The greater the technological separation from the actor, the heavier their responsibility becomes.
From person to target
What happens to humanity when technology can identify, categorize, monitor, and ultimately eliminate individuals without those executing the attacks ever seeing them firsthand? This question was posed by my friend Mohammad Reza Dehshiri in his thought-provoking article. He argues that the threat lies not in artificial intelligence itself but in what he terms artificial inhumanity: machines gaining greater precision precisely as human accountability for their actions diminishes.
This article launches a series building on Dehshiri’s insights, pushing the conversation forward. If the problem is ‘artificial inhumanity,’ then the solution cannot be simply a ‘more ethical’ AI, as if ethics were an add-on module to a pre-designed system. It is crucial to reclaim the concept of human intelligence—a category that technological language has rendered nearly unspeakable—not as a leftover to be preserved amid automation but as a unique kind of knowledge that no boost in computing can substitute.
I refer to it as Human Intelligence with deliberate ambiguity. Intelligence agencies use HUMINT to designate the oldest intelligence discipline, relying on people, sources, languages, and habitual presence in places. Philosophically, it refers to the capacity for understanding. The core argument of this piece is that these two meanings increasingly align, with geopolitical consequences extending far beyond weaponry ethics.
Modern warfare can be broken down into seven steps: detect, identify, classify, predict, prioritize, target, and strike. While each step can be technically improved, the moral dimension of the entire sequence need not improve in tandem. This dynamic results in the ‘datafication of human life’: transforming individuals into mere identities, locations, behaviors, and threat scores.
This is not a theoretical issue. In April 2024, Yuval Abraham’s investigation for +972 Magazine and Local Call exposed ‘Lavender,’ a program used by the Israeli military in Gaza that assessed tens of thousands for suspected links to armed groups. Internal sources reported human verification per name often lasted no longer than twenty seconds, usually just enough to confirm gender. The Israeli Defence Forces have denied these claims, describing Lavender as a database to aid analysts. Even if the official explanation is accepted, the critical issue remains: a human operator was involved, yet their presence failed to convert system recommendations into meaningful judgments.
Cognitive psychology labels this tendency automation bias: the inclination of operators to trust automated outputs as substitutes for personal verification and to overlook contradictory evidence. In 1976, Joseph Weizenbaum—creator of one of the first conversational programs—distinguished between calculating ability and judgment, warning that certain tasks should not be delegated to computers, even if they could perform them. Fifty years later, this division has become increasingly crucial. Computation operates on pre-translated data, while judgment must evaluate the accuracy of that translation.
Dehshiri reinforces this with four distinctions: technical precision differs from ethical precision; accuracy is not justice; automation does not equate to responsibility; technological dominance does not mean moral superiority. I would add a fifth, epistemic distinction that underpins the others: correlation does not equal understanding. Machine-learning systems identify statistical patterns but do not comprehend what a school is, what a child is, or even realize their own ignorance.
Dehshiri highlights a notable example: On 28 February 2026, during early war hours, Shajareh Tayyebeh primary school in Minab, Hormozgan province, was hit multiple times while classes were in session. Iranian authorities report between 150 and 175 fatalities, mostly girls, alongside teachers and some parents; the district prosecutor’s final figure was 156. Washington has neither admitted direct responsibility nor released the Pentagon’s investigation results. Amnesty International labeled the event as a severe intelligence failure at best.
Satellite imagery uncovers a critical factual detail, moving the discussion beyond morality. Western media noted the school was adjacent to a Revolutionary Guards naval base, previously part of the same facility. Among US explanations is outdated targeting data. If confirmed, Minab’s tragedy was not due to a machine error but a database that failed to reflect reality. The building had changed purpose, but the data had not.
No matter how advanced, classification algorithms cannot detect that a former military structure is now a school without updated records. Locals in Minab knew this: parents escorting their children, the schoolbus driver, teachers. This is exactly the type of knowledge HUMINT gathers and that distant warfare often deems unnecessary, due to its slowness, expense, and complexity. Public information hints strongly at US responsibility, though details remain murky—whether a database error, classification mistake, or deliberate risk acceptance. All hypotheses boil down to one question: Who verified the presence of people in that building before the strike?
Beyond ‘human-in-the-loop’
The international conversation has long settled for the human-in-the-loop principle, insisting that a human be involved in decision-making processes.
This standard falls short: someone could be in the loop merely to approve algorithmic outputs without critical scrutiny. Dehshiri suggests a stricter doctrine he names humanity-in-command: the genuine power to question, reject, override, and halt automated workflows.
