AI Safety Benchmarks Lagging

Contrary to widespread assumptions, the notion of a sustained US leadership in artificial intelligence model performance is increasingly challenged by recent data. This is just one of the noteworthy, and perhaps uncomfortable, findings emerging from Stanford University’s latest AI Index Report, published this week.

The comprehensive 423-page annual assessment, curated by Stanford’s Institute for Human-Centred Artificial Intelligence, offers a deep dive into the current landscape of AI. It meticulously examines research output, model capabilities, investment trends, public perception, and the critical domain of responsible AI development. While many headlines have focused on general progress, the report’s most consequential insights lie in areas often overlooked by broader coverage, particularly concerning AI safety and the widening chasm between AI’s expanding functionalities and the rigorous evaluation of potential harms.

Three key findings from the report warrant significantly more attention than they are currently receiving.

The US-China AI Model Performance Gap Narrows Dramatically

The prevailing narrative of a definitive US lead in AI development requires urgent re-evaluation. Stanford’s report indicates that the top performance positions between US and Chinese AI models have been in flux since early 2025. In February of that year, DeepSeek-R1 demonstrated performance levels on par with leading US models. As of March 2026, Anthropic’s most advanced model held a marginal lead of just 2.7%. While the US continues to produce a higher volume of top-tier AI models – 50 in 2025 compared to China’s 30 – and maintains an edge in high-impact patents, China has surpassed the US in publication volume, citation share, and patent grants. China’s contribution to the top 100 most-cited AI papers surged from 33% in 2021 to 41% in 2024. Notably, South Korea leads globally in AI patents per capita, underscoring a broader international push.

The practical implication of these shifts is clear: the assumption of an insurmountable US technological advantage in AI model performance is no longer data-supported. The performance gap that was evident just two years ago has diminished to a margin that fluctuates with each significant model release. Furthermore, the report highlights a critical structural vulnerability within the US AI ecosystem. Despite hosting over 5,427 data centers, a figure more than ten times that of any other nation, the production of virtually all leading AI chips relies on a single entity: TSMC. This concentration of fabrication capabilities in Taiwan, although beginning to diversify with TSMC’s US expansion operational since 2025, represents a significant bottleneck for the entire global AI hardware supply chain.

AI Safety Benchmarking Lags Far Behind, Indicating Growing Risks

While virtually all frontier AI model developers readily report performance metrics on capability benchmarks, the same cannot be said for responsible AI benchmarks. The 2026 AI Index meticulously documents this disparity. The report’s benchmark analysis for safety and responsible AI reveals a stark lack of comprehensive data, with most entries for crucial metrics being conspicuously empty. Only Claude Opus 4.5 reports results across more than two tracked responsible AI benchmarks, and only GPT-5.2 has reported “StrongREJECT” performance. Across key areas such as fairness, security, and human agency, the majority of leading models offer no reported data.

This absence of reported data does not imply a lack of internal safety work by development labs. The report acknowledges that internal red-teaming and alignment testing occur, but emphasizes that “these efforts are rarely disclosed using a common, externally comparable set of benchmarks.” This lack of standardized reporting effectively renders external comparison of AI safety dimensions nearly impossible for most models. The consequences are tangible: documented AI incidents rose to 362 in 2025, a significant increase from 233 in 2024, according to the AI Incident Database. The OECD’s AI Incidents and Hazards Monitor, utilizing a broader automated detection pipeline, recorded a peak of 435 monthly incidents in January 2026, with a six-month moving average of 326 incidents.

Organizational-level governance responses appear to be struggling to keep pace. A joint survey by the AI Index and McKinsey reveals a decline in organizations rating their AI incident response as “excellent,” dropping from 28% in 2024 to 18% in 2025. Those reporting “good” responses also saw a decrease, from 39% to 24%. Concurrently, the proportion of organizations experiencing three to five AI incidents surged from 30% to 50%. The report further identifies a fundamental challenge in the very process of improving responsible AI: advancements in one dimension often degrade performance in another. For instance, enhancing safety might compromise accuracy, or improving privacy could reduce fairness. The absence of a standardized framework for managing these inherent trade-offs, coupled with a lack of necessary standardized data for tracking progress in areas like fairness and explainability, presents a significant hurdle for future development.

Public Anxiety Mounts with AI Adoption, Highlighting an Expert-Public Divide

Globally, a growing majority—59% of surveyed individuals—believe AI’s benefits outweigh its drawbacks, an increase from 55% in 2024. Simultaneously, 52% express nervousness about AI products and services, a two-percentage-point rise in a single year. The simultaneous upward trend in both metrics suggests a public increasingly engaged with AI while simultaneously experiencing heightened uncertainty about its trajectory. This divergence is particularly pronounced when examining AI’s perceived impact on employment. The report indicates that 73% of AI experts anticipate a positive impact on job roles, contrasting sharply with only 23% of the general public—a substantial 50-point gap. On the broader economic front, the disparity is 48 points, with 69% of experts positive compared to 21% of the public. In healthcare, expert optimism remains significantly higher at 84%, versus 44% of the public.

These perceptual gaps are critical, as public trust directly influences regulatory frameworks, which in turn shape AI deployment. In this context, the report flags a striking statistic: the United States reports the lowest level of trust in its own government to regulate AI responsibly among all surveyed nations, at a mere 31%. This stands in stark contrast to the global average of 54%. Southeast Asian countries exhibit the highest levels of trust, with Singapore at 81% and Indonesia at 76%. On a global scale, the European Union garners more trust than either the US or China for effective AI regulation. A median of 53% across 25 countries surveyed in 2025 trusted the EU, compared to 37% for the US and 27% for China. The report concludes its public opinion chapter by noting that Southeast Asian nations remain among the world’s most optimistic about AI’s future. In countries like China, Malaysia, Thailand, Indonesia, and Singapore, over 80% of respondents anticipate AI profoundly changing their lives within the next three to five years. Malaysia, in particular, recorded the largest increase in this sentiment between 2024 and 2025.