Will AI Destroy Humanity? MIT Magazine Publishes Expert Answers

According to reporting by Will Douglas Heaven and Grace Huckins of MIT Technology Review AI, a subscriber-only public discussion was held on the possibility that AI could destroy humanity. It was held on Wednesday, September 17, 2026. The format is an interactive program called Roundtables. The theme was the question, “Could AI really kill us all?” Thirty minutes of discussion was not enough time, and some of the questions received were answered in writing afterward. The respondents were senior AI editor Will Douglas Heaven and AI reporter Grace Huckins.

This article is based on MIT Technology Review AI (All Rights Reserved) and relies on fair quotation under Article 32 of the Japanese Copyright Act. The core of the original is the distinction between individual harm and the extinction of humanity as a whole.

Background and Aim of the Subscriber-Only

Roundtable

The magazine operates Roundtables for direct dialogue with subscribers. The theme this time was extinction risk, the most high-profile issue in the current AI debate. The background is a series of predictions about improving capabilities and alignment problems. Questions from participants focused on the possibility of death, motives, and defenses. The editorial team selected representative questions and published the views of the two side by side. The positions of the two conflict in part and overlap in part. That difference itself reveals the structure of the controversy. A distinctive feature is that no single conclusion is presented. Facts and evaluations need to be read separately.

Huckins begins her discussion from the unavoidable fact of individual deaths. Her position is that the possibility of AI being the cause cannot be denied. In Ukraine, AI-equipped drones are already killing and injuring people. AI-driven cyberattacks targeting hospitals are also highly likely to cause harm. These realities are past the stage of speculation. On the other hand, she rates the death of all humanity as unlikely. She also acknowledges the fact that some intellectuals have continued to warn about it for years. Although they are sometimes described as eccentric, she evaluates them as having a deep technical understanding. She also emphasizes that their past predictions about capabilities and alignment have been accurate. Accuracy does not guarantee the correctness of momentous predictions. Still, she takes the stance that they deserve attention.

A Perspective That Distinguishes Individual

Deaths from Human Extinction

Heaven clearly separates individuals from humanity as a whole. His judgment is that the possibility of individuals dying due to AI is not zero. He envisions unfortunate accidents or near-future anomalies. As an example, he cites cyberattacks on infrastructure by groups of AI agents. Scenarios once dismissed as absurd are gaining plausibility. He also lists the spread of an AI-designed novel pathogen as a candidate. He also touches on a scenario in which the collapse of the world economy leads to conflict and famine. He rates the latter two as unlikely. However, he rejects a scenario in which all of humanity dies. He asserts that such conditions do not exist outside apocalyptic science fiction. His argument is that horror stories can be invented endlessly, but they are not rooted in the actual level of technology. Some argue that preparing for the worst is harmless. He criticizes such catastrophic discourse for obscuring the challenges right in front of us. His logic is that it makes the problems of existing technologies and the responsibilities of developing companies harder to see.

This distinction is also important in policy debates. Individual harm leads to concrete, defensible measures. Protecting hospitals, electricity, water, and other systems is highly urgent. As shown by the Emergence of “HalluSquatting” Attacks Exploiting AI Hallucinations, attack methods are evolving by exploiting hallucinations and automation. Infrastructure defense is becoming more concrete. In contrast, human extinction tends to rely on unverifiable assumptions. It risks misallocating resources. There is a need to discuss the two without confusing them.

Examining the Plausibility of Weaponization

and Biological Threats

The answer on why AI would kill people is concise. In the original, it is expressed as follows.

Someone might tell it to, and it might listen.

This means someone might order it to kill, and AI might obey. Importantly, it does not assume autonomous hatred or will. It points to the danger of AI as a tool that obeys human orders. Concern is particularly strong about biological capabilities. There is fear that pathogen-design assistance capabilities could be used by attackers. As an example, it mentions Aum Shinrikyo, which carried out the 1995 Tokyo subway sarin attack. It envisions a case in which a pathogen more lethal than Ebola and more infectious than measles is designed. Defenders must prepare for every conceivable weapon. Attackers achieve their goal if they succeed in producing just one type. This asymmetry makes defense difficult. The spread of general-purpose AI lowers barriers of expertise. Access to experimental procedures and design knowledge becomes easier. Strengthening information controls and detection systems is a challenge.

