Pith. sign in

REVIEW 4 major objections 4 minor 51 references

Structured AI Decision-Making in Disaster Management

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Structured five-Level decision framework with Enabler agents lets an RL agent beat both argmax benchmarking and experienced human responders on accuracy and consistency.

desk verdict The framework idea is worth a look, but the headline gains don't isolate the structure, and internal inconsistencies undermine the central claims. read the letter →

arxiv 2509.01576 v1 pith:CDZTT2SM submitted 2025-09-01 cs.AI cs.CYcs.SYeess.SY

Julian Gerald Dcruz , Argyrios Zolotas , Niall Ross Greenwood , Miguel Arana-Catania This is my paper
classification cs.AIcs.CYcs.SYeess.SY
keywords structureddecision-makingresponsibleAIdisastermanagementreinforcementlearningEnableragentsmultimodalclassificationhumanevaluationsafety-critical
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

The paper argues that autonomy in safety-critical settings needs structure, not just predictive insight, and demonstrates this with a five-Level decision Scenario for disaster management. Trained Enabler agents score incoming data at each Level, and a reinforcement-learning Decision Maker uses those scores to pick actions, with an option to gather more data at a small cost. Compared with a Benchmark that simply takes the Enabler's highest-confidence answer at every step, the RL agent is 60.94% more stable in making consistently correct decisions across Scenarios. Compared with 61 human victims, volunteers, and stakeholders who worked through the same Scenarios without Enabler insights, the RL agent was accurate 88% vs 63.34% of the time, a 38.93% relative gain.

What carries the argument

The load-bearing machinery is the Scenario tree: five Levels (informative?, humanitarian category?, victim/volunteer damage severity?, satellite damage severity?, UAV damage severity?), each fronted by an Enabler classifier that converts raw data into a confidence-score array. A Decision Maker (A2C RL agent) receives the padded confidence array plus a one-hot Level indicator and a credit counter, chooses a label or 'Gather Additional Data' (cost -1, max 5 credits per Level), and advances only on a correct label (+1 correct, -5 wrong). The Levels force the decision into a fixed, auditable sequence; the Enabler scores give the RL agent structured judgment; the credits let it defer a decision.

What would settle it

Give human operators the same Enabler confidence scores inside Disaster Maestro and compare accuracy; if human accuracy rises from about 63% toward the RL agent's 88%, the structure is not the source of the gain. Independently, train the RL agent on the same observations and actions but with the five Levels flattened into a single decision stage; if the 60.94% stability improvement disappears, the Scenario structure is the load-bearing part.

Watch

Extended reading notes

Core claim

The central claim is that structuring autonomous decision-making as a Scenario of five Levels, each fronted by an Enabler classifier, makes RL decisions both more accurate and more consistent than relying on raw model confidence or on human judgment. The claim is operationalized in a disaster-management setting: Level-1 judges whether incoming image-text data is informative, Level-2 identifies the humanitarian category, Level-3 assesses victim/volunteer-captured damage severity, Level-4 assesses satellite damage severity, and Level-5 assesses UAV damage severity. A reinforcement-learning agent (A2C) receives the Enabler confidence scores plus a Level indicator and a credit counter, and choos

Load-bearing premise

The load-bearing premise is that the human baseline is a fair control—yet humans never receive the Enabler agents' confidence scores, so the reported human-relative gains may measure classifier assistance rather than the Levels/Scenarios structure itself, and the stability benchmark is a single fixed argmax rule.

Editorial extensions

If this is right

  • An RL agent operating inside the five-Level structure achieves 88% mean per-Level accuracy across 1,809 validation Scenarios, versus 82.01% for the argmax Benchmark.
  • The 60.94% stability gain means the standard deviation of accuracy drops from 0.3841 to 0.15, making decisions more predictable for oversight.
  • The framework never requests additional data during evaluation (M.A.D = 0) yet outperforms humans, suggesting the structured Enabler confidence scores carry enough signal once organized into Levels.
  • The authors identify ethical and legal compliance checks, safety mechanisms for irreversible decisions, and validation checks as the next step before deployment in real operations.
  • The same Scenario/Level/Enabler structure is intended to transfer to other safety-critical domains such as aerospace and emergency response.

