Artificial Intelligence (AI) is on track to become the defining technology of the 21st century, according to the latest Stanford HAI AI Index, but its gains will be uneven without deliberate oversight.
The data, it cautions, reveals that progress is accelerating faster than institutions can keep pace.Technically, leading AI models are converging, with open-weight systems narrowing the gap.
But the means of measuring that progress are fraying: benchmarks are saturating, transparency is declining, and independent tests do not always align with developer claims, the report notes.
In science, AI is shifting from assisting discrete tasks to replacing entire workflows.
In medicine, tools such as ambient scribes are scaling across health systems.Regulation, meanwhile, is sending mixed signals.
The European Union is enforcing its AI Act, the US is leaning towards deregulation, while Japan, South Korea and Italy are advancing national frameworks.
Many developing economies, including India, are entering the arena, with AI sovereignty emerging as a unifying theme.
Optimism is rising, but so is unease.Here are some highlights from the study:➡️ AI isn’t plateauing.
Instead, it’s accelerating.➡️ Over 90% of frontier models now come from industry, many matching or beating humans on advanced tasks.
Coding benchmarks neared human parity in a year, while adoption hit 88%, with most students using generative AI.➡️ The US-China AI gap has narrowed to near parity.
China leads in AI publications, citations, and patents, while the US dominates in high-impact patents and output (50 notable models in 2025 vs China’s 30).
South Korea tops patents per capita, and China’s share of the 100 most-cited AI papers rose from 33 in 2021 to 41 in 2024.
In 2024, China accounted for 17.8% of AI publications in 2024, compared to 11.1% from Europe and 7.6% from India.➡️ The US leads in AI data centres (5,427)—10 times more than any other country—but the chip supply chain is fragile.
Most advanced chips rely on Taiwan’s TSMC, creating a critical single-point dependency despite US expansion.➡️ AI shows gold-medal math performance, yet they are unable to read time from traditional clocks (with about 50% accuracy).
AI Agents improved to about 66% task success but still failed one in three tasks.➡️ AI models can outperform scientists in some domains, but scaling isn’t linear.
They beat chemists on average, yet perform poorly in areas like astrophysics.➡️ Robots thrive in labs but struggle at home—only 12% success rate.
Simulations hit about 89%, highlighting the gap between controlled environments and real-world complexity.➡️ The US leads AI investment ($285.9 billion) and startups, but talent inflows are falling sharply—down 89% since 2017, including an 80% drop in the last one year.➡️ India leads in AI skill penetration at 3.0 (nearly triple the global average), followed by the US (2.0) and Germany (1.8), LinkedIn data shows.
But gender gaps persist.
In India, men list AI skills far more than women (3.1 vs 1.9), with a similar, slightly narrower gap in the US (2.1 vs 1.4).➡️ AI’s environmental cost is rising fast.
AI data centre capacity reached 29.6 GW, on a par with New York’s peak demand, while GPT-4o’s annual water use may exceed the drinking needs of 12 million people.➡️ Smaller, specialised models often outperform much larger ones.
Scientific AI is also more collaborative, unlike industry-led general AI.➡️ Open-source AI continues to scale, with 5.6 million projects on GitHub and Hugging Face uploads tripling since 2023.
India remains a growing contributor, representing 5.2% of all projects.➡️ AI is reshaping clinical care, cutting documentation time and burnout, but strong evidence is limited.
Nearly half of studies rely on exam-style data, and only about 5% use real patient data.➡️ Education is lagging, even as AI use surges.
More than 80% students use AI, but policies are unclear and uneven.
Meanwhile, AI skills are growing globally, and more PhDs are staying in academia.➡️ AI sovereignty is rising, with nations investing in domestic capability.
Still, model development is concentrated in the US and China.
Open source is widening participation and enabling more diverse models.➡️ Regional efforts are building language-first AI from scratch, not waiting on global labs.
Projects like SEA-LION and AI4Bharat are creating local data pipelines, tokenisers, and benchmarks.
These initiatives make linguistic inclusion a core design goal, expanding responsible AI beyond major hubs.➡️ India saw the sharpest rise in AI nervousness of any country surveyed.➡️ Responsible AI is lagging capability.
Safety reporting remains patchy: incidents rose to 362 (from 233 in 2024), and improving one dimension (for example, safety) can worsen another, like accuracy.➡️ Experts and the public diverge sharply on AI’s impact—73% vs 23% positive views on jobs.
Trust in AI governance is fragmented, with the US among the least confident in its own regulation.AI TOOL OF THE WEEKBy AI&Beyond, with Jaspreet Bindra and Anuj Magazine𝚃𝚑𝚎 𝙰𝙸 𝚏𝚎𝚊𝚝𝚞𝚛𝚎/𝚝𝚘𝚘𝚕 𝚠𝚎 𝚙𝚛𝚎𝚜𝚎𝚗𝚝 𝚝𝚘𝚍𝚊𝚢: 𝙲𝚕𝚊𝚞𝚍𝚎 𝚏𝚘𝚛 𝚆𝚘𝚛𝚍𝚆𝚑𝚊𝚝 𝚙𝚛𝚘𝚋𝚕𝚎𝚖 𝚍𝚘𝚎𝚜 𝚒𝚝 𝚜𝚘𝚕𝚟𝚎? 𝙰 𝚕𝚊𝚠𝚢𝚎𝚛 𝚛𝚎𝚌𝚎𝚒𝚟𝚎𝚜 𝚊 𝚑𝚎𝚊𝚟𝚒𝚕𝚢 𝚛𝚎𝚍𝚕𝚒𝚗𝚎𝚍 𝚌𝚘𝚗𝚝𝚛𝚊𝚌𝚝 𝚊𝚝 5:00 𝚙𝚖, 𝚊𝚗𝚍 𝚝𝚑𝚎 𝚙𝚊𝚛𝚝𝚗𝚎𝚛 𝚠𝚊𝚗𝚝𝚜 𝚊 𝚌𝚕𝚎𝚊𝚛 𝚜𝚞𝚖𝚖𝚊𝚛𝚢 𝚘𝚏 𝚌𝚑𝚊𝚗𝚐𝚎𝚜, 𝚔𝚎𝚢 𝚛𝚒𝚜𝚔𝚜 𝚊𝚗𝚍 𝚙𝚛𝚘𝚙𝚘𝚜𝚎𝚍 𝚌𝚘𝚞𝚗𝚝𝚎𝚛-𝚎𝚍𝚒𝚝𝚜 𝚋𝚢 9:00 𝚊𝚖.𝚃𝚑𝚒𝚜 𝚒𝚜 𝚗𝚘𝚝 𝚊 𝚏𝚛𝚒𝚗𝚐𝚎....
