MIT Technology Review will convene three of its senior staff—executive editor Niall Firth, senior AI editor Will Douglas Heaven, and AI reporter Grace Huckins—for a subscriber-only roundtable on Tuesday, September 15, to interrogate a question that has migrated from science fiction forums into the internal memos of the world's most valuable companies: could advanced AI kill us all? The session, scheduled for 16:00 BST / 11:00am EST / 8:00am PST, exists because employees at leading AI labs are now saying publicly what they once only whispered: that the technology they are building carries a genuine risk of human extinction. That claim deserves scrutiny, not dismissal—and not panic. Meanwhile, in a laboratory a few miles from the roundtable's intellectual center of gravity, a different kind of existential question is being pursued with equal intensity. Geneticist Yuancheng (Ryan) Lu is obsessed with aging, with eyes, and with the possibility that the two might be reversible.
AI Labs Sound the Alarm: Could Advanced AI Destroy Humanity?
The most striking fact about the current AI extinction debate is who is having it. This is not a fringe conversation confined to philosophers or ethicists. The people building frontier models—employees at the world's leading AI labs—are on record saying that advanced AI could plausibly destroy humanity. When the inventors of a technology assign meaningful probability to its catastrophic misuse or loss of control, the rest of us have an obligation to listen carefully rather than reflexively.
The debate is not settled, and the roundtable format reflects that. Are these employees right? Or is this scaremongering and hype—a way for labs to appear responsibly cautious while racing ahead? These are the two poles of the discussion, and both deserve a hearing.
There is empirical context for the alarm. In a 2022 survey conducted by the research group AI Impacts, roughly 48% of machine learning researchers assigned at least a 10% probability to existential-level harm from advanced AI. That number does not settle the question. A 10% chance of a civilization-ending outcome is, by any standard, an enormous risk—and the fact that nearly half of surveyed researchers were willing to countenance it tells you the concern is not confined to a handful of doomsayers. At the same time, probability estimates from experts are notoriously unreliable, and the methodologies behind them are contested.
The roundtable will explore where these fears come from. Some trace to alignment problems—the difficulty of ensuring that systems more capable than their creators pursue goals compatible with human flourishing. Others focus on misuse: the acceleration of bioweapons development, cyberattacks, or disinformation at scales that overwhelm institutional defenses. Still others argue that the real threat is not a sudden takeover but a slow erosion of human agency as critical decisions are ceded to opaque systems. Whether these fears "hold any water," as the roundtable's framing puts it, is precisely the kind of question that honest inquiry should not pre-judge.
What AI Extinction Risk Actually Means
Start with a definitional problem. "AI extinction risk" is a phrase that collapses several distinct scenarios into one ominous label, and the imprecision does the conversation no favors. The first scenario is deliberate misuse: a bad actor uses advanced AI to engineer a pathogen, destabilize infrastructure, or circumvent safeguards at a speed no human institution can match. The second is loss of control: a system pursuing a goal we did not intend, in ways we cannot anticipate or reverse. The third is structural: not a single catastrophic event but a gradual transfer of decision-making authority from humans to machines until the capacity for self-governance has quietly atrophied.
Read next Top Technology Trends in 2026 You Need to KnowEach scenario implies different remedies. Misuse demands regulation, monitoring, and control of dangerous capabilities. Loss of control demands technical alignment research—the unglamorous work of making systems that reliably do what we mean. Structural erosion demands political and institutional vigilance of a kind that no lab can supply on its own. Lumping them together as "extinction risk" can obscure the fact that some are more tractable than others, and that some are more likely than others.
The 48% figure from AI Impacts is best understood as a signal of expert uncertainty rather than a forecast. When nearly half of a field's practitioners decline to rule out a catastrophic outcome, the rational response is neither dismissal nor despair. It is to treat the question as open, fund the research that could close it, and design governance that performs well under uncertainty. That is a harder posture than either alarmism or complacency—and it is the one the roundtable appears designed to model.
