MIT Technology Review has turned its attention to one of the most charged debates in artificial intelligence: whether the technology poses a genuine extinction-level risk to humanity, and specifically whether it could be weaponized to enable the creation of biological weapons. The outlet's coverage stems from a live Roundtables event it hosted, in which experts gathered to address the question of whether AI could, in some plausible scenario, kill us all.
To understand why this conversation is happening now, it helps to trace how the framing around AI risk has shifted over the past few years. Not long ago, mainstream discussions of AI danger centered on comparatively near-term concerns: algorithmic bias, job displacement, surveillance, the spread of misinformation. Those remain serious and active areas of policy concern. But a separate and older strand of thinking, long associated with academic philosophers and a cluster of researchers loosely grouped under the banner of "AI safety," has moved from the intellectual margins toward the center of public discourse. The shift accelerated as large language models became capable enough to surprise even their own developers, and as figures inside major AI laboratories began signing open letters and making public statements warning that the technology they were building might pose catastrophic risks if misaligned with human intentions or placed in the wrong hands.
The bioweapons thread is particularly significant and deserves its own unpacking. The concern is not primarily that an AI system develops its own agenda and synthesizes a pathogen. The more immediate worry among biosecurity researchers is that AI dramatically lowers the barrier of expertise required for a malicious actor to design or enhance dangerous biological agents. Historically, the technical knowledge needed to work with select agents or to engineer novel pathogens represented a meaningful obstacle. The question researchers are now asking seriously is whether sufficiently capable AI models, given the right prompting, could compress years of specialized graduate training into an accessible interface. Several research groups and government-adjacent bodies have already run preliminary evaluations on this question, and the results have been sufficiently unsettling that some have declined to publish them in full.
This sits within a broader pattern of the AI industry grappling, often uncomfortably, with dual-use concerns it did not fully anticipate at the outset. The architecture that makes large language models useful for medical diagnosis, software development, and education is the same architecture that makes them potentially useful for someone attempting to cause mass harm. There is no clean technical separation. The industry's response has so far been a patchwork of content filtering, usage policies, and red-teaming exercises, none of which independent researchers regard as fully satisfying.
The extinction risk framing is more contested and the likely reading is that it divides even specialists who agree on the underlying technical trajectory. Critics argue that the leap from "this technology is becoming very capable" to "this technology could end human civilization" involves a chain of assumptions that remains speculative and that the urgency around existential risk can crowd out more tractable near-term harms. Proponents counter that the asymmetry of the downside justifies serious attention even if the probability is low. What is notable is that this debate is now happening in venues like MIT Technology Review's own events rather than in niche forums and preprint archives, which suggests the Overton window around these topics has moved considerably.
The consequences of this conversation play out across several groups simultaneously. For AI developers, growing public and regulatory attention to catastrophic risk scenarios creates pressure to demonstrate that safety evaluation is rigorous rather than performative. For governments, particularly those developing AI governance frameworks, the bioweapons question specifically has already influenced some early proposals around model evaluations and access controls. For biosecurity institutions, the implication is that their expertise is newly relevant to a technology sector they had little previous engagement with, which creates both opportunity and strain. And for the public, the challenge is navigating a debate where credible experts hold genuinely divergent views and where the evidence base is still being constructed in real time.
Several things are worth watching in the months ahead. How AI developers respond to the growing literature on biological risk evaluations will be telling, particularly whether they move toward independent third-party audits or continue to rely primarily on internal red teams whose findings are selectively disclosed. Regulatory bodies in the European Union and the United States are both, through different mechanisms, attempting to build evaluation requirements into law, and how they treat the highest-capability models will set a precedent. And the academic community will continue producing research that either tightens or loosens the evidentiary case for catastrophic risk, research that events like MIT Technology Review's Roundtables will increasingly help translate for a broader audience.
