When “Could AI Kill Us?” Gets Asked Out Loud
MIT Technology Review held a live Roundtables event this week built around a question that has migrated from fringe forums to mainstream conference rooms: could AI actually kill us all? Senior AI editor Will Douglas Heaven and AI reporter Grace Huckins fielded the questions that outlasted the event’s allotted time – questions including whether individuals should be personally afraid, whether the danger is real or a public relations strategy by tech companies hungry for regulatory attention, and what functional oversight of AI systems would even look like in practice.
The event didn’t resolve these questions. That’s the point.
What made the session notable wasn’t the panic, but the specificity of the concerns. Attendees weren’t asking vague questions about robots or job losses. They were asking about governance structures, about who monitors the monitors, and about whether the danger is being exaggerated for competitive advantage by the same companies building the systems. Heaven and Huckins published their responses after the event, working through the logic of extinction-level AI risk without defaulting to either dismissal or catastrophism. The answers, deliberately, don’t close the debate.
The Bioweapons Problem Is Already Here
Among the specific harm scenarios that AI researchers take seriously, bioweapons sit near the top. The concern is not theoretical. In 2022, researchers demonstrated that an AI model built for pharmaceutical drug development – a molecule generator – could be redirected toward harm in under six hours, producing 40,000 molecules with potential applications as chemical warfare agents. The tool was designed to save lives. The redirection required minimal effort.
That experiment happened before the current generation of general-purpose AI systems became widely accessible. Today, AI tools can field questions across almost every scientific domain, and parallel advances in gene editing and synthetic biology have lowered the barrier to entry for biotech work broadly. Journalist Jessica Hamzelou, writing for MIT Technology Review’s Checkup newsletter, reports that safeguards exist across the pipeline – but none are absolute, and scientists themselves are split on exactly how serious the threat has become. The disagreement is not about whether the tools exist, but about how much capability a bad actor would still need to bring independently before AI assistance becomes decisive.
The harder problem is that the same openness that accelerates legitimate science accelerates the rest of it too. Gene editing tools that help researchers study infectious disease at a university lab are structurally the same tools that could help someone design a pathogen with enhanced transmissibility. AI doesn’t create that tension – it sharpens it by compressing the time and expertise required to move from idea to molecule.
Security Failures, Math Claims, and a Web Being Eaten Alive
Separate from the existential debate, a set of more immediate AI stories broke this week that deserve attention on their own terms. Security researchers breached OpenAI’s internal systems by exploiting a vulnerability in a third-party forum, reaching an employee’s ChatGPT account and internal code. The breach was accomplished using tools developed by Anthropic – a detail that underscores how the competitive AI landscape creates unexpected interdependencies, where one company’s research becomes another company’s attack surface.
OpenAI is also reportedly expecting to solve the Hodge Conjecture, a major unsolved problem in mathematics. The announcement, if it comes, will arrive in a context already shadowed by controversy – the company’s previous math claims drew significant backlash, and MIT Technology Review has reported that those controversies surface troubling patterns about how mathematical progress in AI is being measured, communicated, and verified. Whether OpenAI has solved the Hodge Conjecture matters less, at this stage, than whether the field has agreed on what “solved” means.
Meanwhile, internal voices at both Microsoft and OpenAI have raised alarms about AI’s effect on the broader web. Reported by 404 Media and surfacing in the New York Times copyright case against OpenAI, employees expressed concern that AI systems scraping and summarizing web content are destroying the business models that fund original reporting and publishing. When AI answers a question directly, fewer users click through to the source – and fewer clicks mean less revenue for the outlets that produced the underlying information. The comments emerged in a legal context, which means they carry weight beyond opinion: they could weaken OpenAI and Microsoft’s defense in the copyright suit, according to Reuters.
Ukrainian forces sank a Russian vessel using an autonomous boat – the first recorded instance of robot watercraft fighting each other in combat – while U.S. firms are separately developing combat-ready humanoid robots, according to the Wall Street Journal. The military applications of autonomous systems are moving faster than most public-facing AI policy discussions acknowledge.
The Questions Nobody Is Rushing to Answer
Taken together, this week’s stories describe an AI landscape where the largest companies are simultaneously building systems capable of serious harm, suffering security breaches through those same systems, and being implicated – by their own employees – in the slow economic strangulation of the information ecosystem those systems were trained on. The MIT Technology Review Roundtables event asked whether AI could kill us. The more immediate question, sitting inside the New York Times lawsuit, is whether AI is already killing the web that made it possible – and what publishers, regulators, and the companies themselves intend to do before there’s no fresh content left to scrape.
