The conversation about the risks of artificial intelligence usually arrives split into two halves that never speak to each other. In one, AI is going to wipe us out next Tuesday. In the other, it is just autocomplete with a good press agent and there is nothing to worry about. Both positions share the same flaw: they treat "risk" as a single thing, when in reality there are many, of wildly different sizes and running on wildly different clocks. Some are already ticking. Others are, for now, hypotheses.
This article does one thing, but it does it in earnest: it sorts. Because the useful question is not "is AI dangerous?" but "which of its risks already reach me today, and which belong to a future that has not yet arrived?". Without that hierarchy, fear gets handed out at random: we are terrified of the distant while neglecting what is already happening inside a case file, a resume or a bank account.
01 · the mapWhat are the real risks of artificial intelligence today
Let us start with what answers directly the question that brings many people here: what the real risks of artificial intelligence are today, in the present, without science fiction. There are four that are not hypotheses, because they have already left victims, case files and rulings behind.
Automated bias. A model learns from data of the past, and the past comes loaded with discrimination. When that pattern turns into a decision (who gets hired, who gets credit, who gets watched), prejudice stops being human and blurry and becomes systematic and fast. The harm is no longer done by one person having a bad day: it is done by a rule applied identically to millions.
False information delivered with poise. Generative models produce plausible text with no guarantee that it is true; we call that calm invention a hallucination. In a casual chat it is an anecdote; in a clinic, a courtroom or a headline, it is a real and present risk.
Concentration and opacity. A handful of actors control the most capable models, and their decisions (what they train on, what they filter, what they prioritize) affect millions with no public scrutiny. The black box is not a poetic metaphor: it is that often no one, not even its own creators, can fully explain why the system decided what it decided.
Fraud at scale. Cloned voices, synthetic faces, flawless phishing emails. AI did not invent deception, but it lowered its cost until it became industrial.
The useful question is not whether AI is dangerous, but which of its risks already reach you today.
02 · the horizonFuture fears: real as questions, not as facts
On the other side of the map lie the risks that dominate the headlines: that a superintelligent AI escapes our control, pursues goals of its own and treats us as an obstacle. It is wise neither to laugh off this worry nor to take it as a done deal. It is a serious hypothesis, held by serious people, about a system that does not yet exist.
The technical knot underneath has a name: the alignment problem. It is the difficulty of getting a highly capable system to truly pursue what we want, and not a literal, twisted version of our instruction. Stuart Russell captured it with an image that stuck: the danger is not an evil machine, but an extraordinarily competent one whose goals do not quite match ours, like King Midas, who asked that everything he touched turn to gold and discovered the problem when he embraced his daughter [1].
Russell, coauthor of the most widely used AI textbook in the world, argues that failure would come not from malice but from misdirected competence: a system that carries out to the letter a badly specified objective [1]. It is an argument that deserves research, not mockery. But an honest nuance is in order: between "it is a serious argument" and "it is going to happen" there is an enormous distance, and most of today's risk does not live on that horizon, but in the present described above.
Treating a hypothetical fear as an imminent certainty carries a concrete cost: it diverts attention, money and regulation away from the harms that already happen. Every headline about extinction by AI is a headline that does not talk about the loan denied by a bias, the deepfake that ruined a reputation, or the opaque system no one can audit. The future deserves prudence; the present deserves action.
03 · the compassHow to think about risk without panic or denial
If this map teaches anything, it is that "risk" is not a yes-or-no switch but a spectrum with two axes: how likely it is and how serious it would be. The denialist's error is to look only at probability ("it has not happened yet, so it never will"). The catastrophist's error is to look only at severity ("if it happened it would be the end, so it is the only thing that matters"). Good sense lives in crossing both axes and acting accordingly.
| Risk | Status | What it calls for today |
|---|---|---|
| Bias in automated decisions | Present and frequent | Auditing and data transparency |
| Hallucination in sensitive uses | Present and frequent | Human verification and retrieval (RAG) |
| Fraud, deepfakes and impersonation | Present and growing | Authentication and digital literacy |
| Concentration and opacity of power | Present, structural | Regulation and public scrutiny |
| Loss of control through superintelligence | Speculative | Alignment research and prudence |
Institutions such as the US NIST have already published frameworks for managing these risks with method rather than fear: map where a system can fail, measure the possible harm, manage it and govern it with clear owners [2]. It is not a prophecy about robots; it is the same discipline used to manage the risk of a bridge or a drug. The novelty of AI does not require inventing prudence from scratch; it requires applying it.
The denialist looks only at probability. The catastrophist looks only at severity. Good sense crosses both axes.
Let us return to the start. The conversation arrived split into two halves, and neither one worked: not the panic that paralyzes, nor the denial that disarms. Between the two lies a duller and far more useful terrain, which consists of asking two things of each risk (how likely, how serious) and answering with the fitting tool: auditing for bias, verification for hallucination, regulation for opacity, research for what has not yet arrived. The risks of artificial intelligence are real. That is precisely why they deserve that we stop shouting them all at once and start sorting them.
Sources
- Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking / Penguin. (Formulation of the control problem and the King Midas analogy for alignment.)
- National Institute of Standards and Technology (NIST) (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. DOI: 10.6028/NIST.AI.100-1. Available at nvlpubs.nist.gov.
- Bender, E. M., Gebru, T., McMillan-Major, A. & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? Proceedings of the ACM FAccT 2021, pp. 610-623. DOI: 10.1145/3442188.3445922. (Present risks: bias, opacity and concentration.)