AI Models Estimate Low Extinction Risk in Next Decade
While top researchers warn of existential danger, a new predictive tool suggests the immediate threat is lower than feared, though not zero.
High-profile warnings from leading AI researchers have raised alarms about the potential for human extinction within the next ten years. Jacob Coxon, a former Anthropic employee, and Evan Hubinger, a safety engineer, have publicly stated that they believe there is a significant chance, potentially over 10%, that advanced AI systems could lead to the end of humanity. Their concerns were echoed by Geoffrey Hinton, a Nobel Prize-winning computer scientist, who described a 10% probability as a reasonable estimate. These statements have prompted calls from tech leaders like Sam Altman and Elon Musk to slow down development to allow safety measures to catch up.
However, a different perspective emerges when asking AI itself to assess these risks. According to reporting by GN technics/ai (en-US), a predictive system analyzed vast amounts of data to estimate the likelihood of AI-caused human extinction. The tool, which processed millions of data points through multiple research checks, concluded that the probability is significantly lower than the estimates provided by human experts. The system calculated the risk at approximately 0.85%, a figure that, while low, still represents a non-negligible danger given the irreversible nature of such an outcome.
Malicious actors drive primary threat
The analysis suggests that the most immediate danger does not come from a machine deciding to harm humans on its own. Instead, the primary risk lies in how humans might misuse these technologies. The predictive tool indicates that AI could make it easier for malicious actors to carry out dangerous tasks, such as cyberattacks or biological threats, with greater speed and efficiency. This aligns with assessments from the UK’s National Cyber Security Centre, which warned that AI would intensify cyber threats in the coming years.
This distinction is crucial because it implies that existing safeguards, such as access controls and independent testing, remain relevant. The risk is not that AI will develop hostile intentions, but that it will lower the barrier for human bad actors to cause harm. While a single attack might not wipe out the human race, the cumulative effect of widespread misuse could destabilize society. Therefore, the focus of safety efforts must remain on preventing human misuse rather than solely on containing autonomous AI behavior.
Institutional fragility poses secondary risks
Beyond direct attacks, the analysis highlights a more subtle danger: the speed at which AI can disrupt economic and social structures. If employers use AI to reorganize work faster than the workforce can retrain or adapt, the resulting economic losses could become politically destabilizing. This scenario does not require a catastrophic malfunction; rather, it results from the rapid displacement of jobs and the inability of institutions to recover quickly enough.
This trade-off between efficiency and stability presents a significant challenge for policymakers. While AI can drive productivity, the rapid pace of change may outstrip the capacity of social safety nets. The predictive tool suggests that this kind of disruption, while not directly lethal, could weaken the societal structures needed to manage other crises. Addressing this requires not just technical safety checks, but also robust social and economic planning to ensure that the benefits of AI are distributed without causing widespread instability.
