Leggi in ItalianoWhenever a new report on artificial intelligence comes out, I am less interested in which model has topped the rankings than in what has actually changed. Stanford’s 2026 AI Index offers some useful answers. It is hardly a celebration of the technology: alongside the advances, it documents failures, risks, costs and promises that still need to be tested. Yet one conclusion seems increasingly difficult to avoid: AI is no longer a technology we discuss mainly in the future tense.It is already changing how we work, conduct research and make decisions. Our institutions, meanwhile, seem to be moving at a different speed.
The technical advances are striking, though far from consistent. In tests where AI agents must operate a computer and complete practical tasks, success rates have risen from about 12% to 66.3%. The margin for failure remains considerable, but so does the progress. In 2025, a model achieved a gold-medal-level performance at the International Mathematical Olympiad, while these systems can still struggle to read an analogue clock. The contradiction is only apparent. AI capabilities do not develop in the same way as human abilities: a system may be exceptional at one task and remarkably fragile at another. Strong benchmark results should never be confused with general reliability.
What interests me even more is what is happening beyond computer science laboratories. AI is being used in protein research, genomics and weather forecasting, where some systems can produce forecasts much faster than traditional approaches. In medicine, tools that draft clinical notes have reduced documentation effort by as much as 83% in one of the healthcare settings examined. This does not mean an algorithm can replace a physician. The report itself notes that, in a review of more than 500 clinical AI studies, only 5% used real clinical data. That distinction matters: answering a medical question correctly is not the same as demonstrating better outcomes for patients.
There is also growing evidence of economic benefits. Some studies report productivity gains of 14–15% in customer support and 26% in software development. But those gains are uneven; on certain tasks requiring judgment and experience, AI may actually slow people down. And there is another finding I would not dismiss: in the United States, employment among software developers aged 22 to 25 has fallen by almost 20% from its 2022 peak. AI cannot simply be held responsible for the entire decline. Still, it raises a question for universities and the legal professions as well: if machines increasingly take on work traditionally given to beginners, how will the next generation become experienced professionals?
From a legal perspective, what concerns me most is the gap between technological development and our capacity to govern it. Documented AI incidents increased from 233 in 2024 to 362 in 2025, while transparency among leading developers remains inadequate. Add to this the environmental costs, the concentration of infrastructure and a technological competition increasingly dominated by the United States and China. For the European Union, the challenge cannot be confined to drafting rules. Europe also needs the capacity to research, develop and exercise meaningful control over the infrastructure on which these systems depend. Regulatory autonomy will have limited value without a credible degree of technological autonomy. This is not an argument for weakening the protection of rights. It is an argument for taking seriously the material conditions required to make that protection effective.
Too much of the public debate still seems trapped between two extremes: the expectation that AI will solve almost everything and the belief that every mistake proves it has little value. Neither position tells us much about what is actually happening. The progress is real, so are the limitations, and the benefits and costs will not be shared equally. The task now is not to choose between enthusiasm and suspicion. It is to distinguish what AI can already do, what it still cannot do, and what we should not delegate to it without adequate safeguards.
Main source: Stanford Institute for Human-Centered AI, AI Index Report 2026, especially chapters 2 (technical performance), 3 (responsible AI), 4 (economy), 5 (science), 6 (medicine), and 8 (policy and governance). Research report. The figures quoted come from different studies, benchmarks and samples and are not directly comparable.