AI is becoming more widespread, more accessible because it is cheaper and more efficient

"AI is a technology that is changing civilisation, not limited to one industry, but is transforming every industry it touches," said Russell Wald, executive director of the Stanford HAI and a member of the AI Index steering committee. "Last year we saw AI adoption accelerate at an unprecedented pace and its reach and impact will continue to grow." "The AI Index provides policymakers, researchers and the public with the data they need to make informed decisions and to ensure that AI is developed with human-centred values."

Significantly, for example, AI performance continues to improve. In 2023, researchers introduced new benchmarks: MMMU, GPQA and SWE-bench to test the limits of advanced AI systems. Just one year later, performance increased dramatically: scores increased by 18.8, 48.9 and 67.3 percentage points on MMMU, GPQA and SWE-bench, respectively. Beyond benchmarks, AI systems have made great strides in generating high-quality video and, in some situations, agical AI models have even outperformed humans. One study (reported by Stanford) caused a stir, according to which AI alone was able to make better diagnoses than doctors (even those assisted by AI itself).

However, reasoning remains a challenge for AI, despite the enthusiastic proclamations of the founders of these companies. Yes, learning-based systems that generate and test hypotheses using symbolic methods perform well, though not superhumanly well, in tasks such as problems in the International Mathematical Olympiad. Above all, Llm still lags behind complex reasoning benchmarks such as MMMU and struggles to reliably solve logically heavy tasks such as arithmetic and planning, even when the correct solutions are provable. This limits their use in high-risk contexts and where accuracy must be maximised.

The US is still ahead in terms of the amount of remarkable models released, but Chinese models are catching up. In January 2024, the best US model outperformed the best Chinese model by 9.26%; by February 2025, the gap had narrowed to just 1.70%. The report found similar results in other benchmarks related to reasoning, mathematics and coding.

Artificial intelligence is becoming more efficient, cheaper and more accessible. Thanks to increasingly capable small models, the cost of inference for a system with GPT-3.5 performance decreased by more than 280 times between November 2022 and October 2024. At the hardware level, costs have decreased by 30 per cent per year, while energy efficiency has improved by 40 per cent each year. Open-weight models are also closing the gap with closed models, reducing the performance difference from 8% to just 1.7% on some benchmarks in just one year. Together, these trends are rapidly lowering the barriers to advanced AI.

The other side of the coin is that energy consumption, although more efficient than output capacity, remains astronomical. And it still costs a lot to produce the most advanced models: the record is held by the Gemini 1.0 Ultra, $192 million (although these are only estimates, as companies do not publish these figures).

The US also dominates investment. By 2024, private investment in AI in the US reached $109.1 billion, almost 12 times China's $9.3 billion and 24 times the UK's $4.5 billion.

AI adoption in companies is also accelerating: 78 per cent of companies said they would use AI in 2024, up from 55 per cent the year before.

Certainly companies are investing in AI because they expect a big return on their investment. But it is fair to remember, as the report does, that companies have not yet seen a transformation that translates into significant savings or new profits. The report, citing a McKinsey survey, shows that of the companies that reported reducing costs, most achieved savings of less than 10 per cent. Of the companies that reported increased revenues due to AI, most reported earnings of less than 5%.

Who knows whether duties, by increasing the costs of developing and operating an AI model, will help to cool companies' enthusiasm.

The Responsible AI (RAI) ecosystem is evolving unevenly. AI-related incidents are on the rise, but standardised RAI assessments remain rare among leading industry model developers. New benchmarks such as HELM Safety, AIR-Bench and FACTS, however, offer promising tools for the assessment of incidents and safety. It is good that global cooperation on AI governance has intensified in 2024, with organisations such as the OECD, the EU, the UN and the African Union publishing frameworks focusing on transparency, accountability and other key AI principles.

Global optimism about AI is on the rise, but there are deep regional differences. In countries such as China (83%), Indonesia (80%) and Thailand (77%), a strong majority believe that AI products and services are more beneficial than harmful. In contrast, optimism remains much lower in places like Canada (40%), the United States (39%) and the Netherlands (36%).

AI and computer science education is growing, but gaps in access and preparation persist. Two-thirds of countries offer or plan to offer K-12 computer science education - twice as many as in 2019 - with Africa and Latin America making the most progress.