Ethical Considerations for the Military Use of Artificial Intelligence in Visual Reconnaissance

Achim Kuwertz-Wolf; Almuth Mueller; Arne Schumann; Fabian Jeschke; Mathias Anneken; Michael Teutsch; Nadia Burkart

REVIEW 2 major objections 7 minor 25 references

Ethical Considerations for the Military Use of Artificial Intelligence in Visual Reconnaissance

T0 review · 2 major / 7 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Military AI decision support can stay ethically and legally compliant if it is built around five consolidated principles—traceability, proportionality, governability, responsibility, and reliability—implemented through explainable AI and…

desk verdict Useful industry white paper mapping ethical principles to concrete ISR interface designs, but the load-bearing XAI faithfulness claim is unvalidated. read the letter →

arxiv 2502.03376 v1 pith:55FYZGBV submitted 2025-02-05 cs.CY cs.CV

Mathias Anneken , Nadia Burkart , Fabian Jeschke , Achim Kuwertz-Wolf , Almuth Mueller , Arne Schumann , Michael Teutsch This is my paper · ORCID

classification cs.CYcs.CV

keywords ethicalAImilitaryvisualreconnaissanceeXplainablehuman-in-the-loopdecisionsupportRulesofEngagementtraceability

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that military AI for visual reconnaissance can be ethically and legally compliant if it is built around five consolidated principles: traceability, proportionality, governability, responsibility, and reliability. The authors review existing civilian and military AI ethics guidelines, then translate these principles into concrete design choices for three decision-support systems—maritime surveillance, camp protection, and urban reconnaissance. Their central claim is that explainable AI (XAI) plus mandatory human confirmation, logging, and rule-based alignment with Rules of Engagement lets a human operator remain the final and sole decision-maker while still benefiting from automation. The paper is a design template: it shows how each principle maps to specific interface features, such as confidence warnings, heatmap overlays, and click-to-confirm audit trails. A sympathetic reader would take it as a proposal for how responsible military AI should be built, not as an empirical proof that the approach works.

What carries the argument

The load-bearing mechanism is the pairing of a consolidated five-principle ethics framework with XAI-driven human-in-the-loop interaction. The framework—traceability, proportionality, governability, responsibility, reliability—serves as a checklist that maps to concrete UI and system features: saliency heatmaps (e.g., RISE) that show which image regions drove a classification, feature-attribution bar plots (e.g., SHAP) for trajectory threat scores, confidence and ambiguity warnings, click-to-confirm logging, rule engines or knowledge bases that encode Rules of Engagement and Standard Operating Procedures, and an optional chat-style assistant using retrieval-augmented generation to cite applicable rules. Together these components are meant to make the AI's reasoning legible enough that a trained operator can spot erroneous evidence (such as a harpoon driving a 'military' label), veto or correct the recommendation, and remain accountable for the final decision.

What would settle it

Take the maritime case: recompute the classification after masking or removing the image region the heatmap highlights (the harpoon); if the military-class confidence does not drop substantially, the explanation does not reflect the model's reasoning, and the claimed operator correction would not work. Alternatively, run a controlled operator study where participants use the described GUI on ambiguous vessels and measure whether heatmaps improve correct identification of the 'erroneous feature'—if accuracy is no better than confidence scores alone, the traceability claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the abstract ethical vocabulary of military AI—transparency, accountability, human control—can be operationalized as a small set of engineering requirements. The five consolidated principles are taken from existing ethical guidelines (civilian ethics guidelines, military ethics principles, and just-war criteria) and then attached to observable system behaviour: traceability through per-layer XAI, enforced user acknowledgement, and logging; proportionality through recommendations weighed against mission rules and the current situation; governability through the operator's ability to override or deactivate the system; responsibility through keeping the human in the loop as the accountable decision-maker; reliability through confidence scores, uncertainty detection, and transparency about low-confidence results. The three use cases demonstrate the pattern: whenever the AI is uncertain, it says so, shows why via heatmaps or feature attributions, asks the operator to confirm, and logs the interaction. The authors conclude that carefully designed AI assistance can support the human user substantially and remain well-aligned with ethical principles, with the human in the loop as the final and sole decision-maker.

