Algorithmic Targeting
Algorithmic targeting refers to the class of AI and machine-learning systems that identify potential military targets by pattern-matching over signal intelligence, communications metadata, movement data, and other surveillance streams. Systems of this kind have been deployed in various forms since the early 2000s and have accelerated dramatically with the capabilities of contemporary machine learning. Suchman's recent work has made algorithmic targeting one of the most consequential case studies in the critical analysis of AI, because the gap between generated plans and encountered situations — her foundational framework — has lethal consequences when outputs are accepted without adequate evaluation. Her analyses describe what she has called 'the algorithmically accelerated killing machine,' where the volume of target nominations overwhelms the capacity of humans in the loop to deliberate.
The Opacity Gradient Problem — Contrarian ^ Opus
There is a parallel reading that begins not with the output classifications but with the substrate question: what makes these systems possible at all? Algorithmic targeting depends on a surveillance architecture so comprehensive that pattern-matching becomes feasible. The infrastructure predates the AI layer — decades of signals intelligence collection, the post-9/11 expansion of metadata capture, the normalization of persistent monitoring. The critique of algorithmic targeting as a classification problem obscures the prior political economy: who built the collection apparatus, under what legal regime, with what public consent or oversight. The AI system is the tip; the mass surveillance is the iceberg.
Moreover, the frame of 'plans treated as actions' assumes the problem is one of misapplication — that better human oversight, slower tempo, more deliberation would restore the proper relationship. But the deeper reading is about institutional capture. Military and intelligence organizations adopt these systems not despite their tendency to overwhelm human judgment but because of it. Speed is the point. Volume is the feature. The systems deliver what the institutions want: operational tempo that competitors cannot match, plausible deniability for outcomes ('the algorithm decided'), and a veneer of precision that satisfies domestic political constraints on visible casualties. The gap between plan and action is not a bug introduced by rushed deployment; it is the design goal of organizations that have decided deliberation is friction to be eliminated. The lethality is not incidental but structural to the institutional logic.
In the AI Story
Algorithmic targeting systems operate on the structure of AI outputs as plans. The system processes surveillance data and produces a classification: this pattern of signals or movements corresponds to a valid military target. The classification is a plan — a proposal about what the data means, based on statistical patterns in training material. Whether the classification is correct for the specific open-world situation it addresses depends on situated knowledge of the specific network, geography, and actors — knowledge that human intelligence analysts have traditionally built through years of working specific theaters.
The systems collapse the time available for evaluation. Traditional intelligence work involved slow, judgment-intensive assessment of ambiguous signals in context. Algorithmic targeting produces classifications faster than human operators can deliberate on them. The pressure to act on the output — to 'prosecute' the target, in the military vocabulary — intensifies. The situated judgment that distinguishes a reliable pattern from a training-data artifact is bypassed. The plan is treated as an action. As Suchman put it in her 2025 AI Now interview, 'the possibilities for judgment, for deliberation, for assessing the validity of the data... basically disappear.'
The consequences have been documented in multiple conflict theaters. In Gaza, the Israeli military's use of systems with names like Lavender, Gospel, and Where's Daddy has been the subject of extensive reporting and legal analysis, with Suchman among the scholars most prominently engaged in the critique. The pattern: AI systems generate target nominations at rates that preclude careful evaluation; human reviewers approve the nominations as procedural checkpoints rather than substantive assessments; civilian casualties accumulate as the structural consequence of speed and automation rather than of individual malice.
Suchman's analysis connects algorithmic targeting to her broader framework. The systems are not autonomous in any meaningful sense — they are sociomaterial assemblages of hardware, software, training data, operational procedures, and institutional pressures. But the reification of 'AI' as the active agent conveniently distributes responsibility: the AI 'chose' the target, the operator only 'approved.' The gap between plan and action is crossed not by situated judgment but by procedural compliance. The lethality is the structural consequence of treating a plan as an action at a tempo that makes evaluation impossible.
Origin
Algorithmic targeting has a history stretching back to Cold War-era pattern-matching over signal intelligence and has accelerated with every advance in computational capability. Contemporary systems draw on the machine learning revolution of the 2010s and on the surveillance infrastructure built out in the post-9/11 decades.
Suchman's specific engagement with the topic has deepened over the past decade, producing a series of essays and interventions in Social Studies of Science, open letters on autonomous weapons, and her 2025 AI Now Institute interview. The work has been cited extensively in international legal and policy discussions of AI in warfare.
Key Ideas
Targeting is classification. Algorithmic targeting systems produce classifications — plans about what surveillance data means — not adjudications of what should actually happen.
Speed eliminates deliberation. When outputs accumulate faster than humans can evaluate them, deliberation becomes procedural approval rather than substantive assessment.
Situated judgment is bypassed. The intelligence analyst's years of domain knowledge about specific networks, actors, and contexts is precisely what the automated system does not have and cannot replicate.
Distributed responsibility. The reification of 'AI' as the decision-maker conveniently distributes accountability; the assemblage of training data, corporate decisions, operational procedures, and individual operators is where accountability actually lives.
The template is general. The structure — AI generates plans, humans approve under time pressure, plans become actions without situated evaluation — applies beyond warfare to medicine, law, finance, and any domain deploying AI at scale.
Appears in the Orange Pill Cycle
Tempo as Political Choice — Arbitrator ^ Opus
The classification point is completely right (100%): algorithmic targeting systems do produce plans, not adjudications, and the distinction matters enormously for understanding what the outputs are and what situated judgment would require. The contrarian infrastructure point is also entirely correct (100%): these systems rest on a surveillance substrate whose political economy and legal normalization are prior questions that the 'AI ethics' frame often elides. The right frame holds both: algorithmic targeting is simultaneously a technical classification problem and the legible surface of a much larger apparatus.
On the question of whether speed eliminates deliberation or whether speed is itself the institutional goal, the answer depends on which causal story you're telling. If the question is 'how do civilian casualties happen,' the entry's account is 80% right: outputs overwhelm human capacity, procedural approval replaces substantive assessment, and the gap between plan and action is crossed without the situated knowledge required. But if the question is 'why do institutions adopt these systems,' the contrarian view is 70% correct: tempo is not an accident of deployment but a strategic objective, and the distribution of responsibility ('the AI decided') is a feature of institutional design, not a misunderstanding to be corrected through better practices.
The synthesis the topic benefits from is this: tempo is a political choice that technical systems enable. Algorithmic targeting makes it possible to treat speed as if it were a neutral operational requirement rather than a decision about how much deliberation a democracy will tolerate before lethal force. The systems don't eliminate human judgment by accident — they are adopted by institutions that have decided judgment at scale is operationally unaffordable. The critique must address both layers: the technical gap between plan and action, and the prior political decision to value tempo over deliberation.
Further reading
- Lucy Suchman, 'Imaginaries of Omniscience: Automating Intelligence in the US Department of Defense' (Social Studies of Science, 2023)
- Lucy Suchman, 'The Uncontroversial "Thingness" of AI' (Big Data & Society, 2023)
- Yuval Abraham, '"Lavender": The AI Machine Directing Israel's Bombing Spree in Gaza' (+972 Magazine, 2024)
- International Committee of the Red Cross, 'Artificial Intelligence and Machine Learning in Armed Conflict' (2019)