything—our electric grids, in particular—to the public internet.
Real reasons to worry about AI-induced physical risks are found with any autonomous or semi-autonomous machine, from surgical and industrial robots to self-driving vehicles. AI, properly trained and deployed, can add extraordinary improvements to autonomous control systems. But sloppy development can compromise safety, and that’s nothing new. Failures in software used for machinery controls or operation have always had the potential to cause physical harm.
Certainly, AI enthusiasts should have to meet the same standards of quality assurance and safety that we have applied for decades to industrial-class software. Those domains have adults in the room and well-established protocols, procedures, and liabilities.
The longer-term physical risks being pondered mainly involve worries that AI will accelerate the ability for bad actors to develop novel, more dangerous kinds of chemical and biological threats. Again, though potentially dangerous and genuine, these are not new classes of risks. Scientists and regulators have wrestled with both since the dawn of chemical and biological sciences. That AI-enhanced detection will improve our ability to find bad actors seems likelier than AI enhancing the physical ability to deploy such weapons.
As far as conventional weapons themselves, the addition of autonomous guidance is not new. The distinction that makes a difference here is scale, especially with drones. Again, it is a normal pattern in history. That does not make it any less serious, but it doesn’t rise to end-of-humanity tropes.
As for energy, water, and land use—the real-world issues that catalyzed the data-center resistance movement—let’s take them one by one.
Regarding energy, we can compare the power needed to move information versus the power required to transport physical goods. By 2035, if the U.S. ends up with all the AI data centers now forecast, the total power required by all of them will be less than one-tenth the total collective power under the hoods of all Class 8 trucks moving goods on the nation’s roads. The data-center demand hardly seems a heavy lift.
As for water and land, data centers use far less of both than golf courses, never mind agriculture. Where were the protesters when Congress enacted mandates and subsidies for corn-ethanol in gasoline? Nearly 200 times more water is used per gallon to produce ethanol than gasoline.
Finally, on land use: the total square footage of warehouses in America is, at present, almost 100-fold greater than for data centers. Again, where were the demonstrators when politicians and promoters started subsidizing solar power installations that blanket farmland and fragile deserts by the square mile with black, glass-covered silicon?
This brings us to cultural and social issues, from disinformation and fake news to social-media misuses, doom-scrolling, and job displacement. All are serious in their particulars; none is unique in character. On the jobs front, even as AI is being rapidly adopted, tracking data show no evidence of negative trends for either layoffs or unemployment. The jobs apocalypse exists only in cherry-picking examples or hyperbolic scenarios. While none can predict the future, history again offers useful lessons. As MIT economist David Autor documented, using data from World War II to date, a period of profound technological changes, over 60 percent of employment today involves work in job categories that didn’t previously exist.
Nvidia CEO Jensen Huang recently remarked on AI fears, giving a “0% chance” of the technology ending the world by 2030, as some critics warn is possible. Nonetheless, the hype machine often runs faster than truth can keep up. That’s another old problem, but one that matters—because politics matters.