Historical Cold War examples illustrate this well. On 26 September 1983, the Soviet Oko early-warning system falsely detected five US missile launches. Duty officer Stanislav Petrov judged this an error and refrained from escalating, reasoning that a true first strike wouldn’t comprise just five missiles. He was correct: the satellite had confused sunlight reflected off clouds with missile launches. Similarly, on 27 October 1962, the Soviet submarine B-59 was about to fire a nuclear torpedo when Vasily Arkhipov refused consent, averting a disaster.
In both instances, the individual contradicted system alerts rather than confirmed them—not because they had additional data but because they questioned the coherence of what they observed given their broader understanding. Petrov knew US nuclear strategies; Arkhipov distinguished between a genuine attack and a threat. This capacity to frame signals within the broader reality, to weigh detail against plausibility, is what I term Human Intelligence. Notably, the Dena was sunk by a submarine; we lack information on what orders its officers received or their freedom to exercise judgment.
This concept needs refinement to avoid being reduced to a vague humanistic notion. Human Intelligence, as defined here, operates on three fundamental levels, each highlighting a structural limitation of artificial intelligence rather than a contingent one.
The first is epistemic. The scholastic tradition differentiated intellectus, the immediate apprehension of truth, from ratio, the discursive reasoning process. Thomas Aquinas (Summa Theologiae, I, q. 79, a. 8) viewed them as facets of a single intellect, where reasoning starts and ends with understanding. Modern machine learning achieves unprecedented ratio, inferring and predicting beyond human scope. Yet, it lacks intellectus: genuine understanding of content. While philosophical in nature, this claim has practical consequences—a system without understanding cannot discern when its world model becomes inaccurate.
The second is practical. Thomist prudentia—right reason applied to action—involves interpreting general principles in particular situations. Machines generalize, classifying cases into categories, while prudence involves recognizing exceptions and anomalies. Warfare, characterized by uniqueness and unpredictability, exemplifies where this distinction matters most. The Vatican’s January 2025 Antiqua et nova note echoes this, underscoring the embodied, relational aspects of human intelligence that cannot be reduced to mere data processing.
The third level is relational, where philosophy meets intelligence practice. Sherman Kent—the founding figure of US intelligence studies—defined intelligence foremost as knowledge rather than organization or activity. HUMINT embodies this knowledge as it passes through people who know local languages, customs, contexts, and can differentiate between a festival and a demonstration. Over recent decades, major powers have increasingly prioritized technical methods like signals intelligence and satellite imagery, neglecting HUMINT. The Minab tragedy illustrates the consequences of this imbalance.
All this does not advocate rejecting technology. Human Intelligence is not opposed to AI as a manual craft is to a mechanic. Rather, it establishes a hierarchy: algorithms serve as tools subordinate to judgments they cannot independently form. From this arises a principle highlighted by Dehshiri that merits operational adoption: as technology further distances decision-makers from consequences, the weight of conscious responsibility must grow, never diminish.
Geopolitical dimensions
At first glance, this might appear a moral dilemma. However, the reality is more complex. The capability to wage algorithmic warfare translates into power distribution, and those writing the rules—if rules emerge—will likely seek the fewest limits on their freedom.
International law reflects this tension. Since 2017, the Convention on Certain Conventional Weapons’ expert group has debated lethal autonomous weapons without producing binding regulations. November 2023 saw the US promote a non-binding political declaration on responsible military AI usage; December brought the first UN resolution on these systems. Since 2021, the International Committee of the Red Cross has advocated banning unpredictable autonomous systems and those targeting humans. In June 2024, Pope Francis called at the G7 summit for never allowing machines to decide a person’s fate. The gulf between such positions and battlefield realities in 2026 is measured in lives lost.
There is also a deeper imbalance: those controlling data determine who is targeted, with the majority of algorithmic warfare victims located outside Western countries—in Gaza, Iran, the Sahel, and Yemen. For nations lacking platform production or database control, maintaining human intelligence becomes critical for sovereignty. States retaining linguistic skills, local knowledge, and HUMINT networks preserve their judgment autonomy; those relying on externally designed systems—often commercial—cede not only data but worldviews. In a shift toward multipolarity, this divide will separate judges from the judged.
This sets the stage for a concrete recommendation to be expanded upon later. In addition to broad standards for meaningful human oversight, international humanitarian law could mandate human verification of civilian status for every fixed target listed, complete with expiry dates requiring updates from direct sources. This baseline would have compelled verification of Minab’s building status before attack.
Technology can control how weapons reach targets, but humanity must determine whether a target should be engaged. The fundamental question is who exercises this humanity, with what knowledge tools, and within which institutions. In Minab, someone should have known the building was a school. The tragedy lies in no one knowing—or no one asking—and that is where our efforts must begin anew.
About the Author