Breakthrough Science: Reversing Eye Aging to Restore Sight
Outside the Whitehead Institute in Cambridge, Massachusetts, Yuancheng (Ryan) Lu's aviator glasses darken automatically in the sun—a small, practical technology that happens to sit on the face of a man trying to reverse one of the most consequential forms of biological decay. Lu is a geneticist, and his obsession is aging. His focus, specifically, is eyes.
The personal dimension is not incidental. Age-related blindness runs in Lu's family, and his own 23andMe test returned a result that sharpened his attention to the problem. That kind of motivation—the scientist who studies what threatens him and the people he loves—has a long history in medicine, and it functions as a credibility signal: this is not an abstract research program but a problem with a face attached.
The stakes are enormous and well documented. Age-related macular degeneration is among the leading causes of irreversible vision loss worldwide. Estimates cited in global health literature put the number of people affected by AMD at roughly 196 million globally by 2020—a figure that reflects both the prevalence of the condition and the simple arithmetic of an aging population. As life expectancy rises, so does the denominator of people at risk.
Lu's work sits at the frontier of age-reversal research, a field built on the observation that aging is not a single process but a collection of biological mechanisms that can, in principle, be manipulated. The eye is an unusually attractive target for this kind of work. It is small, accessible, and partially immune-privileged—properties that make it easier to deliver therapies and observe results than in organs buried deeper in the body. If age-reversal techniques can be made to work anywhere, the retina is a plausible first proving ground.
The reported summary is necessarily thin on the specific mechanisms Lu is pursuing, and that caution is appropriate: age-reversal science is early, contested, and prone to overhyped headlines. What can be said is that the direction of travel is real. Restoring sight lost to age-related disease would be a concrete, measurable win—the kind of outcome that transforms a research program from promising to proven.
The Intersection of AI Safety and Biotech Innovation
The two stories in this edition of The Download are usually treated as belonging to separate universes: one about machines that might escape our control, the other about biology we are learning to control. They are more entangled than that.
Advanced AI is already accelerating biological research—protein folding, drug discovery, the design of experiments—and that acceleration cuts both ways. The same capabilities that might help Lu's lab identify age-reversal targets faster could, in the hands of a bad actor, lower the barriers to engineering dangerous pathogens. This is precisely why the misuse scenario in AI risk discussions often centers on biosecurity. The roundtable's question about whether AI could kill us all has a biological wing, and it is not hypothetical.
There is also a methodological overlap. Both fields are grappling with the limits of expert prediction. Just as AI researchers disagree wildly about the probability of catastrophe, aging researchers disagree about which interventions will translate from mice to humans, and on what timeline. In both cases, the responsible posture is the same: acknowledge uncertainty, insist on evidence, and resist the pressure to declare victory or doom prematurely.
The institutional anchors matter here. MIT Technology Review has covered both AI risk and biomedical research for decades, and the Whitehead Institute is one of the world's leading centers for genetics and genomics. These are not sources given to sensationalism. When they host a conversation about AI extinction or profile a geneticist pursuing age reversal, the signal is that these questions have moved into the mainstream of serious scientific inquiry.
What These Developments Mean for the Future of Technology
Two futures are being negotiated at once. In one, AI systems become powerful enough that their safe governance is the central political problem of the century—and we have not yet built the institutions to solve it. In the other, the biological damage of aging becomes a tractable engineering problem, and conditions like age-related macular degeneration stop being irreversible sentences.
Neither future is guaranteed. The AI extinction debate could resolve into overcaution that stifles beneficial research, or into complacency that leaves the world unprepared. The age-reversal work could stall in the long, unglamorous gap between promising mechanisms and proven therapies—a gap that has swallowed many bold claims before.
What the roundtable and Lu's work share is a refusal to accept the current state of affairs as fixed. One asks whether the most powerful technology humans have ever built might destroy us, and what we should do if the answer is anything other than a confident no. The other asks whether the most universal form of human decline might be slowed, or reversed, starting with the eyes. Both questions are open. Both deserve the kind of evidence-driven scrutiny that MIT Technology Review's roundtable format is designed to provide—and both will be answered, one way or another, by what we choose to do next.
Source: MIT Technology Review
