Stanford University’s annual index report on the state of artificial intelligence notes that the technology is advancing faster than society can keep up with. System capabilities are breaking records, investments have reached unprecedented heights, while security measures and public trust are falling further and further behind.
Endless possibilities
More than 90% of the most notable new models in 2025 were developed by industry, not academic institutions. Some of them are already outperforming humans on competitive-level PhD exams in the natural sciences and mathematics. In one of the key engineering tests, the score rose from 60% to nearly 100% in just one year.
The percentage of organizations that have implemented AI has reached 88%. Four out of five college students are already using generative AI systems in their studies—and that number continues to grow.
Smart and helpless at the same time
Researchers describe what is known as the “uneven frontier” of AI—rapid progress in some tasks and unexpected failures in others. The Gemini Deep Think system won a gold medal at the International Mathematical Olympiad, competing against the world’s top high school students.
However, even the best model correctly reads the time on an analog clock only 50.1% of the time—slightly better than a random guess. AI agents performing real-world computing tasks saw their success rate jump from 12% to 66%—but they still make mistakes in about one out of every three attempts.
Safety measures are not keeping pace
Companies readily report on what their models are capable of—but remain silent about how safe they are. The number of documented incidents related to AI—including misinformation, algorithmic bias, and privacy violations—rose to 362, up from 233 the previous year.
Researchers have identified another problem: making a model safer does not necessarily make it better. Often, the opposite is true—improving security reduces the accuracy of responses, and developers have to strike a balance between these two goals at the same time.
Conquered the world faster than the internet
Generative systems have achieved 53% coverage of the population in less than three years—faster than personal computers or the internet did in their day. The rate of spread generally depends on a country’s income level, but some countries are performing better than might be expected. Among the leaders are Singapore (61%) and the UAE (54%). The United States ranks only 24th, with a rate of 28.3%.
The real benefits of these tools are also growing—specifically, the time, money, and effort that people save every day thanks to AI. Researchers estimate that it is worth $172 billion a year for American users—and the amount per person has tripled over the past year.
Experts’ optimism and public skepticism
Among industry experts, 73% expect AI to have a positive impact on the labor market—but only one in four members of the general public shares this view. A similar gap is evident in assessments of the impact on the economy and the medical field. Experts are optimistic, while the public is much more cautious.
A separate issue is trust in regulators. The United States had the lowest level of public trust in its own government regarding AI regulation among all countries surveyed: just 31%. The European Union is currently viewed by the world as a more reliable regulator than the United States or China.
According to hai.stanford.edu