Reading between the lines

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

  • If human participants saw the Enabler agents' confidence scores in the web app, the reported 38.93% human-relative advantage might shrink substantially; the paper does not test this condition.
  • The Benchmark is a fixed argmax rule, so the 60.94% stability figure is relative to that rule, not to all judgement-based systems; a stronger baseline would be an RL agent without the Level structure.
  • Level-5's Enabler has the weakest macro F1 (65.28%), so improving drone-based damage classification would likely raise end-to-end Scenario accuracy more than improving any other single Level.
  • The three datasets are all from 2017 events, so transfer to newer disasters, other geographies, or SAR/oblique imagery is untested.
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.

Referee Report

4 major / 4 minor

Summary. The paper proposes a structured decision-making framework for autonomous disaster-management decisions, organized into five Levels within a Scenario, with Enabler agents (trained classifiers) providing confidence-score 'judgment insights' and a Decision Maker (A2C reinforcement learning agent or human operator) choosing among classification labels or a Gather-Additional-Data action. The framework is instantiated on CrisisMMD, xBD, and RescueNet data for disaster-phase and post-disaster-phase decisions. The authors compare the RL Decision Maker against a Benchmark agent (argmax over Enabler confidence arrays) and against human participants recruited as victims, volunteers, and stakeholders. The central claims are a 60.94% reduction in standard deviation of per-scenario accuracy relative to the Benchmark, and a 38.93% higher mean accuracy than the aggregate human participants. The paper includes a public code repository and a web-based human evaluation platform.

Significance. If the claims were supported, the paper would make a useful contribution to responsible-AI and disaster-management literatures: it operationalizes a hierarchical decision structure, provides a reusable implementation, and attempts a human-comparison study. Strengths include the public codebase, the explicit hyperparameter search, and the use of standard disaster datasets. However, the headline quantitative claims are not isolated from confounds. The RL agent receives trained Enabler classifier outputs while human participants do not, so the human-relative gain conflates the structure with the information advantage of learned classifiers. The 'judgement-based system' comparator is a fixed argmax policy over the same Enabler arrays, not a structure-free baseline. In addition, key reported numbers are internally inconsistent. The framework concept has potential, but the current evidence does not support the causal attributions made in the abstract and conclusions.

major comments (4)
  1. [§4.4 and §5.3.4 (Table 16)] The 38.93% accuracy advantage is computed between an RL agent that receives Enabler confidence vectors (§4.2, §4.3) and human participants explicitly not provided with Enabler judgment insights (§4.4). This comparison conflates two variables: the Levels/Scenarios structure and the presence of trained classifier outputs. The gain could be entirely due to the Enabler models, which humans never see. A human arm with Enabler outputs, or an RL arm without them, is needed to attribute the improvement to the structured framework. As it stands, the abstract's causal claim ('structured decision-making framework achieves... 38.93% higher accuracy') is unsupported.
  2. [§5.2.1, Table 11; §5.4] The 60.94% stability gain is presented as being 'compared to judgement-based systems,' but the Benchmark agent is a fixed argmax rule over the same Enabler confidence arrays, navigating the same Levels/Scenarios. Thus the comparison isolates A2C versus argmax within one and the same structure, not structured versus judgement-based decision-making. There is no condition that supplies Enabler outputs without the Level/Scenario structure, nor a condition with structure but without Enabler outputs. The headline overstates what the experimental design can establish.
  3. [§5.2.1 (Fig. 12, Table 11)] The reported numbers are internally inconsistent. The text states that 1,736 of 1,809 evaluation Scenarios achieved a perfect tree score of 5. Even if all remaining 73 Scenarios scored the minimum -5, the mean tree score would be at least (1736×5 - 73×5)/1809 ≈ 4.60, not the reported 1.4. The text also says that lower tree scores (2, 3, 4) were due to the agent requesting additional data, while Fig. 12e and Table 11 report M.A.D = 0 (σ = 0) on evaluation. These statements cannot all be true. The result tables and narrative require correction and re-analysis, and this issue is load-bearing for the quantitative claims.
  4. [§4.3 (Table 3, MDP definition)] The observation space is defined as a 9-dimensional array (4 confidence values + 5 one-hot level entries), but the MDP state is later defined as s ∈ R^10 = [c1..c4, l1..l5, qt], including the credit counter qt. If qt is not part of the observation, the process is not Markov from the agent's perspective because the transition and reward for Gather Additional Data depend on qt. If qt is observed, Table 3 and the worked example are incorrect. This ambiguity affects the reproducibility of the RL results and should be resolved.
minor comments (4)
  1. [Fig. 5 caption] The caption says the images are 'used to train Enabler agents across Scenario Levels 3 and 4,' but the surrounding text describes Level-4 (satellite) and Level-5 (drone). The caption level numbers appear to be off by one.
  2. [§5.1.4 and Table 10] Table 8 reports Macro Avg F1 = 0.9981 for Level-4, while Table 10 lists 0.9989. Please reconcile. Also, the text in §5.1.4 gives 'Macro Average F1-score of 99.89%,' which matches Table 10 but not Table 8.
  3. [§4.5, Fig. 12] The metric name is introduced as isGatherAdditionalDataRequested (Section 4.5), but Fig. 12 uses isAdditionalDataRequested. Standardize the terminology across the text and figures.
  4. [§5.3.4] The phrase 'approximately 2.75 times smaller' is ambiguous; suggesting 'about 2.75 times lower' or '63.8% lower' would be clearer.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; headline comparisons are confounded but not circular.