Load-bearing premise

The load-bearing premise is that saliency heatmaps reveal the model's actual decision-relevant evidence, so an operator who sees a harpoon highlighted can correctly conclude that the AI relied on it—if heatmaps are misleading or operators cannot interpret them, the traceability, responsibility, and governability arguments lose their footing.

Editorial extensions

If this is right

If such a system logs every user confirmation and interaction, a mission can be retrospectively audited to establish who knew what and when, which is what traceability and responsibility require in practice.
If the AI is designed to report its own uncertainty—low, close confidence scores or ambiguous behaviour—rather than force a decision, the operator is alerted exactly when human judgment is needed.
If countermeasure recommendations are generated from an ontology of SOPs and RoE, proportionality becomes a checkable property of the system rather than a vague aspiration.
If the operator can override or postpone any AI recommendation, governability holds even when the system's threat assessment escalates.
The authors' conclusion is that this combination of XAI, forced user interaction, and logging is sufficient for ethically compliant military decision support, with the human as final decision-maker.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The framework's ethical promise rests on an unvalidated premise: that heatmap explanations faithfully reflect the model's true decision evidence. A simple test—removing the highlighted region and checking whether the prediction changes—would tell whether the 'harpoon' explanation is genuine or merely plausible.
The same design pattern could be evaluated empirically as an operator-performance question: do officers with heatmap-guided explanations make better or faster decisions than those with confidence scores alone? The paper does not run this study.
The chat-assistant use case introduces retrieval-augmented generation, whose output quality depends on external documents; the paper acknowledges that outdated or biased retrieved information could yield inappropriate suggestions, which is a concrete failure mode for the reliability principle.
The five principles could serve as a requirements checklist for procurement and certification of military AI, potentially turning ethical review from narrative assessment into auditable system tests.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know this is a white paper, not a research preprint. It reads like a well-crafted industry position statement from HENSOLDT/Fraunhofer, and it does something genuinely useful: it takes the standard ethical AI principles (DoD, NATO, Taddeo) and shows what they look like at the level of a screen layout for military visual reconnaissance. The five principles it consolidates—traceability, proportionality, governability, responsibility, reliability—are not new, but the mapping onto concrete GUIs is a useful design checklist for people actually building these systems.

The literature synthesis is accurate and the paper is honest about its own status. I liked that it grounds each principle in the use cases and shows how logging, click-through confirmations, and RoE/SOP-aware suggestions might be implemented.

The soft spot is exactly where the stress-test note lands. Section 3.2 claims the operator can see from RISE heatmaps that the harpoon drove the military classification and can confidently resolve the ambiguity. That premise—saliency maps are faithful enough for an operator to act on—is load-bearing for the traceability, responsibility, and governability arguments. It is not demonstrated. There is no faithfulness evaluation, no deletion/insertion test, no user study, no comparison against ground-truth cause. Published work has shown saliency maps can be fragile. If the heatmap is misleading, the operator is not enabled but misled. That's a substantive gap for a paper whose conclusion is that XAI plus human-in-the-loop achieves ethical compliance.

A second, smaller gap is the assumption that human-in-the-loop itself guarantees ethical decisions. HITL can produce automation bias or over-reliance; the paper does not address that. The use cases are explicitly schematic, so this is not fatal for a white paper, but it means the ethical claims are plausible, not proven.

Minor issues: a missing citation where 'Transformer architecture [ ]' appears in Section 3.2, some typos (e.g., 'RA ' for RAG), and the Jane's database name is corrupted. These are cosmetic.

