# Power Before the Button AI can move power from the visible decision to the selection of options and the story that makes one choice appear natural. - Forfatter: Mikkel Freltoft Krogsholm - Type: Essay - Udgivet: 2026-10-07 - Opdateret: 2026-10-07 - Sprog: en - Emner: ai, power, responsibility, decisions, human control - Kanonisk URL: https://mikkelkrogsholm.dk/en/articles/magten-foer-knappen/ --- On the morning of 28 February, Sobhan wanted to wear sports socks to school. His mother said no and found a pair she thought suited him better. Later, the socks were the only way she could identify his body. That is how [Jyllands-Posten’s article about the attack on Shajareh Tayyebeh primary school](https://jyllands-posten.dk/international/mideast/ECE19697761/er-ai-bedre-end-mennesker-til-at-udpege-bombemaal-ny-data-om-skoletragedie-i-iran-rejser-spoergsmaal/) in Minab in southern Iran begins. Two American Tomahawk missiles struck the school. More than 150 people were killed, including at least 123 children. [Bloomberg reconstructed the path to the attack](https://www.bloomberg.com/graphics/2026-iran-school-attack/) through interviews with more than two dozen current and former US officials, including people familiar with the Pentagon’s as-yet unpublished investigation. Their account is not a story about a machine deciding on its own to bomb a school. It is more disturbing because humans and machines were entangled at every step. The school grounds had once been part of a military site, and the US target database continued to label the location as an Islamic Revolutionary Guard Corps facility. Yet commercial satellite images already showed in 2017 that walls and separate entrances divided the school from the base. Images from 2018 showed colourful walls, a football pitch and markings for children’s games. An analyst recorded the changes in 2019, but the note ended up in a system that was not connected to the database used for targeting. The old classification was fed into Palantir’s AI-supported Maven Smart System together with other possible targets. The school emerged as a recommended target for the first day of the war. According to Bloomberg, some users expected Maven to discover outdated information and contradictions by itself, even though the system was not built to do that. Work that had previously taken hours could now be completed in minutes as the United States struck more than a thousand Iranian targets during the first 24 hours. Palantir denies that the company is responsible for the underlying data or intelligence shortcomings, and no evidence has been published that the software failed technically. That distinction matters. Maven did not invent the old classification. It made it fast, coherent and ready for action. We often talk about *human in the loop* as if human presence by itself solves the problem. A person sees the machine’s recommendation, makes the final assessment and presses the button. The decision is therefore human, and responsibility has somewhere to live. It sounds reassuring. It simply rests on a very narrow understanding of power. ## The hand we can see In 1957, the American political scientist [Robert Dahl offered](https://doi.org/10.1002/bs.3830020303) a simple idea of power: A has power over B when A can get B to do something B would not otherwise have done. This is visible power. Someone wants something, someone else resists, and in the end we can see who got their way. On a battlefield, that understanding almost automatically directs our attention to the final decision. Who authorised the attack? Who gave the order? Who pressed the button? [The UN’s independent fact-finding mission](https://www.ohchr.org/sites/default/files/documents/hrbodies/hrcouncil/sessions-regular/session63/advance-versions/a-hrc-63-61-auv-ffmi.pdf) found reasonable grounds to believe that the attack constituted a war crime. The United States had not done everything feasible to determine whether the school was a military objective, and the failure, according to the mission, went beyond simple negligence. These are important questions. A military cannot wash away its responsibility by pointing to a computer. The model did not enlist, choose the aims of the war or grant itself access to weapons. People built the system, selected the data, established the procedures and decided that it should be used. Yet the visible hand can trick us into overlooking everything that happened before it moved. In Minab, people authorised and carried out the attack. But the target they saw had already been produced through databases, satellite imagery, disconnected IT systems, an algorithmic ranking and a workflow designed for extreme speed. When a model can sort more images, phone data, movement patterns and intelligence than any person could ever review, its most important contribution may not be the answer. It may be the selection of questions the human is then allowed to answer. ## Power over the list A few years after Dahl, [Peter Bachrach and Morton Baratz](https://doi.org/10.2307/1952796) pointed to the second face of power. Power exists not only in the conflicts we can see. It also exists in the ability to determine which issues reach the agenda at all. The person who writes the meeting agenda does not necessarily have the final word. But if they can keep certain problems and proposals out of the room, much has already been decided before the discussion begins. AI makes this form of power enormously scalable. A model can review an amount of material that would require thousands of hours of human labour and reduce it to a short list. That can be a major benefit. It can find connections we miss, detect