AI can destroy humanity – an illustrated guide

An illustrated field guide lays out how advanced AI can cause serious harm, organizing risks by four failure modes - deliberate misuse, accidents, competitive races that cut safety, and misaligned objectives - and walking each risk through how it works, a documented case, the worst credible outcome, and possible defenses. It cites a 2023 survey of almost 2,800 AI researchers whose median assessment put the chance of an “extremely bad” outcome like extinction at 5%, and a 2026 international safety report that calls present decision-making an “evidence dilemma.” Early chapters focus on harms already unfolding; later ones examine emerging or hypothetical catastrophes. The guide repeatedly stresses that experts disagree sharply about probabilities even as some dangers are already measured.

Concrete chapters detail deepfakes (real-time video and voice forgery that enabled a US$25 million transfer from Arup’s Hong Kong office and cloned-voice robocalls), pervasive surveillance (facial recognition errors that produced wrongful arrests, and AI-enabled social control), economic concentration and fragility (large shares of workers exposed to automation, rising datacenter energy use, and Financial Stability Board warnings of vendor dependence), and faster cyberattacks as AI automates vulnerability discovery. Suggested countermeasures include content credentials and disclosure rules, bans on the most dangerous surveillance uses, independent audits, stress tests, retraining and redistribution policies, infrastructure planning, and legal limits on deployment - measures presented as necessary because some harms are documented while others remain speculative but plausible.