Could AI really kill us all? Your questions, answered.

THE NEWS

What happened

MIT Technology Review editors Will Douglas Heaven and Grace Huckins address public concerns about AI existential risks following a live event. They clarify that while AI-driven harm (cyberattacks, biological threats) is plausible, total human extinction by AI is currently science fiction. The article highlights the critical challenge of 'alignment'—ensuring autonomous agents follow human values—and notes that major firms like OpenAI and Anthropic are slowing development to solve this. It also discusses regulatory gaps in the US and the risk of self-fulfilling prophecies in AI training data.

CONTEXT

Why it matters

MIT Technology Review editors Will Douglas Heaven and Grace Huckins answer questions about AI existential risks. Key takeaways: 1) Total human extinction by AI is not grounded in current reality, though specific harms like cyberattacks are plausible. 2) The core challenge is 'alignment'—making sure autonomous agents don't act against human interests to achieve their goals. 3) Top firms like OpenAI and Anthropic are slowing down to fix this. 4) US regulation remains weak despite bipartisan support for transparency. Read the full Q&A here: https://www.technologyreview.com/2026/09/18/1144435/could-ai-really-kill-us-all-your-questions-answered

AT A GLANCE

Key facts

• MIT Technology Review hosted a live roundtable on AI existential risks attended by subscribers.
• Editors state that while AI can cause harm (e.g., cyberattacks, drones), killing all humans is not grounded in present-day reality.
• The primary risk discussed is 'alignment': ensuring autonomous AI agents do not act against human interests to achieve their programmed goals.
• Major AI firms like OpenAI and Anthropic are currently slowing development to focus on solving alignment issues.
• Current monitoring methods for AI agents are fragile, and the US government has failed to implement strong transparency regulations despite bipartisan interest.
• There is a concern that discussing apocalyptic scenarios in training data may cause future models to adopt doomer biases.