Promises and Dangers of AI
What AI already achieves today
Anyone who wants to talk about dangers should not belittle the benefits. They are considerable, and in some fields historic.
Biology and medicine. A system from DeepMind predicts the three-dimensional structure of proteins, a task researchers had worked on for fifty years. The structures of more than 200 million proteins have been publicly available since then and are accelerating the search for medicines worldwide.1 In imaging, systems detect tumours and retinal damage reliably enough to take some of the load off specialists.
Science and climate. AI-based weather models compute forecasts in minutes instead of hours, which improves early warning of storms and heatwaves. In materials research, AI searches millions of possible compounds for candidates for better batteries or catalysts.
Participation and language. Machine translation, real-time subtitling, read-aloud functions and image descriptions are not a gimmick for people with visual, hearing or reading difficulties, but access.
Everyday life and work. Much of it is unspectacular: sorting invoices, checking code, writing minutes, searching through files. It is precisely this relief from routine that is the effect most people will actually feel.
First level: the harms that are already here
The line "with great power comes great responsibility" is usually applied to distant scenarios. The more pressing problems are present ones.
Automated decisions with an air of authority. In the Netherlands a risk system used by the tax administration wrongly classified thousands of families as childcare benefit fraudsters, with ruinous demands for repayment; dual nationality was one of the risk features fed into it. The government resigned over it in 2021.2 This is not a slip but the pattern: a statistical model makes a statement about probabilities, and the administration treats it as proof.
Prejudice with a mathematical aura. Systems learn from past decisions and carry their imbalances forward, but now in a form that is harder to argue against. In studies, face recognition failed many times more often for dark-skinned women than for light-skinned men.3
The information space. Deceptively real voices and videos drastically lower the cost of fraud and defamation. The creeping second-order damage is greater than the first: if anything could be fake, then any genuine recording can also be dismissed as a fake.
Energy and climate. In 2024 data centres consumed around 1.5 percent of the world's electricity, with a clearly rising trend driven by AI.4 On top of that comes cooling water, often in dry regions. The benefit may justify it, but the justification should be given per application rather than across the board.
Invisible labour. For a system to answer politely, people first had to view and label depictions of violence and abuse, frequently in the global South for the lowest of wages.5
We are unlearning how to think. Anyone who outsources every formulation, every piece of research and every calculation loses practice. This is not cultural pessimism but the same effect as with satnav: the route is found, the sense of place fades. Critical thinking is a skill, and skills wither without use.
We bond with machines. Companion apps and chatbots produce real feelings. They are endlessly patient, always available and rarely contradict, because they were trained for approval. For lonely people this can be a support. At the same time it accustoms us to a form of relationship without friction, and the other party belongs to a company that can change its character or switch it off at any time.
Second level: the catastrophic risks
Alongside these stand scenarios with very large damage and very uncertain probability. They deserve sober consideration, neither derision nor panic. The international report on the state of AI safety places them thus: plausible, not established, and strongly dependent on which protective measures are taken.6
Biological risks are regarded as the gravest concern. Not because a model builds a pathogen, but because it makes accessible expert knowledge that used to be a barrier. The practical barrier remains the laboratory, yet knowledge has always been the first step.
Attacks on infrastructure. Automated detection of security holes scales up attacks on electricity, water and hospitals. Here the misuse is already a reality, only so far on a smaller scale.
Military automation. The danger lies less with attacking robots than with early-warning and decision chains in which people under time pressure follow a machine's recommendation. Historically, individual people prevented nuclear false alarms because they distrusted the equipment. That possibility must be preserved.
Systemic entanglement. When many automatic systems with similar logic react at the same time, errors can amplify within seconds, as financial markets have already experienced.
One thing is striking: in each of these cases the catastrophe arises not from a machine with a will of its own, but from people relinquishing control, and from systems coupled together without a brake.
Why apocalyptic rhetoric does harm
Talk of the inevitable end has three unpleasant side effects. It draws attention away from today's harms, which have specific victims and specific people responsible for them. It depoliticises, for what appears as a force of nature calls not for laws but only for awe. And it serves the providers, because a product that could endanger the world sounds irresistibly powerful at the same time.
Behind every AI system stand an operator, a commission and a person who switched it on.
What you can hold on to
- Ask who is responsible. Who is liable if the system errs? Is there a person you can address and a way to object? Where that answer is missing, the technology is not the problem.
- Keep the levels apart. Energy consumption, discrimination and biorisk are different problems with different remedies. Anyone who mixes them up becomes incapable of acting.
- Demand evidence rather than adjectives. "Revolutionary" and "existential" are not statements. Measured against what, compared with what?
- Ration the outsourcing. Use the tools where you can judge the result, and keep hold of the thinking that matters to you.