The more technology distances the actor, the more responsibility must increase.
Contact us: info@strategic-culture.su
From person to target
What becomes of humanity when technology becomes capable of recognising, classifying, tracking and ultimately killing human beings without those carrying out the strikes ever having seen them? My friend Mohammad Reza Dehshiri asked this very question, and his answer is a powerful one. The danger, he writes, is not artificial intelligence but artificial inhumanity – that is, the possibility that machines will become more accurate precisely as humans cease to feel responsible for what those machines do in their name.
With this article, I am launching a series that takes Dehshiri’s analysis as its starting point and attempts to take it further. If the problem is ‘artificial inhumanity’, the answer cannot lie in a ‘more ethical’ AI, as if ethics were a module to be installed as an afterthought to a system that has already been designed. We need to reclaim a category that technological jargon has rendered almost unmentionable: human intelligence, understood not as a remnant to be protected amidst the complexities of automation, but as a form of knowledge with its own distinct characteristics, which no increase in computing power can replace.
I shall call it Human Intelligence, and the ambiguity is deliberate. In intelligence service jargon, HUMINT refers to the oldest of the intelligence disciplines: that which relies on people, sources, languages and the patient frequenting of places. In philosophical terminology, the expression denotes the faculty of understanding. The thesis of this first contribution is that these two meanings now converge, and that their convergence has geopolitical implications that extend far beyond the debate on the ethics of weaponry.
We can describe the operational chain of contemporary warfare in seven stages: detect, identify, classify, predict, prioritise, target, strike. Each link can be technically refined without the sequence, taken as a whole, being morally refined. The resulting phenomenon is called the ‘datafication of human life’: the individual reduced to an identity, a location, a behavioural pattern, a threat score.
This is no abstraction. In April 2024, an investigation by Yuval Abraham for +972 Magazine and Local Call described ‘Lavender’, a system used by the Israeli armed forces in Gaza to assign tens of thousands of people a probability of affiliation with armed groups. According to internal sources cited in the investigation, the human verification of each name often amounted to no more than twenty seconds – the time needed to ascertain that the target was male. The Israeli Defence Forces have disputed this account and described Lavender as a database designed to support analysts. Even if one accepts the official version, however, the crux of the matter remains: a human operator was present, and their presence was not sufficient to transform the system’s recommendation into a judgement.
Cognitive psychology has a name for this phenomenon, automation bias: the tendency of those working with an automated system to treat its recommendations as a substitute for their own verification, and to fail to seek out the information that might contradict them. Back in 1976, Joseph Weizenbaum – who had built one of the first conversational programmes in history – already distinguished between the ability to calculate and the ability to judge, and argued that there are tasks that should not be entrusted to a computer even if the computer were capable of performing them. Half a century on, this distinction has become even more pressing. Computation works on what has already been translated into data. Judgement must decide whether that translation has been faithful.
Dehshiri drives the point home with four distinctions: technical precision is not ethical precision; accuracy is not justice; automation is not responsibility; technological superiority is not moral superiority. I would add a fifth, of an epistemic nature, upon which the others depend: correlation is not understanding. A machine-learning system recognises statistical patterns in a dataset. It does not know what a school is, because it does not know what a child is, and it does not even know that it does not know this.
L’iraniano cites an interesting case study: On 28 February 2026, in the early hours of the war, the Shajareh Tayyebeh primary school in Minab, in Hormozgan province, was struck several times during school hours. The figures provided by the Iranian authorities range between 150 and 175 deaths, the vast majority of whom were girls, along with dozens of teachers and a few parents; the district prosecutor put the final death toll at 156. Washington has never directly acknowledged its responsibility nor published the findings of the Pentagon’s investigation. Amnesty International has described the incident, at best, as a serious intelligence failure.
There is one detail, revealed by satellite imagery, which shifts the analysis from the moral to the factual plane. According to Western media, the school stood next to a naval base belonging to the Revolutionary Guards and had previously formed part of it; among the explanations circulating in US circles is that of out-of-date targeting data. If this hypothesis were confirmed, Minab would not be a case of a machine making a mistake, but of a database that had ceased to correspond to reality. The building had changed. The data had not.
No classification algorithm, however sophisticated, can detect that an old military building has become a school if no one has recorded the change. Everyone living in Minab knew this: the parents who took their daughters to school every morning, the schoolbus driver, the teachers. This is precisely the kind of knowledge that HUMINT gathers and that remote warfare tends to regard as superfluous, because it is slow, costly and difficult to scale. Based on publicly available sources, US responsibility appears highly probable, whilst the exact circumstances remain unclear and confidence on this point is low: a database error, a classification error, or a conscious acceptance of the risk. The three hypotheses, however, all converge on the same question. Who, prior to the strike, checked whether there were people in that building?
Beyond ‘human-in-the-loop’
The international debate has long been content with the principle of human-in-the-loop, whereby a human being is part of the decision-making loop.
This formula is insufficient: a person may be present in the loop yet merely rubber-stamp the algorithm’s recommendation. He proposes replacing it with a more demanding principle, which he calls humanity-in-command: the effective ability to question, reject, override and halt the automated process.
The history of the Cold War offers two examples that clarify what this means in practice. On 26 September 1983, the Soviet early-warning system Oko signalled the launch of five ballistic missiles from the United States. The officer on duty, Stanislav Petrov, dismissed the alarm as a false alarm and did not report it as an ongoing attack, reasoning that a first American strike would never consist of just five missiles. He was right: the satellite had mistaken the sun’s reflection off the clouds for missile launches. Twenty-one years earlier, on 27 October 1962, in the Caribbean Sea, the Soviet submarine B-59, isolated and under fire from signal depth charges, was on the verge of launching a nuclear-tipped torpedo. The launch required the consent of the three senior officers on board, and Vasily Arkhipov refused to give it.
