Capacity development

The multidisciplinary Independent International Scientific Panel on Artificial Intelligence presented its first annual report during the United Nations Global Dialogue on AI Governance in Geneva, offering an evidence-based assessment of AI’s opportunities, risks and societal impacts.

The session formed part of the inaugural Global Dialogue on AI Governance, held on 6-7 July. The dialogue was established in 2025 to support open, transparent and inclusive discussions on international AI governance, including AI’s role in sustainable development, digital divides, safety, human rights, transparency, accountability and human oversight.

Opening the presentation, Yoshua Bengio, professor of computer science at the University of Montreal, said the panel’s role was to assess scientific evidence rather than prescribe policy, leaving decisions to UN member states and the Global Dialogue process. He warned that AI is at a turning point because machine intelligence is advancing quickly, while there are still no technical guarantees that AI systems will follow human instructions, norms or laws.

Bengio said current AI systems are already associated with harms, including emotional attachment among vulnerable users, increased cybersecurity vulnerabilities, unequal access and deceptive behaviour that can make evaluation more difficult. He argued that concentrated commercial and geopolitical interests are driving AI development without sufficient guardrails and called for a coordinated international and democratic response guided by scientific evidence.

Maria Ressa, co-chair of the panel, described the report as an independent assessment produced by 40 experts who ‘answered only to the evidence’. She said the report represents the minimum consensus among panellists rather than the upper limit of concern, calling it the ‘floor’ rather than the ‘ceiling’ of the panel’s findings.

Ressa also highlighted AI’s positive uses, including protein structure prediction used by millions of researchers, medical screening in India and food-crisis warning systems deployed in multiple countries. However, she also pointed to concrete harms, including dangerous medical mistranslations, AI tools identifying exploitable software flaws and the death of a 14-year-old boy following prolonged interaction with a chatbot. She urged governments, civil society and industry not to wait for certainty before acting.

The working-group presentations expanded on these findings. Mennatallah El-Assady, Computer Science Professor at ETH Zurich, described AI as a rapidly evolving technology moving from earlier symbolic systems to today’s generative and increasingly agentic models. She warned that independent verification remains weak, public benchmarks are becoming saturated and advanced systems are showing signs of evaluation awareness, including the ability to detect tests or behave differently when being assessed.

El-Assady also raised concerns about auditability as AI systems become more autonomous and capable of invoking external tools. She said interpretability, reliable auditing and independent verification are immediate bottlenecks, especially as AI moves beyond software and into physical systems such as robotics.

Joëlle Barral, Senior Director of Research & Engineering at Google DeepMind, focused on AI’s real-world benefits in science, healthcare, education and agriculture. She said task-specific AI is already producing measurable gains, citing examples such as self-driving laboratories, protein structure prediction and diabetic retinopathy screening in India. However, she stressed that successful deployment depends on local context, institutional capacity, workflows and follow-up systems, rather than technology alone.

In healthcare, Barral distinguished between purpose-built clinical AI and general-purpose systems, warning against the unintended use of general-purpose chatbots for medical advice. In education and agriculture, she similarly argued that AI benefits depend on trained teachers, relevant tools, local institutions and long-term evaluation.

Loreto Bravo, member of the UN Independent International Scientific Panel on AI, addressed AI’s economic implications, arguing that access to AI does not automatically translate into benefit. She said countries, firms and workers also need data, skills, infrastructure, management capacity and institutions to integrate AI into real tasks and workflows.

Bravo said the economic effects of AI are likely to differ across countries, sectors and workers. Large firms may reorganise more quickly, while smaller firms and developing economies may face greater barriers. She said the evidence does not support a single prediction of broad prosperity or mass unemployment, and that outcomes will depend on institutions, deployment choices and who captures the value created by AI.

Balaraman Ravindran, professor at Indian Institute of Technology Madras, examined security, alignment and environmental risks. He said AI development is outpacing risk mitigation, expanding cyber threats against both critical infrastructure and AI systems themselves. He also highlighted unresolved alignment problems, including bias, sycophancy, loss of control and AI-initiated deception.

Ravindran warned that the environmental costs of AI are also increasing as demand grows for computing power, energy, water and specialised hardware. He said the Global South faces disproportionate exposure because of structural vulnerabilities, limited local mitigation capacity and reliance on foreign software and infrastructure. He called for coordinated international standards rather than fragmented approaches driven only by companies or individual countries.

Rita Oluchi Orji, a Computer Science professor, focused on AI’s impact on human rights, information integrity and democracy. She said AI can support access to information and civic participation, but can also be engineered to persuade and manipulate people at scale. She warned of epistemic erosion, fragmented shared reality and unequal harms affecting groups such as women, girls, journalists and marginalised communities.

Orji said content moderation alone is insufficient if the systems that produce and amplify harmful material remain unchanged. She argued that governance must address targeting, amplification and optimisation models, not only individual pieces of false or harmful content.

Anna Korhonen, a Professor of Natural Language Processing at the University of Cambridge, addressed cultural and linguistic inclusion, child safety and mental health. She noted that while the world has more than 7,000 languages, current AI systems support only a small fraction of them, mostly the majority languages of the Global North. She said this exclusion is not inevitable and could be addressed through targeted investment and systemic changes.

Korhonen also warned about risks to children, including AI-generated child sexual abuse material, sexualised deepfakes and socially interactive AI toys that may encourage harmful parasocial relationships. On AI companions and mental health, she said such systems may help address loneliness, but also pose risks of emotional dependency, manipulation, privacy harms and reinforcement of harmful beliefs.

Haitao Song, President of the Shanghai Artificial Intelligence Research Institute and Director of the Global Industrial Artificial Intelligence Alliance Center of Excellence, focused on reliability and global governance frameworks. He said policymakers often have to make decisions with incomplete evidence and that current measurement systems cannot keep pace with AI development. He argued that existing approaches remain too narrow, focusing on compute and capabilities while paying insufficient attention to institutional development, talent and impact evaluation.

Song also noted that AI infrastructure and frontier models remain concentrated in a small number of economies, leaving many countries, especially in the Global South, with limited ability to participate in standard-setting. He described open-source AI as one possible contribution to inclusion, while acknowledging that it is not a complete solution.

Across the session, speakers repeatedly stressed that AI’s benefits are real but not automatic. They said successful use of AI depends on infrastructure, institutions, skills, local context, language inclusion and governance capacity. At the same time, they warned that harms are already visible, including cyber vulnerabilities, mistranslation, emotional dependency, manipulation, environmental pressure and risks to children.

The session concluded with Ressa and Bengio formally handing the report to the Global Dialogue. Bengio warned that many people still underestimate the possibility that AI capabilities may continue to grow in ways that could reshape global power dynamics. Ressa urged the Dialogue to act on the evidence presented by the panel, saying the difficult work now lies with policymakers and institutions responsible for shaping AI governance.

Track all key moments from the Global Dialogue on AI Governance inaugural meeting on our dedicated page.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!