Who should read it: military AI developers, procurement officers, ethics advisory boards. It is a good prompt for thinking through interface-level ethics. It does not deserve formal peer review as a research paper because there is no empirical or formal contribution to referee. I would handle it as a non-archival industry white paper. If the authors want it to become a research result, they need to add a user study and a faithfulness evaluation of the XAI method. As it stands, read it, cite it if you work in this space, but don't treat its causal claims as established.

Referee Report

2 major / 7 minor

Summary. This white paper argues that responsible military use of AI in visual reconnaissance can be achieved by consolidating existing ethical guidance into five principles — traceability, proportionality, governability, responsibility, and reliability — and applying them to AI-powered decision-support systems that integrate eXplainable AI (XAI) and human-in-the-loop oversight. The paper reviews civilian guidelines (FATE, Jobin et al.) and military frameworks (U.S. DoD, NATO, Taddeo et al. for the U.K. DSTL), derives the consolidated principles in Section 2.3, and maps them onto three use cases: maritime surveillance from a submarine optronic mast (Section 3), military camp protection using wide-area motion imagery and PTZ cameras (Section 4), and land-based reconnaissance in an inhabited area supported by a chat-like AI assistant using knowledge models and retrieval-augmented generation (Section 5). The central claim, stated in the introduction and conclusion, is that careful design combined with XAI and human oversight can align military AI decision support with ethical and legal guidelines while keeping the human-in-the-loop as the final decision-maker. The paper is explicitly framed as a white paper with schematic GUIs and makes no claim of a full empirical evaluation.

Significance. If the central claim holds, the paper provides a concrete industrial design template: it translates abstract ethical principles into specific interface and process features (confidence warnings, heatmap overlays, mandatory click-through acknowledgments, audit logging, RoE/SOP-based recommendations). This concreteness is the main strength and goes beyond purely theoretical treatments of military AI ethics, and the literature synthesis in Sections 2.1–2.2 is accurate and correctly anchored in external authorities rather than derived from the authors' own prior work. The technical specifications are detailed enough to be testable: Section 3.2's claim that RISE heatmaps reveal the harpoon as the cause of the military-class score is a falsifiable empirical prediction, and Section 5.2 candidly lists limitations of retrieval-augmented generation (data quality, filtering). The significance is bounded, however, by two factors: the five principles are a synthesis rather than a genuinely new framework, and the load-bearing empirical premise that operators can reliably read saliency maps to detect and correct model error is asserted rather than demonstrated.

major comments (2)

[Section 3.2 (Fig. 3) and Section 3.3] The traceability analysis rests on an unvalidated empirical premise about the faithfulness of RISE heatmaps. Section 3.2 claims that 'the operator can discern that the harpoon on the whaling ship led to the suspiciously high score for the military class' and can 'confidently resolve the situation,' and Section 3.3 then concludes that traceability is 'achieved by introducing XAI to each layer,' with heatmaps showing 'the regions of interest in the image leading to the system's decision.' The paper offers no evidence for either component of this premise: (1) that the RISE saliency map is a faithful attribution of the classifier's output (no insertion/deletion evaluation, no comparison against known decision-relevant object parts, no model-randomization sanity check), and (2) that a human operator can correctly interpret the map and act on it (no user study, and the example is a single handpicked sample). The same unvalidated assumption applies to the SHAP attributions in Section 4.2, where no evaluation of attribution correctness is reported. Because published work shows that saliency maps can be insensitive to model parameters and can highlight artifacts rather than true decision evidence, an unfaithful heatmap would actively mislead the operator, which would collapse the governability argument ('as a prerequisite, the system must be traceable for a human user to understand the AI results and their consequences') and the responsibility argument ('the use of XAI methods and visualization enables the human user to understand the support provided by the AI system'). The authors should either soften the demonstrative wording, add explicit validation criteria (faithfulness metrics against ground-truth part annotations and an operator-comprehension test), or reframe the passages as design intentions rather than demonstrated system behavior.
[Section 2.3 (Governability), Section 5.3, Section 6] The human-in-the-loop argument equates procedural override capability with meaningful human control. Governability is defined in Section 2.3 as the ability of 'a human user to override an AI-based decision at any point or completely deactivate the respective system,' and the conclusion asserts that 'the human-in-the-loop shall still be the final and sole decision-maker.' Yet in the flagship scenario of Section 5.3 the AI assistant asserts 'enemy group now eligible target' and the operator 'verifies this and approves the target' (see Figure 7), a workflow in which the AI performs the normative assessment and the human role is reduced to confirmation. The paper does not address the known failure modes of human oversight — automation bias, vigilance decrement under time pressure, and deference to a system that states conclusions with high confidence — which are central to the meaningful-human-control literature that the paper itself cites through Taddeo et al. [23]. The central claim that careful design plus XAI can keep military AI decision support ethically aligned therefore requires either an engagement with the conditions under which override is actually exercised, or an explicit limitation stating that the framework presupposes well-trained, non-deferential operators and makes no claim about automation-bias effects.

