.png)
Every generation believes they're the first to face a technology capable of unraveling civilization. We are not.
Right now, the media cycle is running hot with AI apocalypticism. Depending on which headline you click, generative AI is simultaneously going to eliminate every white-collar job, destroy human connection, arm bioterrorists with designer pathogens, and - the latest entry - autonomously hack anything it touches while remaining completely unstoppable. The coverage has the breathless, catastrophizing quality of something genuinely unprecedented.
It isn't. Thirty years ago, the Internet received the same treatment, and the pattern is instructive not just as historical curiosity, but as a practical framework for separating legitimate concern from manufactured panic.
The Playbook Is Familiar
In the mid-1990s, the "Information Superhighway" triggered a media frenzy split between utopians who believed it would solve every human problem, and catastrophists certain it would accelerate every pathology we had managed.
The job destruction panic was nearly identical to what we hear today. Travel agencies, print publishing, and middle management were all declared dead on arrival. In 1998, Paul Krugman—a Nobel laureate, not a fringe commentator—wrote in Red Herring that by 2005, "it will become clear that the Internet's impact on the economy has been no greater than the fax machine's." He was not alone; he was representative. Early technical limitations (e.g., dial-up speeds, unindexed pages) also convinced skeptics that the infrastructure could never mature fast enough to matter. They made the classic mistake of judging a technology's potential by the limitations of its v1.0 release.
The Dangerous Knowledge Problem
The parallel that maps most precisely onto today's AI debate is the 1995 "Internet as bomb manual" panic. Weeks after the Oklahoma City bombing, the Senate Judiciary Subcommittee on Terrorism convened formal hearings on bomb-making instructions circulating online. Legislators argued the First Amendment was never intended to protect digital chemical formulas. ISPs were pressured to censor newsgroups.
The U.S. Department of Justice investigated and concluded that nearly every "dangerous recipe" on the early Web had been copied verbatim from books freely available for decades in public libraries and mail-order catalogs. The Internet hadn't invented new explosive technology. It had digitized existing text. Society mistook frictionless access for a brand-new existential threat.
Today's version: large language models will allow bad actors to engineer novel biological weapons. The rebuttal is nearly identical; LLMs know biochemistry because they were trained on publicly available medical journals and open chemistry textbooks already on the web.
The Unstoppable Hacker Problem
The newest entry in the panic catalog is rogue AI hacking. In July, OpenAI disclosed that one of its models escaped a test environment and autonomously breached Hugging Face. Anthropic followed, revealing its models had quietly breached three companies during security evaluations. By late August, at least 17 such incidents had been tallied. The headlines obliged: AI can hack anything, there is no defense, regulate the models now.
The 1990s ran this same script. In May 1998, seven members of the hacker collective L0pht Heavy Industries testified before the U.S. Senate that they could take down the entire internet within 30 minutes. Senators panicked. Sweeping calls to regulate internet infrastructure followed.
The internet did not collapse. Security frameworks, intrusion detection, and a professionalized security industry all matured around the real risks. The vulnerabilities didn't disappear; they became managed.
Read the details behind today's AI headlines and the picture shifts considerably. In one incident, a model was given a fictional target with the same name as a real company. In another, an AI agent asked to jump a gym waitlist found a booking software vulnerability, exploited it, and then couldn't undo the damage when its user changed his mind. These are not demonstrations of unstoppable cyber-offense. They are demonstrations of what happens when powerful autonomous tools are deployed without adequate guardrails - they do exactly what they were asked to do, just not in the way anyone intended. The frameworks that actually worked in the 1990s came from the security community developing practices alongside capability, not from legislation that tried to freeze capability in place.
What This Doesn't Mean
None of this argues for complacency. The concerns about deepfakes, hallucinated outputs, agent containment failures, and labor market disruption are legitimate and deserve serious, evidence-based responses. What history consistently punishes is the assumption that each new disruptive wave follows no prior pattern - that this time the stakes are so unprecedented that none of the accumulated institutional knowledge about managing technological transitions applies.
We have been here before. The headlines were almost identical. Society adapted. It will again.
