In 2023, the British government commissioned the International Scientific Report on the Safety of Advanced AI. In January 2025, ahead of the AI Action Summit in Paris, the final report was published. Given the constant evolution of AI, the international team of researchers led by Yoshua Bengio has now published its first interim report.
The update summarises the latest developments in AI technology. It states that AI models are currently more powerful than ever. They can now solve high-level tasks. These advances are linked to the development of reasoning models. AI models structure their responses in intermediate steps. This demonstrates that factors such as training and additional computing power, as well as scaling the model size, contribute to making AI models more powerful. The report also discusses the current situation in the practical application of AI, the increased security measures employed by developers and the new challenges in the area of control. For instance, AI models are becoming increasingly adept at discerning whether they are being assigned a task or if the objective is merely to evaluate the model. The report analyses these developments and identifies new issues in the field of AI safety.
More details can be found in the first key update of the AI Safety Report 2025.