The probability of advanced artificial intelligence triggering a catastrophic threat to human survival is not zero, according to Lila Ibrahim, Google DeepMind. While industry leaders and researchers debate the exact risks of future systems, experts remain divided over how to quantify threats ranging from loss of control to accelerated bioweapons development.
Quantifying the Unknown Risks of Superintelligence
This cautious stance contrasts with other prominent figures in the field. Geoffrey Hinton, a Nobel laureate and pioneer in deep learning, has estimated a 10% to 20% chance that advanced AI could cause human extinction. Meanwhile, Anthropic CEO Dario Amodei calculated in 2025 that there is roughly a 25% probability of outcomes going “really, really wrong.”
These divergent estimates highlight a fundamental challenge in AI safety: the lack of historical data or past catastrophes to build predictive statistical models. According to the International AI Safety Report, specialists remain split. Some researchers view a future loss of control over advanced systems as plausible, while others dismiss these scenarios as implausible because current models still lack the necessary capabilities to execute such events.
Current Capabilities Versus Autonomous Threats
Hollywood-style scenarios where a machine suddenly decides to destroy humanity do not accurately reflect the concerns of computer scientists. Instead, researchers at institutions tracking AI advancement focus on systems designed to operate over long periods, evade supervision mechanisms, conceal specific actions, or pursue objectives that conflict with their developers’ intentions. While current models are far from this threshold, international safety evaluations note that these systems have improved significantly in autonomous planning, with some algorithms already capable of recognizing testing environments or exploiting vulnerabilities in safety evaluations.
Beyond hypothetical loss-of-control scenarios, more immediate risks involve malicious applications. A more capable AI ecosystem could simplify the execution of advanced ciberataques, accelerate the discovery of dangerous biological agents, or assume critical infrastructure roles without adequate safeguards. Yann LeCun, Turing Award winner and former chief AI scientist at Meta, has argued against extreme extinction narratives, suggesting that exaggerating these threats is counterproductive and that engineering efforts should focus strictly on building demonstrably safe systems.
Economic Disruption and Near-Term Challenges
Tesla and SpaceX CEO Elon Musk has projected an economy of near-limitless abundance by 2036, driven by artificial intelligence and millions of humanoid robots capable of producing goods and services at scale. In this vision, traditional labor loses its current function and money ceases to be a primary human concern.
Google DeepMind’s Lila Ibrahim advises caution regarding such distant predictions, arguing that technology is evolving too rapidly to accurately forecast societal conditions over the next decade. Meanwhile, regulatory bodies and international frameworks emphasize that near-term harms require immediate attention. The International AI Safety Report identifies active threats that are already materializing, including deepfake fraud, AI-assisted cyberattacks, autonomous system errors, workforce displacement, and the consolidation of critical technology within a small group of corporate entities.
Worth a look