International AI Safety report
In January 2025, the first international AI Safety report was published. This document emerged from discussions at the 2023 Global AI Safety Summit. This report lays down the current AI advanced capabilities as well as the emerging risks.
Developed through an unprecedented international collaboration of over 100 AI experts, the report provides critical insights into the rapid advancements in AI capabilities, emerging threats, and the strategies required to manage AI risks effectively.
Contents
- Key Findings
- Capabilities of General Purpose AI
- Risks
- Technical Approaches to risk management
- Conclusion
Key Findings
Rapid Advancements in General-Purpose AI
General-purpose AI has seen significant improvements in recent years, with enhanced capabilities in scientific reasoning, programming, and multimodal applications. AI agents, which can autonomously act and make decisions with minimal human oversight, are gaining traction, posing both opportunities and challenges.Emerging and Established Risks
The report identifies several well-documented harms, including scams, biased outputs, privacy violations, and misinformation. However, new risks are also emerging, such as AI-enabled cyberattacks, potential biological and chemical threats, and concerns about societal control over increasingly autonomous AI systems. Some AI companies have already increased their assessment of biological risks from low to medium.The ‘Evidence Dilemma’ in AI Policymaking
The pace of AI development presents a significant challenge for policymakers, who must balance potential benefits with emerging risks. Due to the rapid evolution of AI, decision-makers often lack sufficient scientific evidence to predict risks accurately, complicating regulatory responses.Challenges in Risk Identification and Management
Risk assessment in AI remains difficult due to limited access to proprietary AI systems, technical opacity, and the diverse contexts in which AI operates. While interpretability techniques and adversarial training have improved, they remain insufficient for ensuring fully reliable AI behavior.
Capabilities of General Purpose AI
General Purpose AI (GPAI) refers to Artificial Intelligence models or systems that can perform a wide range of tasks rather than being specialised for one specific function. The versatility of this AI application holds capabilities which other AI applications simply cannot surpass. With these capabilities come inevitable challenges which outweigh all others. Listed below are the capabilties and challenges realting to GPAI.
Current Capabilities
Input and Output modalities. Modalities refer to the kinds of data that AI can receive as input and produce as output. General Purpose systems exist for 9+ modalities. Such as:
- Text and Code: Engage in interactive dialogue and write short computer programmes
- Audio and Speech: Engage in spoken conversation and emulate human voices
- Image: Describe the contents of images with high accuracy, generate images according to descriptions
- Video: Transcribe/Describe the contents of videos and generate short videos according to instructions
- Robotic Actions: Plan out robotic movements, but not yet control the
- Proteins and other molecules: Perform a range of tasks useful to biologists e.g. predicting protein folding’s.
Capabilties in the coming years
The pace of progress for GPAI has been on an exponential scale over recent years. The progress is tested with benchmarks relating to specific domains., such as maths, software and natural language processing. For example, a case study used the math benchmark alongside GPAI to test a wide range of problems. In 2021, this benchmark was released. General purpose AI systems initially scored around 5% but just 3 years later the model 01 reached 94.8% which matches the score of expert human testers. This case study can be used to highlight the pace of growth relating to AI models and is potential ability to surpass human intelligence across a wide range of areas. The following image, shows AI’s performance against human performance across a wide range of benchmarks.
As you can see, the capabilities of GPAI are not even. As such, the differing strengths and weaknesses of humans and GPAI models lead to a challenging comparison. With this, it means that while AI can surpass human performance on certain benchmarks but it it vital to remember that AI still lacks the deep conceptual understanding and abstract reasoning processes relating to humans. One common area in which GPAI highlights its weakness is in relation to reasoning. Humans are able to apply their reasoning to cope with novel scenarios
AI models generally memorise patterns rather than apply abstract reasoning and thinking and as such struggle to perform real world tasks. In these cases, human performance surpasses all AI capabilities.
From this, efforts are being deployed to overcome these weaknesses. These are discussed below.
To extend the model capabilities of GPAI, researchers have discovered ‘scaling laws’. These are mathematical relationships that quantify the relationship between inputs of the AI training process and the capabilities of the model on performance tasks. This has shown that the model performance improves with increased computational resources.
Algorithmic improvements allow GPAI to be trained with fewer resources. The techniques and training methods have been consistently improving over time. The computational efficiency of AI techniques for training have increased 10X every 2-5 years.
Risks
The following chapter dives into all of the upcoming and emerging risks associated with AI, specifically GPAI.
General Purpose AI has been seen to be exploited by malicious individuals who use AI to generate fake content in order to harm individuals. Despite there being mitigation strategies in place, these have serious limitations. There is also a lack of reliable evidence which supports these claims of exploitation which suggests development is needed in these areas.
Here are just an overview of the crimes committed through the exploitation of AI in todays society.
- Realistic AI generated content to demand and leverage money against others
- AI generated content to impersonate trusted authority figures in order to commit financial fraud.
- AI generated content to sabotage individuals in their in their personal/professional lives.
As of right now, there is no single robust solution to detecting and reducing the spread of AI-generated content. This highlights a significant limitation as AI is outpacing detection methods.
Loss of control scenarios are hypothetical future scenarios in which a GPAI system comes to operate outside of anyone’s control. Due to the mixed evidence surrounding this area, experts share disagreement on the likelihood of an active loss of control within the next 7 years. This suggests that there is most likely not a definite timeframe for preparation. However, it is plausible that as AI is gaining more and more attention and development the likelihood of an ‘active loss of control’ is growing. Especially with competitive pressures pushing unconsidered advancements ahead.
Risk Management Approach
The report outlines various technical methods aimed at mitigating AI risks, emphasizing that no single approach (especially in the context of GPAI) is fool-proof. Some of the key strategies include:
-
Risk Taxonomy and Assessment: Establishing systematic frameworks to categorize and evaluate AI risks based on severity and likelihood.
-
‘Defense in Depth’ Strategy: Layering multiple protective mechanisms to safeguard against AI failures, inspired by methodologies from nuclear safety and infectious disease control.
-
Participatory Risk Management: Engaging diverse stakeholders, including domain experts, regulators, and affected communities, to improve risk identification and mitigation.
-
Governance Mechanisms: AI companies are encouraged to adopt internal decision-making panels and independent advisory boards to oversee AI safety considerations.
-
AI Safety and Reliability Engineering: Drawing from established safety-critical industries like aerospace and nuclear energy, AI developers should integrate structured safety evaluations, traceability mechanisms, and robust auditing frameworks.
Policy Implications
Given the report’s findings, policymakers and regulatory bodies must consider the following actions:
Global AI Coordination: Governments should enhance cross-border collaboration to standardize AI safety protocols and share risk assessments.
Stronger Transparency Measures: AI companies should be required to disclose more information on their models’ capabilities, training data sources, and safety mechanisms.
Legally Enforceable Risk Thresholds: Defining clear regulatory thresholds for AI risks, similar to those in other high-risk industries, would ensure greater accountability.
Pre-emptive AI Risk Management: Regulatory frameworks should focus on proactive measures rather than reactive responses, considering the speed of AI advancements.
Conclusion
The International AI Safety Report 2025 underscores the urgent need for a multifaceted approach to AI governance, combining technical risk management with comprehensive policy interventions. As AI continues to evolve, international cooperation, transparency, and proactive safety measures will be crucial in ensuring AI’s benefits are maximized while its risks are minimized. The choices made today will shape the future trajectory of AI and its impact on global society.
For questions or further information, contact our team at sales@globalailaw.com.