In neither case did the man approve a recommendation from the system. He contradicted it. And he did not do so because he had more data: indeed, the data were the problem. He did so because he knew how to ask himself a question that no procedure required him to ask, namely whether what he was seeing made sense in the context of everything else he knew about the world. Petrov was familiar with American nuclear doctrine; Archipov knew the difference between an attack and a threat. It is this ability to compare the signal with the bigger picture, the detail with what is plausible, that I call Human Intelligence. It is worth noting that the Dena was sunk by a submarine. We do not know what its officers were asked, nor what scope they had for judgement.
This proposal needs to be clarified, as there is a risk of reducing it to a generic humanistic appeal. Human Intelligence, as I understand it, is structured across three distinct levels, each of which marks a structural – rather than contingent – limit of artificial intelligence.
The first level is epistemic. The scholastic tradition distinguished between intellectus, the act by which the mind immediately grasps a truth, and ratio, the discursive process of moving from one element to another: for Thomas Aquinas (Summa Theologiae, I, q. 79, a. 8), the two are not distinct faculties but aspects of a single intellect, in which reasoning begins with an understanding and returns to an understanding. Current machine learning systems realise a ratio of unprecedented power, capable of inferring, combining and predicting on scales inaccessible to the human mind. They lack the aspect of intellectus, that is, a grasp of the meaning of what they process. This is a philosophical thesis and must be presented as such, but it has precise practical consequences: a system devoid of understanding cannot know when its own model of the world has become false.
The second aspect is practical. Prudentia, in the Thomist definition, is right reason applied to action, and its specific task consists in applying the universal principle to the particular case. The machine generalises: its value lies in mapping the individual case back to a class. Prudence does the opposite, because it must recognise what, in this particular case, falls outside the class. War, which is the realm of the singular and the unforeseen, is the arena where this difference carries the most weight. The Vatican’s Antiqua et nova note of January 2025 took up this line of thought, emphasising the embodied and relational nature of human intelligence and its irreducibility to a mere processing function.
The third level is relational, and it is here that philosophical meaning meets that of the intelligence services. Sherman Kent, who effectively founded the field of intelligence studies in the post-war United States, defined it first and foremost as knowledge, even before organising or activity. HUMINT is the form of this knowledge that passes through the other as a person: someone who speaks their language, knows their customs, can read a neighbourhood, and can distinguish a festival from a rally. Over the last three decades, the major powers have progressively shifted resources towards technical disciplines, from signals intelligence to satellite surveillance, right through to data fusion platforms. Minab illustrates the cost of this imbalance.
None of this implies a rejection of technology. Human Intelligence does not stand in opposition to AI in the same way that a manual stands in opposition to a mechanic. It establishes a hierarchy: the algorithm as a tool subservient to a judgement it cannot produce on its own. And from this it derives a principle, formulated by Dehshiri, which deserves to become an operational criterion: the more technology distances the actor from the consequences of their own actions, the more the conscious responsibility of the decision-maker must increase, not decrease.
Geopolitically speaking
So far, the issue might seem to be one of moral philosophy. It is not merely that. The distribution of algorithmic warfare capability is a distribution of power, and the rules governing it will be written – if they are written at all – by those with an interest in limiting them as little as possible.
The state of international law confirms this. The group of government experts established under the Convention on Certain Conventional Weapons has been discussing lethal autonomous weapons systems since 2017 without having reached a binding instrument. In November 2023, the United States promoted a political declaration on the responsible military use of AI, which is by definition non-binding; in December of the same year, the United Nations General Assembly adopted the first resolution dedicated to these systems, whilst the International Committee of the Red Cross has been calling since 2021 for a ban on unpredictable autonomous systems and those designed to target human beings. In June 2024, speaking at the G7 summit in Borgo Egnazia, Pope Francis called for no machine ever to be placed in a position to decide on a person’s life. The gap between these positions and the reality on the battlefields of 2026 is measured in casualties.
There is also a deeper asymmetry. Those who control the data decide who is a target, and the victims of algorithmic warfare are largely found outside the West: in Gaza, Iran, the Sahel and Yemen. For states that do not produce the platforms or control their databases, human intelligence thus becomes a matter of sovereignty. A country that maintains linguistic expertise, regional knowledge and human intelligence networks retains its autonomy of judgement; a country that entrusts its worldview to systems designed elsewhere – often by private companies – surrenders it along with its data. In an order tending towards multipolarity, this distinction will separate the actors who judge from those who are judged.
This leads to a concrete proposal, which I shall develop in future articles. Alongside the general principle of meaningful human oversight, international humanitarian law could codify a specific obligation for human verification of the civil status of every fixed infrastructure included on a list of targets, with an expiry date beyond which the information becomes invalid and must be renewed using direct sources. This is a minimum standard. Applied to Minab, it would have required someone to ask whether that building was still what the records stated it to be.
Technology can determine how a weapon reaches its target, whilst it is up to humanity to decide whether that target should be struck. The question underlying all this is who, in practical terms, exercises that humanity, with what tools of knowledge and within which institutions. In Minab, someone should have known that there was a school there. The fact that nobody knew, or that nobody asked, is the point from which we must start again.