dangers earlier and draw attention to alternatives that human habits or prejudices kept out. But the same ability can narrow the world. When the system filters, ranks and recommends, it also determines what is not seen. If the human enters only after the sorting, the machine has already helped to set the agenda. This applies far beyond targets in war. A doctor sees the diagnoses the system considers most likely. A bank adviser sees the customers the model has flagged as risky. A hiring manager sees the applicants who passed the automated screening. A minister sees a summary of thousands of consultation responses. In all four cases, a person can make the formal decision while being cut off from large parts of the material on which it rests. That changes what it means to choose. There is a difference between choosing among options and having influence over which options appear in front of you. ## The third face [Steven Lukes took the analysis of power one step further](https://plato.stanford.edu/entries/feminist-power/). Power can also shape how we see the world: our sense of what is natural, realistic and in our own interest. The strongest power need not win a conflict. It can help determine which conflicts we are able to notice in the first place. Here AI becomes more than a sorting machine. A system rarely presents raw data alone. It gives the data names, categories and probabilities. A person may be presented as high risk. A building may appear to be a likely target. One explanation may be labelled the most relevant. The words can be technically precise, but they also create a story about the situation: Who constitutes the problem? What counts as evidence? Which action appears sensible? When a number appears on a screen, it easily takes on the character of a fact found in the world. But the number is the result of a long chain of human and machine choices. Someone decided which data count. Someone formulated the system’s objective. Someone chose which errors are most costly and which can be accepted. The model learned patterns within that frame and turned them into a ranking, which in turn shapes the human gaze. AI does not need a political will of its own to have this effect. When I say that AI can acquire power, I therefore do not necessarily mean that the model wakes up with interests of its own. The power lies in the whole circuit connecting data, model, interface, institution and human being. AI can be an instrument of power, but it can also take on an independent function within the circuit because its selection and wording become something people organise their actions around. I do not think we can settle the question once and for all by calling AI either a tool or a power holder. It can be both. It depends on how it is used and on how much of our own judgement we hand over to it. ## Criticism requires more than mistrust It is easy to conclude that we simply need to use AI critically. Read the answer carefully. Check the sources. Remember that the model can be wrong. The problem is that uncritical use does not arise only from laziness or lack of training. Automated systems change the situation in which we make decisions. In 2010, researchers Raja Parasuraman and Dietrich Manzey [reviewed the research on automation bias](https://pubmed.ncbi.nlm.nih.gov/21077562/). They found that people can overlook problems the system fails to flag and act on recommendations that are wrong. This affected novices and experts alike and could not be eliminated by a simple instruction to pay attention. That makes sense. If a system is fast and usually right, trusting it becomes rational. If the workload is also high, investigating everything from scratch becomes expensive. The human controller can become the slowest part of a process designed for speed, and every rejection may demand more explanation than every approval. Minab shows how this can become organisational. According to Bloomberg’s sources, Pentagon staffing intended to limit civilian casualties had been cut by around 90 per cent; at the responsible Central Command, the team had been reduced from ten people to one. No civilian-harm specialist reviewed the school before the missiles were fired. The principle of *human in the loop* was still satisfied on paper, but some of the people who were meant to challenge the result had been removed from the process. [The EU AI Act tries to make human oversight more concrete](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng). For high-risk systems, the responsible person must, among other things, be able to understand the system’s limitations, interpret its output, remain aware of automation bias and have the ability to override or stop the system. The rules do not apply to military weapons systems, but they point to something general: a person does not have real control merely by being near a decision. [The International Committee of the Red Cross](https://www.icrc.org/en/document/icrc-position-autonomous-weapon-systems) similarly argues that human control over autonomous weapons requires concrete limits on target type, time, geography and scale, as well as the possibility of timely intervention and deactivation. Presence is not the same as control. Control requires time to investigate the basis. Access to what the system filtered out. Authority to say no. The ability to stop the process without being punished for delaying it. And an accountability trail that makes visible who selected the target, the data, the model and the limits. I have previously written about the difference between using AI as [a crane that takes over the work and a spotter that helps the person become