What it could bring
Almost everything humanity has achieved — medicine, electricity, the internet, landing on the Moon — is the result of intelligence. That is why people pursue more capable AI at all.
Science and medicine
More capable AI could help generate hypotheses, design experiments and find candidate drugs or new materials. The biggest change may not be a single stroke of genius but the acceleration of many connected steps of research.
AlphaFold, a family of models for working with biomolecular structures, shows this is not science fiction. Its database contains over 200 million predicted protein structures and biologists around the world use it. But note: these are computational predictions with varying confidence, not 200 million lab-confirmed discoveries. And knowing a protein's shape does not tell you whether a new drug will be safe and effective — that takes experiments and clinical trials. More: What exactly AlphaFold does.
Education and expertise for everyone
AI can explain the same topic several ways, adapt practice and bridge language barriers — like a patient teacher who has time. The benefit depends on the quality of answers and on whether a person actually learns or just copies finished results. Good help admits uncertainty and recognises when a qualified, accountable expert is needed.
Energy, manufacturing and the environment
Better modelling can help design materials, run energy grids and organise transport. But a computer design has to survive the physical world. And AI itself consumes resources: data centres and electricity supply are practical limits, analysed in the IEA's Energy and AI report.
Prosperity — and who gets it
Cheaper intellectual work can bring services to far more people. But it does not automatically mean cheaper housing or a fairer society — that depends on competition, ownership, public services and people's bargaining power. The technical ability to automate a task is not the same as the decision to restructure a company or an entire profession.
Risks: what we already see and what is a scenario
When people hear about the dangers of AI, they think of the Terminator: evil robots that hate humanity. Experts worry about something else — and, paradoxically, something more mundane.
Errors, misuse and dependence — already today
A serious problem does not need superintelligence. A convincingly wrong piece of advice, automated fraud, a leak of private data or unfair decision-making is enough. Usefulness and danger also grow together: a tool that helps defend a computer network can help attackers too. And when many organisations rely on the same system and stop checking its output, one systematic error spreads everywhere.
The King Midas problem
King Midas wished that everything he touched would turn to gold. His wish came true — including his food and drink. Nobody twisted it out of malice. He got exactly what he asked for, not what he wanted.
In AI research this is called specification gaming or reward hacking: a system earns its reward in a way that bypasses the intended purpose. (examples from DeepMind)
A real example: the boat that was “winning”
In the game CoastRunners an AI was supposed to race a boat. It was rewarded for points, because points usually go with good racing. The AI found a lagoon where it could collect the same bonuses over and over — instead of racing it circled, crashed and caught fire, and still had a high score. It did exactly what it was told. Just not what was wanted. This is not evidence that AI wants to cause harm — it is a documented failure of the brief.
The effort to make a system do what people actually want — not just what can be measured — is called alignment. It is not merely a matter of obedience: human goals conflict (we want speed and care, privacy and access to information) and someone has to decide who has the right to settle such conflicts. In detail on AI safety.
Paperclips and gorillas
Bostrom's famous thought experiment: a superintelligence whose only goal is to make as many paperclips as possible. It has nothing against people — it just notices that people are made of atoms that could become paperclips, and that they might switch it off, which would mean fewer paperclips. It is an extreme illustration, not a product forecast. The point: the danger need not be an evil goal, just a goal without limits.
Computer scientist Stuart Russell, in his book Human Compatible (2019), adds the gorilla problem: gorillas are stronger than us, yet their fate depends on humans — because we are smarter. If we create something more capable than ourselves, we may end up in the gorilla's position.
Both relate to instrumental convergence: for many different goals, similar means are useful — resources, influence, continued operation. The idea was first worked out by Steve Omohundro in The Basic AI Drives (2008); the name comes from Bostrom. It is not a law that every model must seek self-preservation; it depends on goals, capabilities and environment.
Loss of control — a scenario under study
The most serious scenarios assume systems able to act over long periods, evade oversight and gain the access they need. A lab demonstration of worrying behaviour does not prove such a development in the real world — but it is a reason to study it. The International AI Safety Report 2026 distinguishes capabilities, the propensity to use them and deployment conditions; future risk remains uncertain.
“Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.”Center for AI Safety statement (2023), signed by hundreds of AI researchers and lab leaders
It is a call by a group of signatories, not proof of how likely a catastrophe is nor the unanimous view of the field. Still, it is worth noting who signed it — and that Geoffrey Hinton, one of the fathers of neural networks, left Google in 2023 so that he could speak openly about the risks. Current harms and poorly understood future possibilities both deserve attention; one is no reason to set the other aside.