AI capabilities moving faster than ability to measure and govern them – UN Panel

According to the scientific panel constituted by the United Nations, AI has proven to be unique from other emerging technologies due to its unprecedented adoption globally and economies of scale that drive it.

Speaking at the launch of the Panel’s report, United Nations Secretary-General António Guterres warned that a lack of governing principles would water down the benefits of the technology.

“The more AI advances without shared rules, the less say governments and people will have in the outcome,” said Guterres.

Similar sentiments were shared by Maria Ressa, the Co-Chair of the Independent International Scientific Panel, who argued that while AI is transformative, failure to move towards measuring impact and harmonising governance structures would fail humanity.

“The risks — to societies, to security, and to our species — are too high, and the forces driving AI forward are not the forces that will deliver its benefits,” said Ressa.

Yoshua Bengio, also the Panel’s co-chair observed that AI capabilities are outpacing both scientific understanding and governments’ ability to adapt.

The safety puzzle

The scientific panel, commissioned in August 2025, established impending challenges in the measurement and evaluation of opportunities, risks and impacts of AI.

This, the report says, is due to information asymmetry in safety validation between companies and society. It cites the current practice, where AI companies retain proprietary visibility of the systems they have built.

“Safety evaluation methodologies are currently designed largely by the companies being evaluated. While there are some legally mandated disclosures, government experts primarily receive the testing data that developers choose to share,” the panel alludes to a lack of third-party assessments.

The panel also notes that evaluation data sets for frontier models are increasingly kept private as they score almost perfectly on an increasing number of standardized tests.

These models, the panel found, are capable of active deception during assessments as they have an evaluation awareness to understand when they are being tested.

“Assessment methodologies that are calibrated to an agent’s capacity for independent action, impact on its operational environment, and emergent behaviour are underdeveloped,” the Independent International Scientific Panel’s report highlights.

To avert the evaluation risks, the scientific panel advocates for dynamic execution-based tests, continuous assessments and interpretability methods to understand what is going on inside AI models.

Additionally, the panel also highlighted the inequities in AI development, which is currently concentrated in the USA and China, as the rest of the world only accounts for 10% of available models.