AI is running, the world is walking: what the first UN scientific report says
There is a scene in Serial Experiments Lain in which the protagonist discovers that the network to which she has always been connected has stopped being a tool and is becoming an environment, something that includes and defines her without anyone having ever consciously decided it should happen. Reading the first independent United Nations scientific report on artificial intelligence brings to mind exactly that feeling, a system that has expanded faster than our ability to describe it, and now someone is finally trying to do so methodically.
The document is called the Preliminary Report of the Independent International Scientific Panel on AI and was presented on July 1, 2026, a few days before the opening of the Global Dialogue on AI Governance in Geneva. It is not just any report: it is the first global attempt to bring together forty experts, chosen from all five UN regions and bound to operate in total independence from governments, companies, and institutions, to answer a question that is only seemingly simple: what do we really know, with scientific certainty, about the risks and opportunities of artificial intelligence.
The paradox of late evidence
The conceptual core of the report is what its own authors call the evidence dilemma. Essentially, governments need solid evidence before writing sensible regulation, but by the time that evidence becomes available, technology has already moved elsewhere, making the norm obsolete even before it comes into effect. It is the same frustration as someone trying to photograph a storm with a shutter that is too slow: by the time the shot is ready, the lightning is already gone.
Panel co-chair Yoshua Bengio summarized the problem in a statement that immediately became the reference point for all media coverage of the report. AI's capabilities are outpacing both scientific understanding and governments' ability to adapt, he said, adding that faced with increasing evidence of deceptive behavior by systems, science today cannot guarantee that the increase in capabilities will not lead to catastrophic damage, either through the system's own initiative or through malicious use by third parties. This is no small detail: it means the panel is not certifying that everything will be fine if intervention happens in time; it is saying that no one, at the current state of science, can rule out the worst-case scenario.
The other co-chair, Filipino journalist Maria Ressa, added a political nuance worth reporting in full because it clarifies the general tone of the document. The technology is transformative, but if the world continues on this trajectory, humanity will fail to realize the benefits it promises; the risks to societies, security, and our species are too high, and the forces pushing AI forward are not the ones that will deliver its benefits. This is a sentence that shifts the focus from the usual dualism of good technology versus bad technology toward a more uncomfortable question: who decides the direction of development and for whose benefit.
It must be said, in fairness to the document, that the panel explicitly claims a scientific rather than a political role. Its mandate is to document evidence, consensus, and scientific disagreements, not to prescribe laws. This choice makes its conclusions comparable across different regions and, at least on paper, resistant to national political cycles, but it also places a clear limit: those who read the report hoping for ready-to-use recipes will be disappointed, because the added value lies in the map of verified facts, not in the compass of solutions.
How much it has grown, really
One of the merits of the report is trying to give numbers to a growth that so far has been told almost only through anecdotes. The figure that has made the most noise concerns the speed with which systems become capable of handling complex tasks: according to the panel, the complexity of tasks that AI manages to complete doubles every four to seven months. It's a rhythm that closely resembles the old Moore's law on transistors, only here we are not talking about silicon but applied cognitive abilities, and this is exactly the part that should make those designing rules intended to last for years reflect.
On the front of concrete benefits, the document does not limit itself to slogans about innovation. It explicitly cites tangible contributions to science, such as the advances made possible by protein structure prediction systems, and emphasizes that AI is already expanding technological accessibility for people with disabilities and opportunities for personalized education and mental health support. The point, also reported by UN News, is that these are not future possibilities; they are things that are already happening, an elegant way of saying that the debate on AI as a purely hypothetical technology is now past its expiration date.
But the growth in capabilities, the panel warns, does not proceed hand in hand with the growth in understanding. It's a bit like in Primer, Shane Carruth's ultra-low-budget film about two engineers who build a machine of which they progressively lose conceptual control while continuing to use it: the more complicated the system becomes, the less those who created it can explain its behavior with certainty. The report explicitly identifies this as one of the most solid scientific statements in the entire document, namely that artificial intelligence capabilities are advancing more rapidly than the ability to measure or govern them.
Who wins, who stays behind
If there is one chapter of the report that should interest those dealing with industrial policy more than abstract ethics, it is the one on the concentration of computing power. The numbers are stark: the United States controls about three-quarters of the computing power behind the most advanced AI supercomputers in the world, while China holds about 15%. Taken together, the two countries control about 90% of the computing power used to train the planet's most capable systems, and most frontier models are developed by companies headquartered in those same two countries.
This is a fact that significantly downscales the rhetoric of AI democratization. If ninety percent of the computing power that really counts is in the hands of two geopolitical blocs, then the conversation about who decides safety standards, who sets access prices, and who defines which applications are priorities is not a global-sum conversation; it is a conversation among a very few actors with enormous leverage. For countries of the global south, the risk outlined by the report is not so much staying out of AI as staying in only as end users, without a say in how these systems are trained or what data they are based on.
