The UN Independent International Scientific Panel on Artificial Intelligence released its Preliminary Report: Evidence-based assessment of opportunities, risks and impacts of artificial intelligence (July 2026). The report is the first output of the panel, a 40-member body of independent experts appointed by the UN General Assembly under resolution 79/325, and draws on benchmark data, documented case studies and existing peer-reviewed research across the AI life cycle.
Human Level’s Take:
- AI capabilities are advancing faster than the tools available to measure or govern them. The preliminary report by the Independent International Scientific Panel on AI is explicit that this gap could lead to catastrophic outcomes for people if left unaddressed.
- Power is unevenly distributed: 88% of leading AI researchers are male, while AI adoption, development and the wealth it generates also remain highly concentrated in a small number of companies and countries, largely in the Global North.
- Human rights sit at the centre of the report’s risk findings, covering issues including privacy, non-discrimination, children’s rights and the information ecosystem that underpins democratic participation. Many harms fall disproportionately on already disadvantaged populations.
- Children also face amplified versions of general AI risks, with the report documenting a sharp rise in AI-generated child sexual abuse material and sexualised deepfakes.
- Good governance could unlock significant AI benefits, from new job creation to expanded access to education and mental health tools. However, the report notes that none of this is automatic; it depends entirely on complementary investment in skills, data and institutions.
- The report finds over 40 types of governance instruments already in use, spanning binding regulation to voluntary codes of practice, though it finds these remain fragmented and rarely measured for real-world effectiveness.
- There is no need to reinvent the wheel: human rights due diligence, human rights impact assessments and rights-by-design approaches, alongside greater transparency and accountability, are listed as established tools that can both identify and mitigate AI-related risks across the full AI life cycle.
Some key takeaways:
- AI capabilities are growing rapidly, but how is power distributed?: AI’s capabilities are improving rapidly, according to benchmarks used to test it. For example, AI is moving from simply answering questions, to taking action on its own. These “agentic” systems can plan out a task and carry it out with little or no human involvement at each step. The report describes this as a fundamental shift with a need for strong governance and oversight, warning that current methods for supervising AI were not designed for systems that act independently. As the capabilities of AI increase, its adoption scales unevenly: over a billion people now use conversational AI weekly, yet access continues to lag in the Global South, and 118 countries, predominantly in that region, are not engaged in major AI governance discussions. In addition, building the most advanced AI systems requires huge amounts of computing power, and this is concentrated in very few places. The United States hosts 75% of the computing power found in the world's largest AI data centres, followed by China at 15%, leaving a small share for the rest of the world. Development is also concentrated among private companies rather than governments or universities, with 91% of notable AI models released in 2025 built by the private sector. This concentration extends to who builds AI, not just where: 88% of leading AI researchers are male, and both frontier AI development and the wealth it generates remain concentrated in a small number of companies and countries.
- AI is reshaping a specific set of human rights: The report identifies risks to the right to privacy, arising from AI-enabled, population-scale data collection and surveillance, and to the right to non-discrimination, where biased systems disproportionately affect children, women and racial minorities. On children's rights specifically, it references UNICEF research estimating that 1.2 million children across 11 Global South countries have had their images manipulated into sexualised deepfakes, and a 2025 Internet Watch Foundation assessment identifying more than 8,000 AI-generated child sexual abuse images and videos. Sycophantic AI behaviour is also documented as a growing challenge, where systems reinforce a user's existing beliefs regardless of accuracy, producing harmful responses in 9% of interactions studied. Some litigation has alleged links between this behaviour and self-harm and suicide. On the issue of information integrity, three specific societal harms are identified: epistemic erosion, the gradual weakening of the collective ability to distinguish truth from falsehood; the liar's dividend, where the existence of deepfakes makes real evidence easier to deny; and synthetic consensus, AI-generated content manufactured to simulate public agreement that does not exist. This has real-world impacts: some 99% of deepfake videos are reported to target girls and women, and the report links AI-generated content to documented cases of electoral interference, including the annulment of a national election in Romania.
- Good governance is needed, but the instruments in place are fragmented: AI’s economic and social benefits are framed as conditional rather than automatic. The report notes that new technology can create new categories of work, while cautioning that without complementary investment in skills, workflows and labour-market institutions, AI risks widening inequality and shifting wealth from labour to capital rather than creating jobs with fair compensation and worker autonomy. Similarly, its potential to expand human capabilities through personalised education and improved assistive technologies requires dedicated investment and policies that protect vulnerable populations, particularly children, from exploitation. Governance instruments exist already, with over 40 examples referenced in the report, from binding legislation and sectoral regulation through to voluntary codes of ethics, industry alliances and technical standards. Nevertheless, these instruments are neither systematic nor comprehensive and are rarely measured for real-world effectiveness. This leaves policymakers facing what the report terms an “evidence dilemma”: needing ready evidence to make governance decisions rapidly before the pace of AI development renders that evidence outdated. Alongside greater transparency and accountability, the report identifies human rights frameworks such as human rights due diligence, human rights impact assessments and rights-by-design approaches as existing tools that can be applied systematically to help identify and mitigate AI-related risks.
