In November 2023, the United Kingdom hosted the first global AI Safety Summit, which convened a diverse coalition of world leaders, technologists, academics, and civil society groups to accelerate international action on the safe and responsible development of frontier AI. In addition to producing the Bletchley Declaration on AI safety, the Summit also commissioned a “State of the Science” report to help create international consensus on the present and potential future risks, opportunities, and capabilities of general-purpose AI. The interim version of this international collaboration was unveiled at the AI Seoul Summit in May 2024.
I contributed an analysis on current issues regarding the underrepresentation of non-Western languages and cultures in AI systems, which contributes to the growing AI divide and analyzed various aspects of bias in AI. Such contributions fit within the broader discussion on AI capabilities—assessing and understanding general-purpose AI, risks from such systems, and methods for mitigating such risks—and must be highly considered as global debates continue.
This report underscores the necessity for broader global cooperation in defining AI risks and developing robust mitigation solutions. As the U.K. government prepares for the final report to be published ahead of the AI Action Summit in France in February 2025, there is the need for a concerted effort to improve international cooperation in AI, particularly by including governments, researchers, and civil society advocates from the Global South. Equitable inclusion of diverse perspectives will be crucial to contend with the risks and harms from AI and democratize the benefits of these systems.
Underrepresentation and AI bias
Harmful bias and underrepresentation in AI systems have been challenges since well before the current AI boom. These problems have persisted with general-purpose AI systems and will likely remain for the foreseeable future. AI outputs are biased if they are skewed based on protected characteristics, such as gender, race, or religion. When these outputs inform high-stakes decisions, they create unfair outcomes that lead to tangible harm for members of marginalized groups in areas from job recruitment to financial lending to health care. A prominent source of AI bias extends from underrepresentation in training data, where data about certain groups are limited or non-existent within datasets, leading to lower-quality model predictions, or reflecting pre-existing human biases. Most of the language datasets used to train general-purpose large language models are in English and primarily represent Western cultures, which have been proven to lower the performance on prompts referencing non-Western cultures and societies. Furthermore, AI systems may exhibit intersectional bias, where bias compounds because an individual has multiple marginalized characteristics (e.g., a low-income woman of color). Despite extensive research, reliable methods to fully mitigate such discrimination remain elusive.
Global AI divide
Not only can AI create unequal outcomes for individuals and demographic groups, but the technology can also exacerbate inequalities between nations. Frontier AI research and development primarily occurs in Western countries, such as the U.S., EU, and U.K., and China, which means that other countries, particularly those in the Global South, will have less agency over and reap fewer benefits from continued AI advancements. This “AI divide” occurs largely along the fault lines of existing global socioeconomic disparities and, without deliberate efforts to share the benefits of AI, stands to compound them. Substantial barriers that prevent these countries from benefiting from general-purpose AI include lower digital skills literacy, limited access to vital computing resources, infrastructure challenges, and economic dependence on entities in higher-income countries. Because general-purpose AI development is dominated by a few companies, particularly those based in the U.S., the systems used worldwide will likely reflect the values, cultures, and goals of Western countries and their large corporations. In addition, the recent trend towards developing ever-larger, more powerful general-purpose AI models could also exacerbate global supply chain inequalities, place demands on energy usage, and lead to harmful climate effects, which also worsen global inequalities.
The State of the Science of advanced AI
Broadly, the report concludes that, according to many metrics, the capabilities of general-purpose AI are advancing rapidly. Current undebated capabilities include helping humans write code, maintaining coherent conversations, and solving textbook math and science questions. Whether there has been significant progress on fundamental challenges such as causal reasoning is debated among researchers. Experts disagree on the expected pace of future progress of general-purpose AI capabilities, variously supporting the possibility of slow, rapid, or extremely rapid progress. Overall, there is limited understanding of the capabilities and inner workings of general-purpose AI systems, especially as they are suited to generate other external harms.
For example, malicious actors can use AI for large-scale disinformation and influence operations, fraud, and scams. Malfunctioning general-purpose AI can also cause harm, for instance, through biased decisions with respect to protected characteristics like race, gender, culture, age, and disability. Future advances in general-purpose AI could pose systemic risks, including labor market disruption and economic power inequalities. Experts have different views on the risk of humanity losing control over AI in a way that could result in catastrophic outcomes. Several technical methods (including benchmarking, red-teaming, and auditing training data) can help to mitigate risks, though all current methods have limitations, and improvements are required.
The final version of the report, expected to provide a comprehensive update on the rapidly evolving science of AI safety, is scheduled for release at France’s AI Action Summit in February 2025. This updated consensus of this continually evolving field will powerfully shape global AI governance strategies in the coming year.
Click here to access the full report for more detailed insights and recommendations from experts around the world.