minor comments (7)

[Section 3.2] The citation after 'a modern Transformer architecture' is an empty bracket '[ ]'; the intended reference appears to be [6] (Dosovitskiy et al.).
[Throughout] The manuscript contains many OCR-style transcription errors: missing initial letters (e.g., ' owever,' ' urthermore,' ' ence,' ' RA '), 'Jane's ighting Ship Database' missing the F in 'Fighting,' 'ethically unresponsive harm' for 'irresponsible,' 'severance of the choice,' and 'immanent' for 'imminent.' A full proofreading pass is needed before publication.
[Section 2.3 (Reliability)] Folding the DoD principle 'equitable' into reliability as a matter of training-data bias conflates fairness with technical robustness; the paper should justify this mapping or discuss fairness separately, since an unbiased but unreliable system and a biased but reliable one are distinct failure modes.
[Section 3.2 / Figure 3] Since the classifier's top classes are 'training ship' and 'patrol boat' — neither of which is a fishing boat — the operator's identification of the vessel as a fishing boat is a human semantic inference from the harpoon feature, not a model output; the text should make this distinction explicit.
[References] Reference [2] lists the year as 2012 for a DAIO Study with volume number 23—12, which implies 2023; please verify and correct.
[Section 4.3] The subsection header 'Responsibility.' uses a period instead of a colon, breaking the formatting pattern of the adjacent subsections.
[Section 6] The conclusion lists the domains as 'sea, air, and land,' but Use Case 2 is a land-based camp-protection scenario whose sensor happens to be airborne; this should be phrased as sensor domain rather than operating domain.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the ethical framework is sourced from external authorities and the use cases are illustrative; self-citations are confined to technical feasibility and are not load-bearing.

full rationale

This white paper makes no quantitative predictions and contains no fitted parameters, so the fitted-input and self-definitional circularity patterns do not apply. The consolidated principles in Section 2.3 are explicitly derived from external sources (Jobin et al. [10], DoD [4], NATO [15], Taddeo et al. [23]) rather than from the authors' own outputs. The three use cases are schematic scenarios designed to illustrate how the principles could be implemented; they are not independent empirical tests, and the paper does not claim they validate the principles. The only self-references appear in technical feasibility discussions (e.g., maritime detection benchmarks [14], WAMI tracking [19], WAMI moving-object detection [20], and knowledge-based reasoning [12]) and are not used to justify the ethical framework or the conclusion. The unvalidated assumption that RISE heatmaps faithfully reveal the classifier's decision evidence is a substantive correctness and validation gap, but it is not circularity: traceability is not defined in terms of RISE, and the heatmap demonstration is an asserted capability rather than a result derived from the paper's own principles. Accordingly, no step in the paper's derivation chain reduces to its inputs by construction.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The framework has no fitted parameters and introduces no entities. The central argument rests on four domain assumptions: human oversight is sufficient for ethical decisions, XAI saliency maps faithfully reflect model reasoning, confidence scores indicate uncertainty, and ROE/SOPs can be reliably encoded. The paper itself flags the fragility of the last assumption in Section 5.2.

