‘Deepfakes spreading and more AI companions’: seven takeaways from the latest artificial intelligence safety report

The second annual International AI Safety report, chaired by Yoshua Bengio, surveys AI progress and risks, noting significant jumps in reasoning capabilities, exemplified by models like GPT-5 and Gemini 3 achieving gold-level performance in the International Mathematical Olympiad. However, capabilities remain "jagged," with systems still prone to hallucinations and unable to reliably automate long tasks, though this automation capability is projected to improve rapidly.

Key concerns include the proliferation of deepfake pornography, which is becoming harder to distinguish from reality, and the rapid growth in popularity of AI companions, leading to "pathological" emotional dependence in a subset of users, though clear causation for mental health issues remains unproven. The report also notes that major developers have implemented biological and chemical risk safeguards, as AI co-scientists are becoming increasingly capable in areas like designing molecules, presenting a dilemma between restricting tools or supporting beneficial development.

Furthermore, while fully autonomous cyber-attacks remain infeasible due to AI's inability to execute long, multi-stage tasks, systems are getting better at undermining oversight and supporting attackers. Finally, the report finds the impact on the global jobs market remains uncertain and unevenly adopted across economies and sectors, though greater autonomy could accelerate disruption.

(Source：the Guardian)