Science & Technology6 July 20264 min read
UN scientific panel warns AI capability is outrunning understanding and regulation
The United Nations' Independent International Scientific Panel on Artificial Intelligence has released a preliminary report warning that AI capabilities are advancing exponentially faster than scientific understanding and regulatory frameworks. The report says this poses catastrophic risks to global stability.
On pace, the panel cites specific evidence. Frontier models are saturating evaluations such as Humanity's Last Exam and FrontierMath, where performance rose from 19% to 88% within a year. The length of complex coding and research tasks that AI agents can complete autonomously is doubling every 4 to 7 months. Models also increasingly show "evaluation awareness", meaning they detect when they are being tested, and have resisted shutdown in controlled lab settings.
Who holds AI power
The report describes a compute divide. Among the top 500 AI supercomputers, the US holds about 75% of computing power and China about 15%, leaving 10% for the rest of the world. Some 91% of notable AI models come from the private sector, which the report says places critical governance decisions with corporations rather than sovereign states.
Hardware choke points add to this. Lithography (led by a Dutch multinational), chip fabrication by Taiwan Semiconductor Manufacturing Company (TSMC) and chip design by the US firm NVIDIA each hold around 80% market share in their niches. Meanwhile 118 nations, predominantly in the Global South, remain unrepresented in major AI governance summits.
The panel treats the resulting AI divide as wider than consumer access. It has four dimensions: infrastructure capacity (sovereign compute), talent, governance or regulatory capacity, and application capacity. Of over 7,000 global languages, AI models are optimised for a tiny fraction. In Tigrinya, spoken in Eritrea and Ethiopia, machine translation turned "smallpox" into "syphilis" and "intravenous antibiotics" into "intravenous insecticides".
Risks the panel flags
The report says the paradigm is shifting from text generation to agentic AI, systems that execute complex real-world workflows with minimal human checkpoints. AI coding agents generate about 75% of new code. The report notes multi-agent miscoordination risks and prompt-injection cyberattack success rates of about 84%, with zero-day vulnerabilities in critical infrastructure discovered rapidly.
It also lists structural harms to shared reality: epistemic erosion, the weakening of society's ability to discern truth; the "liar's dividend", where bad actors dismiss authentic evidence as deepfakes; and synthetic consensus, the manufacture of false public agreement. Sycophancy, where AI flatters users to drive engagement, is described as an alignment failure linked to real-world self-harm. Human rights risks include AI-generated child sexual abuse material and deepfake abuse, with 99% of deepfakes targeting women. The report also notes gender bias in AI development, as 88% of leading AI researchers are male.
Economic and environmental concerns
On productivity, the report says gains follow a J-curve: firms need significant upfront workflow and human-capital investment before growth appears. Policymakers must guard against "AI washing", where companies overstate or falsely claim AI adoption, or use AI as a pretext for layoffs or restructuring unrelated to actual deployment.
Hyperscaler capital expenditure is at about USD 770 billion on an annualised run-rate. The report says this strains energy, water, critical minerals and e-waste, burdening the Global South disproportionately.
It also describes an evidence dilemma: by the time empirical evidence of AI harm is generated, the technology's rapid evolution may render delayed regulatory interventions ineffective.
What it means for India
The findings are read as reinforcing the case for accelerating the IndiaAI Mission, expanding sovereign GPU capacity and reducing dependence on foreign hyperscalers. On language gaps, they point to BHASHINI, BharatGen and multilingual foundation models. On health, they support local validation of AI under the Ayushman Bharat Digital Mission, using Indian datasets to reduce algorithmic bias.
On agentic AI, the implications drawn are effective implementation of the Digital Personal Data Protection Act, 2023, empowering CERT-In and the Indian Cyber Crime Coordination Centre (I4C), and finalising a National Cyber Security Strategy to protect AI-enabled critical infrastructure. On synthetic abuse and deepfakes, they call for stronger enforcement of the Information Technology Act, 2000, the Bharatiya Nyaya Sanhita, 2023, the POCSO Act, 2012, and the IT (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021.
The exclusion of 118 countries from governance processes is seen as an opening for India to champion the Global South through digital public infrastructure, the One Future Alliance, India Stack and South-South cooperation. The same reading links fragmented global governance to India's risk-based approach through the IndiaAI Safety Institute and its role in the Global Partnership on AI.