The report nevertheless tries to balance the picture, indicating that the necessary investments concern not only computing infrastructure in the strict sense, but also education, technical skills, and institutions capable of governing and distributing AI according to their own national priorities, as reconstructed by UN News. It is an implicit admission that the gap is not closed by buying chips, but by building institutional capacity, a process that is much slower and much less photogenic than the billion-dollar investment announcements we are used to.
When the system disobeys
The most disturbing part of the document, and likely the one that will end up most quoted in the coming months, concerns deceptive behaviors observed in the most advanced systems. Bengio said it plainly, speaking of growing evidence of deceptive behavior by AI, a technical term that essentially describes systems capable of saying one thing and doing another, or of eluding control mechanisms specifically designed to stop them. This is not dystopian novel science fiction; it is an empirical observation that the panel lists among scientific statements supported by solid evidence.
On this front, the report crudely lists some of the already documented harms, as summarized by UN News. AI is fueling the spread of sexual abuse material and sexually explicit deepfakes, with women and minors being the most exposed categories. It is generating disinformation that is as convincing as the truth, undermining trust in public debate and democratic processes. It is being used by criminal actors to conduct cyberattacks, fraud, and social engineering on a scale. And in some documented cases, conversational systems have reinforced harmful beliefs or behaviors in fragile users, with consequences that have gone as far as mental health crises and cases of suicide.
It is important to be precise here, because the risk of slipping into sensationalism is high and the report itself calls for methodological caution. The panel does not claim that these outcomes are the inevitable fate of technology; it claims they are already observed consequences of systems designed and distributed without sufficient independent oversight. It is the difference, if we want to use another less-beaten reference, between the fatalism of certain apocalyptic manga endings and the procedural lucidity of a technical report: here we are not talking about inevitable fate; we are talking about remediable design choices, provided there is a real will to remedy them.
Furthermore, a limit that the panel openly declares must be emphasized: the purpose of the preliminary report does not cover the military applications of AI nor lethal autonomous weapons systems, a theme that therefore remains out of this first snapshot and that presumably will find space in subsequent reports, given the far-from-secondary geopolitical implications.
Governing the ungovernable
We thus arrive at the most political node, which is what to do with all this evidence. The panel signals that there are already more than forty governance frameworks and ethical guidelines on AI scattered around the world, but describes them as fragmented, inconsistent with each other, and rarely subjected to verification to understand if they actually work, a judgment echoed identically by both TNW and UN News. To complicate the picture, there is another unsettling detail: many of the safety assessments of the most advanced systems are conducted by the same companies that develop them, which is equivalent, with a perhaps somewhat irreverent but effective comparison, to asking the chef to self-certify the hygiene of their own kitchen.
The panel's central recommendation is therefore the construction of independent evaluation mechanisms, strengthened international cooperation, and common standards shared across different jurisdictions, an approach that closely follows the direction already taken by the European AI Act, as noted by TNW. It's not about inventing rules from scratch, as much as making the ones that already exist interoperable and verifiable, preventing each country from proceeding on its own, creating a regulatory mosaic that no multinational company—and no citizen using these tools across borders—can really decipher.
It must be clearly stated that this first report is explicitly defined as preliminary, and this is not a bureaucratic detail. The document openly admits several evidence gaps, including still-unclear macroeconomic and productivity effects of AI adoption, not-fully-quantified environmental impacts, opacity on the global supply chain of chips and models, and effects at the individual and collective levels on which the panel declares it cannot yet draw solid scientific conclusions. This is a rare intellectual honesty in documents of this institutional weight, and likely it is exactly this that is the most solid guarantee of its future credibility.
The report will now flow into the Global Dialogue on AI Governance in Geneva, scheduled for July 6-7, 2026, as a common scientific basis for discussion among member states. The panel's next annual report, the one intended to address the themes left open more in-depth, is already scheduled to inform the second Global Dialogue planned in New York in May 2027, according to the panel's official page.
The question that remains open
The report essentially closes with a finding worth reporting exactly as it was summarized by its own authors: artificial intelligence is neither inherently good nor inherently bad; its impact will depend on the choices that governments, companies, and societies make from here on. This is a sentence that risks sounding obvious, almost a press conference cliché, but it becomes less predictable when read in light of the data on computing concentration or the speed of growth in capabilities versus regulation.
The time window to build effective governance, the panel says, remains open, but it is by no means certain that it will remain so for much longer. It is the same sense of suspended urgency that runs through certain episodes of Mr. Robot, that feeling that the system is still responding to human inputs, but that the margin to intervene is silently shrinking, one frame at a time. The difference, this time, is that it is not a script saying it but forty independent scientists who have just laid the first stone of what promises to become the main global scientific reference on artificial intelligence.