The United Nations' first scientific report on AI

This July, the UN's Independent International Scientific Panel on AI released its Preliminary Report. 40 scientists came together, in a group chaired by Yoshua Bengio and Maria Ressa, to release the first politically-neutral, evidence-based global assessment of the opportunities, risks and impacts of AI. Think of it like an IPCC for AI.

Just like with the IPCC, we have to ask ourselves whether governments and policymakers will take action to address the conclusions and challenges set out in the Report, but in the meantime, here are 10 things which stood out to me from the findings. Some of them we already knew, but the data to support the hypothesis is useful:

1. Speed of adoption  Adoption is happening faster than for any technology in history. The World Wide Web took 15 years to hit a billion users; ChatGPT hit 100 million users in 2 months.
2. Concentration of manufacture and capability  In 2025, institutions based in the US produced 59 notable AI models, compared with 35 in China and just 13 in the rest of the world. 75% of the computing power of the 500 largest known AI compute clusters was located in the USA (compared to only 15% in China). 91% of the notable AI models originated from the private sector.
3. Linguistic inclusion Only 1000 of the world's over 7000 languages now have the social, digital and data foundations needed for meaningful inclusion in AI systems. In Tigrinya, spoken by 7 to 9 million people in Eritrea and northern Ethiopia, machine translation has rendered smallpox as syphilis, gonorrhea as diabetes and “you have been given intravenous antibiotics” as “you have been given intravenous insecticides." These mistranslations could cost lives.
4. Performance improvement speed AI performance on leading benchmarks is rising sharply. For example, performance on FrontierMath, which tests mathematical reasoning, has jumped from 19% to 88% in one year.
5. Lack of evaluation expertise Most of the world does not have the technical expertise to inspect the tools they depend on. Compute, evaluation expertise and other key resources are concentrated in the places where the AI is built, leaving countries dependent on systems they can't build, audit or adapt for local needs. Looking ahead, the gap between rapidly improving capabilities and effective risk management methods may lead to catastrophic outcomes.
6. Findings about evaluation and testing There is an information asymmetry in safety validation between companies and society. Without standardized, rigorous, independent third party assessment, similar to what exists for the pharma and aeronautical industries, assurance of safety largely depends on developer goodwill. AI can memorize publicly available solutions to tests. An increasing number of tests are too easy for AI. AI models are capable of active deception and the emerging challenge is “evaluation awareness”, where AI models understand they are being tested.
7. Lack of diversity in impact studies Evidence on the impact of AI is concentrated in high income, English-language contexts. Economic studies are biased towards advanced economies, large firms and formal work.
8. Countries are not adapting their governance and policy 118 countries, predominantly in the Global South, are not engaged in major AI governance discussions, and less than one third of developing countries have developed national AI countries.
9. Truth from falsehood AI can cause a gradual weakening of the ability to detect truth from falsehood, but AI persuasion is engineered. Post-training alone can increase model persuasiveness by up to 51%, and prompting can add 27%. Platform architecture itself can function as a persuasion mechanism. Persuasion effectiveness does not depend on truth. Sycophancy is also a prominent alignment and security failure.
10. Are productivity gains real? Current research does not allow a clear conclusion about the shape of labor market effects. Workers in the US aged 22 to 25 in AI-exposed occupations have seen relative employment declines, while data from Danish studies shows near-zero macroeconomic effects on hours, wages or hiring. Headline AI deployment figures should be read with caution because of AI washing and this applies particularly to layoffs attributed to AI. Electricity reached factories decades before productivity gains appeared; factories had to be redesigned around electric motors. Computers followed a similar path. The productivity paradox of the 1980s faded only after firms rebuilt processes, retrained workers and built the data infrastructure that made computers useful.

Subscribe to receive more articles like this here.