full rationale

The paper's central claims are empirical performance comparisons, not derivations. The RL agent is trained on Enabler confidence outputs from training splits and evaluated on Enabler outputs from validation splits, so the reported 88% accuracy is not measured on the same records used to train the policy. The 'judgement-based' comparator is an explicit argmax over the same confidence arrays (Section 5.2.1); while weak, it is an empirical baseline rather than a quantity that equals the RL result by construction. The human comparison is confounded: Section 4.4 states humans were not given Enabler insights, so the 38.93% gain (Table 16) may be due to the trained classifiers rather than the Levels/Scenarios structure, and the 60.94% stability gain is relative to the fixed argmax rule, not a general class of judgement-based systems. There is also a validation-selection concern: hyperparameters were tuned on a 500-scenario subset of the validation set (Section 4.3) and final numbers are reported on the full validation set (Section 5.2.1), which biases absolute figures upward. These are methodological/validity problems, not circular derivations: no quantity is defined in terms of the claim, no fitted parameter is renamed as a prediction, and no load-bearing self-citation appears. Therefore no significant circularity; score 2 reflects the confounds.

Assumptions & free parameters 5 free parameters · 5 assumptions · 3 invented entities

The framework rests on hand-picked reward/credit values, hyperparameters tuned on the validation set, and post-hoc label relabeling. The conceptual entities (Enabler agents, Levels, Scenarios) are introduced by the paper without independent verification. The axioms are reasonable domain assumptions but are not validated against real disaster-management workflows.

free parameters (5)
  • Reward values for RL = +1 correct, -5 wrong, -1 gather
    Chosen by hand in Section 4.3; affect the learned policy.
  • Credits per level = 5
    Chosen by hand in Section 4.3; limits the Gather Additional Data action.
  • A2C hyperparameters = gamma=0.995, n_steps=128, ent_coef=0.02, vf_coef=0.5, max_grad_norm=1, lr=5e-4
    Tuned via grid search on a 500-scenario validation subset (Section 4.3); these affect the RL policy.
  • Class relabeling choices = e.g., mild damage to little or no damage, injured/dead to affected individuals, vehicle damage to infrastructure/utility
    Post-hoc label transformations in Sections 4.2.1 and 4.2.2 that change the classification problem and affect Enabler performance.
  • Observation padding length = 4
    Confidence arrays padded to length 4; set in Section 4.3.
assumptions (5)
  • domain assumption Enabler agents provide useful judgment insights for the RL agent
    Section 4.2 assumes the trained classifiers produce informative confidence scores for downstream decision-making.
  • ad hoc to paper The five-level tree structure captures the disaster-management decision process
    Section 3 defines Levels and Scenarios without validation that this structure matches real stakeholder workflows.
  • domain assumption Human participants recruited via social media are representative of disaster stakeholders
    Section 4.4 describes recruitment through Facebook/LinkedIn and close contacts; representativeness is not established.
  • domain assumption CrisisMMD, xBD, and RescueNet are representative of disaster-management data
    Section 4.1 treats these benchmarks as sufficient for both training and evaluation; the paper acknowledges some limitations (e.g., temporal homogeneity, class imbalance).
  • ad hoc to paper The RL environment's random resampling models real-world data acquisition
    In Section 4.3, Gather Additional Data fetches a new random record, which may not reflect how real responders obtain new information.
invented entities (3)
  • Enabler agent
    purpose: AI model that provides judgment insights to the Decision Maker at each Level
    The paper introduces this concept; it is a trained classifier, not an independently verified entity.
  • Levels
    purpose: Stages in a Scenario representing critical decision points
    Organizational construct of the framework.
  • Scenarios
    purpose: Episode consisting of five Levels
    Organizational construct of the framework.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Structured AI Decision-Making in Disaster Management." pith.science (2026). https://pith.science/paper/CDZTT2SM