assumptions (4)

domain assumption Human-in-the-loop oversight renders military AI ethically acceptable
Repeatedly stated, e.g., Section 6: 'the human-in-the-loop shall still be the final and sole decision-maker'; no evidence on automation bias or accountability.
domain assumption Saliency explanations faithfully reveal decision-relevant features
Section 3.2 assumes the RISE heatmap showing the harpoon reveals why the model gave a military label; no faithfulness validation.
domain assumption Low confidence across classes is a valid uncertainty signal
Section 3.3 treats similarly low scores as the system 'detecting and reporting its own uncertainty'; calibration is not established.
domain assumption Rules of Engagement and SOPs can be encoded in knowledge models or RAG
Use Case 3 depends on this; Section 5.2 concedes knowledge models are hard to generalize and RAG can retrieve biased or incorrect information.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Ethical Considerations for the Military Use of Artificial Intelligence in Visual Reconnaissance." pith.science (2026). https://pith.science/paper/55FYZGBV

@misc{pith2026250203376,
  author       = {Pith},
  title        = {Pith review of: Ethical Considerations for the Military Use of Artificial Intelligence in Visual Reconnaissance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/55FYZGBV}},
  note         = {Machine review of arXiv:2502.03376}
}

read the original abstract

This white paper underscores the critical importance of responsibly deploying Artificial Intelligence (AI) in military contexts, emphasizing a commitment to ethical and legal standards. The evolving role of AI in the military goes beyond mere technical applications, necessitating a framework grounded in ethical principles. The discussion within the paper delves into ethical AI principles, particularly focusing on the Fairness, Accountability, Transparency, and Ethics (FATE) guidelines. Noteworthy considerations encompass transparency, justice, non-maleficence, and responsibility. Importantly, the paper extends its examination to military-specific ethical considerations, drawing insights from the Just War theory and principles established by prominent entities. In addition to the identified principles, the paper introduces further ethical considerations specifically tailored for military AI applications. These include traceability, proportionality, governability, responsibility, and reliability. The application of these ethical principles is discussed on the basis of three use cases in the domains of sea, air, and land. Methods of automated sensor data analysis, eXplainable AI (XAI), and intuitive user experience are utilized to specify the use cases close to real-world scenarios. This comprehensive approach to ethical considerations in military AI reflects a commitment to aligning technological advancements with established ethical frameworks. It recognizes the need for a balance between leveraging AI's potential benefits in military operations while upholding moral and legal standards. The inclusion of these ethical principles serves as a foundation for responsible and accountable use of AI in the complex and dynamic landscape of military scenarios.

Figures

Figures reproduced from arXiv: 2502.03376 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 22 canonical work pages

[23]

Ethical Principles for Artificial Intelligence in National Defence

Mariarosaria Taddeo, David McNeish, Alexander Blanchard, and Elizabeth Edgar. Ethical Principles for Artificial Intelligence in National Defence. Philosophy & Technology, 34(4):1707–1729, December 2021

work page 2021
[1]

Autonomous weapon systems and jus ad bellum

Alexander Blanchard and Mariarosaria Taddeo. Autonomous weapon systems and jus ad bellum. AI & SOCIETY, March 2022

work page 2022
[2]

Master and servant - defense ai in Germany

Heiko Borchert, Torben Schütz, and Joseph Verbovszky. Master and servant - defense ai in Germany. In DAIO Study, volume 23—12. Defense AI Observatory, 2012

work page 2012
[3]

Trajectory-based pattern of life analysis

Hua-mei Chen, Erik Blasch, Nichole Sullivan, and Genshe Chen. Trajectory-based pattern of life analysis. In Proceedings of IEEE International Conference on Image Processing (ICIP), pages 2591–2595, Bordeaux, France, 2022

work page 2022
[4]

