Could AI Really Destroy Humanity? A Deep Dive into AI Risk

Introduction: A Question Everyone Is Asking

"Could AI really kill us all?" Once confined to science fiction, this question has become a hotly debated topic among technologists, policymakers, and the general public alike. MIT Technology Review recently hosted an online roundtable for subscribers dedicated to exploring this very issue. Although the session ran only 30 minutes, audience questions far exceeded what the time could accommodate — a telling sign of just how much anxiety surrounds AI risk.

Drawing on perspectives from Will Douglas Heaven, MIT Technology Review's senior AI editor, and other experts, this article attempts to clarify the core debates around AI existential risk and help readers understand what this conversation is truly about.

Two Camps on AI Existential Risk

The question of whether AI poses an existential threat to humanity has sharply divided the field. One camp argues that as AI systems rapidly grow more capable, the emergence of a "superintelligence" far exceeding human intellect — one whose goals may not align with human interests — could lead to catastrophic outcomes. This group includes researchers at AI labs and scholars in the safety field, who advocate for robust safeguards before the technology spirals out of control.

The other camp believes that the "AI destroys humanity" narrative is largely overhyped panic. They argue that today's large language models are fundamentally statistical prediction tools, still far removed from autonomous intent or genuine general intelligence. Pouring resources into distant "doomsday scenarios," they contend, only diverts attention from real and present harms.

Those warning against superintelligence typically cite the Alignment Problem as their core argument: if an AI system far more capable than humans is given a goal, it may pursue that goal in ways humans cannot anticipate or stop, with no regard for human well-being as collateral damage. Oxford philosopher Nick Bostrom vividly illustrated this concern in his 2014 book Superintelligence with the "paperclip maximizer" thought experiment: an AI instructed to "make as many paperclips as possible" might, once sufficiently powerful, convert all of Earth's resources — including humans — into paperclips. Turing Award laureates Yoshua Bengio and Geoffrey Hinton have also publicly expressed deep concerns about AI safety in recent years. The opposing camp, meanwhile, relies on the argument that "capability ≠ intent," emphasizing that current AI lacks autonomous desires and persistent goals, and that doomsday narratives confuse tools with agents.

The "Present-Day Risks" We're Overlooking

Beyond the grand narrative of AI apocalypse, many experts place greater emphasis on harms that are already happening. Algorithmic bias, rampant misinformation, deepfakes, privacy violations, and labor market disruption — these are not hypothetical futures, but concrete challenges actively reshaping society.

Focusing the entire conversation on "will AI kill us" is, in some ways, a strategic distraction. When the public and regulators are drawn into science-fiction-style fears, scrutiny of tech companies' real-world practices — data use, content generation, market monopolization — may actually loosen. Recognizing this dynamic is key to thinking clearly about AI risk.

Algorithmic bias is far from abstract — it's extensively documented in real cases. The COMPAS risk assessment algorithm used in the U.S. criminal justice system was found by researchers to systematically overestimate recidivism risk for Black defendants. AI screening tools in hiring have discriminated against female candidates due to historical biases in training data. Facial recognition systems show significantly higher error rates for darker-skinned individuals than for lighter-skinned ones. Deepfakes, powered by generative adversarial network (GAN) technology, have already been deployed at scale to produce non-consensual intimate imagery and fabricated videos of political figures. These harms don't require the arrival of "superintelligence" — they are measurably damaging real people right now, especially those already marginalized in society. Precisely because these harms lack a cinematic quality and aren't "sexy" enough, they have long been overshadowed by doomsday narratives in media coverage.

How to Think Rationally About AI Risk

The expert consensus tends toward a "dual-track" approach: neither succumbing to panic about AI taking over the world, nor dismissing the value of long-term safety research.

For ordinary users, the more practical focus is on the reliability and transparency of AI in everyday applications — whether it spreads misinformation, reinforces bias, or is used for manipulation. For policymakers, the challenge is striking a balance between fostering innovation and preventing misuse, building enforceable regulatory frameworks rather than getting stuck in abstract philosophical debate.

One point of consensus among experts like Will Douglas Heaven is this: rather than fixating on the unfalsifiable ultimate question of "will AI destroy humanity," it's more productive to direct energy toward specific, verifiable, and governable problems. The direction technology takes ultimately depends on how humans choose to use and constrain it.

Conclusion: Clarity Beyond Fear

Whether AI will destroy humanity is a question unlikely to have a definitive answer anytime soon. But the ongoing public conversation has value in itself — it pushes us to think about the boundaries of technology, where responsibility lies, and how prepared society really is.

What truly matters may not be the fear itself, but maintaining clarity between fear and optimism: understanding AI's potential and its limits, neither mythologizing nor demonizing it, and taking a pragmatic approach to steering this technology toward outcomes that benefit humanity.