better](https://mikkelkrogsholm.dk/en/articles/spotteren-ikke-kranen/). That difference is also about power. The critical user does not merely ask for an answer but uses the machine to widen their view and challenge their first understanding. The uncritical user may receive an excellent answer while losing sight of how the question was framed. ## The storyteller on the other side Yuval Noah Harari has described language as the operating system of human civilisation. Money, nations, corporations and laws can organise millions of people because we can share stories about things that do not exist in nature in the same way as stones and trees. In a [talk about AI and the future of humanity](https://www.youtube.com/watch?v=LWiM-LuRe6w), he says that AI, through its mastery of language, has gained a master key to our institutions. His fear is not only that the machine can spread false stories. It can create and adapt narratives, learn our weaknesses and build a form of artificial intimacy that makes influence personal. I previously explored that thought in [*Schrödinger’s AI*](https://mikkelkrogsholm.dk/en/articles/schrodingers-ai/). Humans build societies with stories about hope, danger and the good life, and AI can now participate in producing the stories about its own future. In [*When Stories Grew Hands*](https://mikkelkrogsholm.dk/en/articles/da-fortaellingerne-fik-haender/), the story took another step: the agent could not merely describe the world but act on its description of it. The three faces of power show what happens between those two movements. First, AI helps make a decision. Then it determines which decisions reach the table. Finally, it helps shape the language we use to understand what is on the table and why it appears reasonable. That does not mean every conversation with a chatbot is hidden manipulation. AI can make us wiser, show us counterarguments and open perspectives we would not have found ourselves. I use it myself as a thinking partner. But that possibility depends on whether the conversation expands my view or merely delivers a more persuasive version of the first story. In uncritical use, AI gets to define a great deal for us. It does not need to command. It is enough for its answer to become the starting point, its categories to become our categories and its omissions to disappear so quietly that we never notice them. ## The path to the button A decision does not begin when a finger moves. It begins when something in the world is registered as relevant. It continues when data are gathered, categories are chosen, options are sorted and a particular explanation is made visible to the person who will eventually act. On the screen, the human may still appear as the final authority. But if that person cannot see what the system rejected, understand the story behind the ranking or genuinely oppose the recommendation, the human may be only the last link in a decision whose most important parts have already been made. Sobhan’s socks tell us nothing about algorithms. They tell us what exists on the other side of our abstract words about targets, probabilities and decision support. A system can process a school as a point in a stream of signals. For the family, it was a morning, an argument about clothes and a child who was supposed to come home again. Human responsibility therefore lies not only in the right to press the button or refrain. It lies in power over the entire path to the button: what we make visible, what we allow to disappear, and which story makes the final action look like the natural choice. --- ## Sources and further reading - Bloomberg: [*Inside the US ‘Kill Chain’ That Destroyed an Iranian School*](https://www.bloomberg.com/graphics/2026-iran-school-attack/), 18 September 2026. - Jyllands-Posten: [*Er AI bedre end mennesker til at udpege bombemål?*](https://jyllands-posten.dk/international/mideast/ECE19697761/er-ai-bedre-end-mennesker-til-at-udpege-bombemaal-ny-data-om-skoletragedie-i-iran-rejser-spoergsmaal/), 5 October 2026. - Amnesty International: [*Four months after horrific Minab school airstrike, accountability delayed*](https://www.amnesty.org/en/latest/news/2026/06/usa-four-months-after-horrific-minab-school-airstrike-accountability-delayed/), 25 June 2026. - UN Independent International Fact-Finding Mission on Iran: [report A/HRC/63/61](https://www.ohchr.org/sites/default/files/documents/hrbodies/hrcouncil/sessions-regular/session63/advance-versions/a-hrc-63-61-auv-ffmi.pdf), 2026. - Robert A. Dahl: [*The Concept of Power*](https://doi.org/10.1002/bs.3830020303), 1957. - Peter Bachrach and Morton S. Baratz: [*Two Faces of Power*](https://doi.org/10.2307/1952796), 1962. - Steven Lukes: *Power: A Radical View*, 2nd edition, 2005; [Stanford Encyclopedia of Philosophy on the three dimensions of power](https://plato.stanford.edu/entries/feminist-power/). - Raja Parasuraman and Dietrich H. Manzey: [*Complacency and bias in human use of automation*](https://pubmed.ncbi.nlm.nih.gov/21077562/), 2010. - European Union: [Regulation 2024/1689, Article 14](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng), on human oversight. - ICRC: [*ICRC position on autonomous weapon systems*](https://www.icrc.org/en/document/icrc-position-autonomous-weapon-systems), 2021. - Yuval Noah Harari: [*AI and the future of humanity*](https://www.youtube.com/watch?v=LWiM-LuRe6w), Frontiers Forum, 2023. - Previous essays: [*Schrödinger’s AI*](https://mikkelkrogsholm.dk/en/articles/schrodingers-ai/), [*When Stories Grew Hands*](https://mikkelkrogsholm.dk/en/articles/da-fortaellingerne-fik-haender/) and [*The Spotter, Not the Crane*](https://mikkelkrogsholm.dk/en/articles/spotteren-ikke-kranen/).