Responsible Artificial Intelligence Strategy and Implementation Pathway

DoD Responsible AI Working Council. Responsible Artificial Intelligence Strategy and Implementation Pathway. https://media.defense.gov/2022/Jun/22/2003022604/-1/-1/0/Department-of-Defense- Responsible-Artificial-Intelligence-Strategy-and-Implementation-Pathway.PDF, 2022

work page 2022
[5]

AI Principles: Recommendations on the Ethical Use of Artificial Intelligence by the Department of Defense

US Defense Innovation Board (DIB). AI Principles: Recommendations on the Ethical Use of Artificial Intelligence by the Department of Defense. https://media.defense.gov/2019/Oct/31/2002204458/-1/- 1/0/DIB_AI_PRINCIPLES_PRIMARY_DOCUMENT.PDF, 2020

work page 2019
[6]

An image is worth 16x16 words: Transformers for image recognition at scale

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

arXiv 2010
[7]

Rise of Artificial Intelligence in Military Weapons Systems, 2020

Fraunhofer Group for Defense and Security VVS. Rise of Artificial Intelligence in Military Weapons Systems, 2020

work page 2020

Show all 25 references

[8]

Retrieval-Augmented Generation for Large Language Models: A Survey

Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Meng Wang, and Haofen Wang. Retrieval-Augmented Generation for Large Language Models: A Survey. In arXiv:2202.01110, 2023

2023 arXiv
[9]

For the sake of privacy: Skeleton-based salient behavior recognition

Thomas Golda, Johanna Thiemich, Mickael Cormier, and Jürgen Beyerer. For the sake of privacy: Skeleton-based salient behavior recognition. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3983–3987. IEEE, 2022

2022
[10]

Artificial Intelligence: the global landscape of ethics guidelines

Anna Jobin, Marcello Ienca, and Effy Vayena. Artificial Intelligence: the global landscape of ethics guidelines. Nature Machine Intelligence, 1(9):389–399, September 2019. arXiv:1906.11668 [cs]

2019 arXiv
[11]

A transformer-based late-fusion mechanism for fine- grained object recognition in videos

Jannik Koch, Stefan Wolf, and Jürgen Beyerer. A transformer-based late-fusion mechanism for fine- grained object recognition in videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 100–109, 2023

2023
[12]

Applying Knowledge-Based Reasoning for Information Fusion in Intelligence, Surveillance, and Reconnaissance

Achim Kuwertz, Dirk Mühlenberg, Jennifer Sander, and Wilmuth Müller. Applying Knowledge-Based Reasoning for Information Fusion in Intelligence, Surveillance, and Reconnaissance. In Multisensor Fusion and Integration in the Wake of Big Data, Deep Learning and Cyber Physical Sys...

2018
[13]

A unified approach to interpreting model predictions

Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems 30, pages 4765–4774. Curran Associate...

2017
[14]

A benchmark for deep learning based object detection in maritime environments

Sebastian Moosbauer, Daniel König, Jens Jäkel, and Michael Teutsch. A benchmark for deep learning based object detection in maritime environments. In Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Long Beach, CA, USA, 2019

2019
[15]

Summary of the NATO Artificial Intelligence Strategy

North Atlantic Treaty Organization (NATO). Summary of the NATO Artificial Intelligence Strategy
[16]

RISE: Randomized input sampling for explanation of black- box models

Vitali Petsiuk, Abir Das, and Kate Saenko. RISE: Randomized input sampling for explanation of black- box models. In British Machine Vision Conference (BMVC), 2018

2018
[17]

Video processing from electro-optical sensors for object detection and tracking in a maritime environment: A survey

Dilip K Prasad, Deepu Rajan, Lily Rachmawati, Eshan Rajabally, and Chai Quek. Video processing from electro-optical sensors for object detection and tracking in a maritime environment: A survey. IEEE Transactions on Intelligent Transportation Systems, 18(8):1993–2016, 2017

1993
[18]

