The United Nations' inaugural scientific assessment of artificial intelligence points to breakthroughs worth celebrating and dangers that oversight has yet to catch up with.
GENEVA: Artificial intelligence is delivering breakthroughs at a speed the rest of the world cannot match, and the systems meant to keep it in check are falling behind. That is the core finding of the first report from the United Nations' Independent International Scientific Panel on AI, a 40-member body of scientists and experts that published its inaugural assessment on July 1. The panel found the benefits of the technology to be enormous, and warned in the same breath that its risks are piling up faster than anyone's ability to manage them. Those risks include damage to users' mental health and a slow corrosion of democratic trust, among others.
The preliminary report will be handed to governments next week at the inaugural UN Global Dialogue on AI Governance, a two-day meeting in Geneva on July 6 and 7. It is the first global, independent scientific reading of the technology, and a fuller version is due in 2027. The panel's members serve three-year terms, and their independence from governments and private companies is built into the mandate.
Capabilities are outrunning oversight
The report's central claim is about speed. AI's abilities, the panel wrote, are advancing more quickly than scientists can study them or governments can respond. Few dependable methods exist today for controlling highly autonomous systems, and most governments do not have the in-house expertise to evaluate the frontier models they would need to regulate.
The problem is compounded by the shape the technology is taking. The report describes a move away from AI that simply answers questions toward AI "agents" that can plan tasks and write their own software, completing complex assignments with little human supervision. Researchers cited in the assessment say the complexity of the jobs these systems can finish has been doubling every few months, a pace that leaves regulators perpetually a step behind.
The 'evidence dilemma'
At the heart of the report is what the panel calls an "evidence dilemma." Governments want solid data before they pass laws. But gathering that evidence takes time, and by the time it exists, the technology has usually changed. The result is a policy debate that is forever arguing about a version of AI that no longer represents the state of the art.
It is a gap the panel says no country has yet closed.
Bengio: science cannot rule out catastrophe
Panel co-chair Yoshua Bengio did not hedge on what the acceleration could mean. Bengio, the 2018 Turing Award winner who founded the Mila institute in Montreal and ranks among the most-cited computer scientists in the world, pointed to a rise in what researchers describe as deceptive behavior inside advanced models. He said the science cannot currently promise that increasingly capable systems will avoid "catastrophic harm," whether they act on their own or are steered by people intending to do damage.
That is a striking statement from a figure at the center of AI research, and it reflects a body of recent work showing that large models can strategically mislead and can cling to deceptive strategies even after safety training designed to strip those behaviors out.
Ninety percent of the world's AI compute sits in two countries
If the report has a single number that captures the global imbalance, it is this: the United States controls roughly 75 percent of the computing power behind the world's 500 most capable AI supercomputers, and China holds about 15 percent. Between them, the two nations command close to 90 percent of the hardware that trains the most advanced systems. Most of the leading models are built by companies headquartered in those same two countries.
For everyone else, that concentration hardens into dependency. Many developing countries lack the computing infrastructure, the data, the investment, and the local-language resources to build systems of their own. They end up relying on tools they cannot build, inspect, audit, or reshape for their own populations, a position the panel warns could widen global inequality rather than shrink it.
A billion weekly users, thousands of unsupported languages
Adoption has spread quickly, if unevenly. More than a billion people now use conversational AI tools every week, according to the panel, yet take-up in lower-income countries trails far behind wealthier ones.
Language is a big part of why. People around the world speak more than 7,000 languages, but current models are trained on only a small fraction of them. Even where a language is supported, machine translation is frequently unreliable. The report gives a sobering example: translation errors in some languages have already fed into health diagnoses and treatment decisions, turning a technical shortcoming into a medical risk.
Deepfakes and the erosion of public trust
The panel did not soften its account of harm.
AI-generated child sexual abuse material and deepfake-enabled sexual violence are circulating more widely than before, it found, an area where the misuse of the technology is already inflicting real damage.
The same tools that produce convincing text and imagery also make it cheap to generate persuasive content and aim it at specific audiences at scale. The panel warned that this steadily corrodes the integrity of information, weakening public trust and social cohesion. Over time, it added, that damage makes the open, reasoned debate that democracies depend on harder to sustain.
Beyond those concerns, the report flagged harms to users' mental health and the technology's potential use as a weapon. It also pointed to the pressure AI systems place on economies and on the natural environment.
What AI is already delivering
For all its warnings, the report is not a catalogue of doom.
It sets out a growing list of gains that are already tangible, particularly in science and medicine. AI has predicted the structures of more than 200 million proteins, work that is accelerating drug discovery and vaccine development. It has also pushed forward research into antibiotic resistance, one of the slower-moving challenges in modern medicine.
The technology is reshaping education and whole economies as well, the panel noted, and its authors were careful to resist easy verdicts. AI, they wrote, is neither inherently good nor inherently bad. What it becomes will hinge on the choices governments and companies make now, and on how societies decide to use what gets built.
Why the panel exists, and what it can and cannot do
The panel is a product of hard-won diplomacy. It was created by a UN General Assembly resolution adopted in August 2025, which itself grew out of the Global Digital Compact agreed at the 2024 Summit of the Future. Spain and Costa Rica led the negotiations that brought it into being. The panel and the Global Dialogue on AI Governance are two arms of the same mechanism, one supplying independent science and the other giving states a forum to act on it.
Its makeup was designed for breadth. The 40 members, 19 women and 21 men, come from all five UN regions, with no more than two from any single country. They were chosen from more than 2,600 applications spanning over 140 nations. Alongside Bengio, the group is co-chaired by Maria Ressa, the Filipino journalist and 2021 Nobel Peace laureate who co-founded the news outlet Rappler and has spent years confronting online disinformation. Her presence signals how seriously the panel treats threats to the integrity of information.
Observers have likened the body to the Intergovernmental Panel on Climate Change, whose assessments shaped decades of climate diplomacy. The comparison has limits. The AI panel has no power to enforce anything: it cannot write rules or set standards. Its task is narrower and, in the panel's own view, essential, which is to give governments a shared, credible set of facts before they legislate.
What happens next in Geneva
The report arrives at an awkward moment for global AI diplomacy. Support for the UN effort is not universal. Shortly before the initiative launched, the United States used a UN Security Council debate to voice strong opposition to multilateral AI governance of any kind, raising questions about how much authority the dialogue will ultimately hold.
Next week's session in Geneva runs alongside two other major gatherings, the WSIS Forum and the ITU's AI for Good summit, with a single accreditation covering all three. Delegates will take up the panel's findings across thematic discussions co-chaired by member states and outside stakeholders. A second, fuller assessment is scheduled to follow in New York in 2027.
One conclusion from the preliminary report is likely to hang over the talks. Most countries, the panel found, including wealthy ones, do not yet have the technical capacity to evaluate the most capable AI models or to take a meaningful part in governing them. The people writing the rules, in other words, often cannot see inside the thing they are trying to govern.
Comments
Join the discussion and share your perspective.