Biological threats differ in nature from nuclear and cyber threats. Manufacturing facilities could be relatively small-scale. Control after spread is extremely difficult. Prior monitoring and material management are therefore emphasized. Capability evaluations and disclosure standards by developing companies are also at issue. Although on a different level from rights disputes such as the Round Hill Lawsuit Against Suno and Anthropic for Over $1 Billion, there is a common challenge of managing training data and weights. The question is how to balance capability transparency with misuse prevention.

Concerns That Doomsday Narratives Downplay

Immediate Challenges

Heaven warns against leaning toward catastrophism. His criticism is that attention gathers on distant extinction while current harms are overlooked. Killing by drones is already a reality. Attacks on medical institutions directly cost lives. Impacts on the labor market and information environment are also underway. Accountability and safety evaluations by developing companies remain inadequate. Doomsday narratives push such issues into the background. They distort the prioritization of countermeasures. Huckins is also dismissive of stockpiling canned goods or befriending bunker owners. She shows a stance of not contributing to exaggeration. Both share a calm assessment rather than alarmism. The difference lies in how they face uncertainty. Huckins accepts the weight of predictions that proved correct. Heaven prioritizes measures grounded in reality. This tension is at the heart of the debate.

In defensive practice, basic hygiene is also important. Establishing verifiable tools and operational procedures reduces harm. Interest in transparent maintenance tools such as the New Open-Source CCleaner Alternative FluentCleaner illustrates part of this. Accumulated efforts to strengthen endpoint defenses determine the resilience of infrastructure as a whole. Everyday operational improvements can be more effective than grand narratives.

Safety Measures and the Challenge of Defender

Asymmetry

Defensive asymmetry cannot be resolved by technology alone. Institutions and international cooperation are needed. One example is managing the distribution of pathogen-related knowledge and reagents. Restrictions on AI-generated procedures are also debated. Excessive restrictions would hinder research freedom and medical progress. Setting the scope is difficult. In cyber defense, strengthening systems at hospitals and local governments is urgent. Talent shortages and budget constraints are common barriers. Export controls and operational norms are relevant to drone autonomy. Diversion of civilian technology makes regulation difficult. Progress in alignment research is also essential. Criteria for obeying and refusing orders need to be clarified. Standardization of evaluation methods is lagging. Continued public discussion is a prerequisite for understanding. This Roundtables event is one such attempt.

Editorial Opinion

On short-term impact. This summary is practical in that it separated individual harm from extinction. In the next 3 to 6 months, defense investment for hospitals and infrastructure will likely be at the center of debate. Pressure to disclose evaluations of biological capabilities will also likely intensify. Companies will likely be pressed to explain capability limits. Calm measures that avoid exaggeration will likely gain support.

On the long-term perspective. In 1 to 3 years, the design of order obedience and refusal will likely become a competitive axis. Management of pathogen-related knowledge will likely require an international framework. Without progress in standardizing verification methods, distrust will likely remain. Balancing transparency with misuse prevention will likely determine market trust. Balancing the benefits of technology with control will become a challenge.

Questions from the editorial team. How to close the asymmetry of defense is the question. The reality that a single success is enough for attackers should be faced squarely. There is no universal solution. Which should be prioritized, strengthening everyday defenses or institutional design. Discussion of concrete measures beyond the extinction debate will likely be needed.

References

• “Could AI really kill us all? Your questions, answered.”, by Will Douglas Heaven, Grace Huckins — MIT Technology Review AI, 2026-09-18T11:29:22.000Z (ARR)
• Source URL: https://www.technologyreview.com/2026/09/18/1144435/could-ai-really-kill-us-all-your-questions-answered/