Grad-cam: Visual explanations from deep networks via gradient-based localization

Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618–626, 2017

2017
[19]

Appearance and motion based persistent multiple object tracking in wide area motion imagery

Lars Sommer, Wolfgang Krüger, and Michael Teutsch. Appearance and motion based persistent multiple object tracking in wide area motion imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3878–3888, 2021

2021
[20]

A survey on moving object detection for wide area motion imagery

Lars Sommer, Michael Teutsch, Tobias Schuchert, and Jürgen Beyerer. A survey on moving object detection for wide area motion imagery. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2016

2016
[21]

A comprehensive analysis of modern object detection methods for maritime vessel detection

Raphael Spraul, Lars Sommer, and Arne Schumann. A comprehensive analysis of modern object detection methods for maritime vessel detection. In Artificial Intelligence and Machine Learning in Defense Applications II, volume 11543, pages 13–24. SPIE, 2020

2020
[22]

Accepting Moral Responsibility for the Actions of Autonomous Weapons Systems—a Moral Gambit

Mariarosaria Taddeo and Alexander Blanchard. Accepting Moral Responsibility for the Actions of Autonomous Weapons Systems—a Moral Gambit. Philosophy & Technology, 35(3):78, August 2022

2022
[24]

FCOS: Fully Convolutional One-Stage Object Detection

Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: Fully Convolutional One-Stage Object Detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627– 9636, 2019

2019
[25]

ATR Algorithm Development Image Database

Defense Systems Information Analysis Center (DSIAC). ATR Algorithm Development Image Database. https://www.dsiac.org/resources/available-databases/atr-algorithm-development-image-database

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [23]

Ethical Principles for Artificial Intelligence in National Defence

Mariarosaria Taddeo, David McNeish, Alexander Blanchard, and Elizabeth Edgar. Ethical Principles for Artificial Intelligence in National Defence. Philosophy & Technology, 34(4):1707–1729, December 2021

work page 2021

[2] [1]

Autonomous weapon systems and jus ad bellum

Alexander Blanchard and Mariarosaria Taddeo. Autonomous weapon systems and jus ad bellum. AI & SOCIETY, March 2022

work page 2022

[3] [2]

Master and servant - defense ai in Germany

Heiko Borchert, Torben Schütz, and Joseph Verbovszky. Master and servant - defense ai in Germany. In DAIO Study, volume 23—12. Defense AI Observatory, 2012

work page 2012

[4] [3]

Trajectory-based pattern of life analysis

Hua-mei Chen, Erik Blasch, Nichole Sullivan, and Genshe Chen. Trajectory-based pattern of life analysis. In Proceedings of IEEE International Conference on Image Processing (ICIP), pages 2591–2595, Bordeaux, France, 2022

work page 2022

[5] [4]

Responsible Artificial Intelligence Strategy and Implementation Pathway

DoD Responsible AI Working Council. Responsible Artificial Intelligence Strategy and Implementation Pathway. https://media.defense.gov/2022/Jun/22/2003022604/-1/-1/0/Department-of-Defense- Responsible-Artificial-Intelligence-Strategy-and-Implementation-Pathway.PDF, 2022

work page 2022

[6] [5]

AI Principles: Recommendations on the Ethical Use of Artificial Intelligence by the Department of Defense

US Defense Innovation Board (DIB). AI Principles: Recommendations on the Ethical Use of Artificial Intelligence by the Department of Defense. https://media.defense.gov/2019/Oct/31/2002204458/-1/- 1/0/DIB_AI_PRINCIPLES_PRIMARY_DOCUMENT.PDF, 2020

work page 2019

[7] [6]

An image is worth 16x16 words: Transformers for image recognition at scale

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

arXiv 2010

[8] [7]

Rise of Artificial Intelligence in Military Weapons Systems, 2020

Fraunhofer Group for Defense and Security VVS. Rise of Artificial Intelligence in Military Weapons Systems, 2020

work page 2020