Stanford AI Index 2026: AI Advances Faster Than Regulations

A Revealing Report on the State of AI

In a context where artificial intelligence is omnipresent in technological discussions, the annual AI Index 2026 report from Stanford University is published today, offering valuable insights. This document attempts to untangle the relentless flow of information and speculation surrounding AI. While some predict a slowdown in AI development, the report shows that models continue to improve at a steady pace. Public adoption of AI surpasses that of personal computers and the Internet, and companies in the sector are generating revenue at an unprecedented rate, even as they invest heavily in costly infrastructures like data centers and chips.

This acceleration comes at a price. AI-dedicated data centers currently consume 29.6 gigawatts of energy, enough to power the state of New York during peak demand. Additionally, the annual water usage to operate OpenAI's GPT-4 model could exceed the drinking water needs of 12 million people. The chip supply chain, crucial for AI, remains extremely vulnerable, with a notable dependence on TSMC in Taiwan for the manufacturing of the most advanced chips.

United States and China: An Intense Technological Rivalry

In the global competition for AI supremacy, the United States and China are vying for the top spot. According to Arena, a platform that compares the performance of language models, the gap between the two countries is narrowing. In 2023, OpenAI dominated with ChatGPT, but this advantage diminished in 2024 with the arrival of new models from Google and Anthropic. In February 2025, the R1 model from DeepSeek, a Chinese lab, temporarily matched ChatGPT. By March 2026, Anthropic took the lead, closely followed by Google and OpenAI, while Chinese models like those from DeepSeek and Alibaba are not far behind.

The United States stands out for the power of its AI models, its capital, and its 5,427 data centers, housing most of the world's AI data centers. In contrast, China excels in research publications, patents, and robotics. This intense competition is pushing companies like OpenAI and Google to be less transparent about their training processes, complicating the study of model safety by independent researchers.

Continuous Improvement of AI Models

Despite predictions of stabilization, AI models continue to progress at an impressive pace. They are now achieving performance levels comparable to, or even exceeding, those of human experts in fields such as science, mathematics, and languages at the doctoral level. The SWE-bench Verified benchmark has seen the scores of the top models rise from 60% in 2024 to nearly 100% in 2025. Furthermore, an AI system has successfully produced weather forecasts autonomously.

However, AI is not without its weaknesses. Models, which primarily learn from textual and visual data, exhibit uneven intelligence. Robots, for example, succeed in only 12% of household tasks. Autonomous vehicles are more advanced, with Waymos operating in several U.S. cities. AI is also expanding into professional sectors like law and finance, although no model currently dominates these industries.

Inadequate Testing Methods

The advancements in AI must be interpreted with caution, as current benchmarks struggle to keep pace with technological progress. Some benchmarks are poorly designed, such as one that evaluates models' mathematical capabilities with an error rate of 42%. Others can be circumvented: models trained on specific test data can achieve good results without becoming more intelligent.

The actual use of AI often differs from its testing, and high performance on benchmarks does not guarantee effectiveness in the real world. For complex technologies like AI agents and robots, benchmarks are still rare. Additionally, AI companies are sharing less and less information about their training methods, and independent tests sometimes reveal results that differ from those announced. "The lack of information about your model's performance on a benchmark can say a lot," emphasizes Yolanda Gil, co-author of the report.

Impact of AI on Employment

The rapid adoption of AI is having repercussions on the job market. In less than three years, more than half of the global population is using AI, a rate of adoption faster than that of personal computers or the Internet. Approximately 88% of organizations have integrated AI, and four out of five university students are using it.

Although it is still early to fully assess the impact of AI on employment, some studies already indicate notable effects. A 2025 study from Stanford shows that employment among software developers aged 22 to 25 has dropped by nearly 20% since 2022. While other economic factors may also play a role, AI appears to be contributing to this trend.

Employers anticipate a tightening of hiring, with one-third of organizations expecting a reduction in their workforce due to AI, particularly in service operations, supply chain, and software engineering. AI has increased productivity by 14% in customer service and 26% in software development, but these gains are not reflected in tasks requiring more judgment. The overall economic impact of AI remains to be determined.

Public Perceptions and Experts on AI

Opinions on AI are divided around the world. According to an Ipsos survey cited in the index, 59% of people believe that AI will bring more benefits than problems, but 52% express nervousness about it. A Pew survey reveals a significant gap between experts and the public regarding the future of AI. While 73% of experts believe that AI will have a positive impact on work, only 23% of the American public shares this view. Experts are also more optimistic about the impact of AI on education and health, although they agree with the public on its potential negative effects on elections and personal relationships.

In the United States, trust in the government to regulate AI is low. An Ipsos survey shows that Americans are more concerned about insufficient regulation than about excessive regulation.

The Challenges of Regulating AI

Governments around the world are struggling to regulate AI, although some progress has been made. The initial prohibitions of the EU AI Act, which ban the use of AI in predictive policing and emotion recognition, have come into effect. Japan, South Korea, and Italy have also enacted national AI laws. In the United States, despite a federal trend toward deregulation, state legislatures have passed a record 150 AI-related bills.

California has enacted significant laws, such as SB 53, which imposes security disclosures and protections for whistleblowers. New York has adopted the RAISE Act, requiring AI companies to publish security protocols and report critical incidents. However, Yolanda Gil points out that regulation is lagging behind technology, as our understanding of AI systems is still limited.