@misc{pith2026250901576,
  author       = {Pith},
  title        = {Pith review of: Structured AI Decision-Making in Disaster Management},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CDZTT2SM}},
  note         = {Machine review of arXiv:2509.01576}
}
read the original abstract

With artificial intelligence (AI) being applied to bring autonomy to decision-making in safety-critical domains such as the ones typified in the aerospace and emergency-response services, there has been a call to address the ethical implications of structuring those decisions, so they remain reliable and justifiable when human lives are at stake. This paper contributes to addressing the challenge of decision-making by proposing a structured decision-making framework as a foundational step towards responsible AI. The proposed structured decision-making framework is implemented in autonomous decision-making, specifically within disaster management. By introducing concepts of Enabler agents, Levels and Scenarios, the proposed framework's performance is evaluated against systems relying solely on judgement-based insights, as well as human operators who have disaster experience: victims, volunteers, and stakeholders. The results demonstrate that the structured decision-making framework achieves 60.94% greater stability in consistently accurate decisions across multiple Scenarios, compared to judgement-based systems. Moreover, the study shows that the proposed framework outperforms human operators with a 38.93% higher accuracy across various Scenarios. These findings demonstrate the promise of the structured decision-making framework for building more reliable autonomous AI applications in safety-critical contexts.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 40 canonical work pages

  1. [1]

    Procedia Environmental Sciences 20, 658–667 (2014) https://doi.org/10.1016/j.proenv.2 014.03.079

    Raungratanaamporn, I.-s., Pakdeeburee, P., Kamiko, A., Denpaiboon, C.: Government-communities collaboration in disaster management activity: Inves- tigation in the current flood disaster management policy in thailand. Procedia Environmental Sciences 20, 658–667 (2014) https://doi.org/10.1016/j.proenv.2 014.03.079 . The 4th International Conference on Sust...

    work page doi:10.1016/j.proenv.2 2014

  2. [2]

    Machine Learning and Knowledge Extraction 4(2), 446–473 (2022) https://doi.org/10.3 390/make4020020

    Linardos, V., Drakaki, M., Tzionas, P., Karnavas, Y.L.: Machine learning in dis- aster management: Recent developments in methods and applications. Machine Learning and Knowledge Extraction 4(2), 446–473 (2022) https://doi.org/10.3 390/make4020020

    work page 2022

  3. [3]

    In: Proceedings of the Twenty- Eighth Florida Artificial Intelligence Research Society Conference

    Roberts, M., Apker, T., Johnson, B., Auslander, B., Wellman, B., Aha, D.W.: Coordinating robot teams for disaster relief. In: Proceedings of the Twenty- Eighth Florida Artificial Intelligence Research Society Conference. AAAI Press, Hollywood, FL (2015). An extended version of this paper appears as: Roberts, M., Vattam, S., Alford, R., Auslander, Apker, T...

    work page 2015

  4. [4]

    lessons learned from three case studies in aviation, social care and road vehicles

    Salvini, P., Reinmund, T., Hardin, B., Grieman, K., Ten Holter, C., Johnson, A., Kunze, L., Winfield, A., Jirotka, M.: Human involvement in autonomous decision- making systems. lessons learned from three case studies in aviation, social care and road vehicles. Frontiers in Political Science 5 (2023) https://doi.org/10.338 9/fpos.2023.1238461

    work page arXiv 2023

  5. [5]

    International Journal of Public Health and Clinical Sciences (2019)

    Arifah, A., Tariq, M., Juni, M.H.: Decision making in disaster management cycle of natural disasters: A review. International Journal of Public Health and Clinical Sciences (2019)

    work page 2019

  6. [6]

    Ergonomics 43, 1153–1166 (2000)

    Smith, W., Dowell, J.: A case study of co-ordinative decision-making in disaster management. Ergonomics 43, 1153–1166 (2000)

    work page 2000

  7. [7]

    Scientific Reports 14(1), 29033 (2024) https://doi.org/10.1 038/s41598-024-80586-w

    Amiran, A., Behnam, B., Seyedin, S.: Ai-based model for site-selecting earthquake emergency shelters. Scientific Reports 14(1), 29033 (2024) https://doi.org/10.1 038/s41598-024-80586-w

    work page 2024

  8. [8]

    Scientific Reports 13(1), 8091 (2023) https://doi.org/10.1038/s41598-023-34874-6

    Shukla, D., Azad, H.K., Abhishek, K., Shitharth, S.: Disaster management ontology- an ontological approach to disaster management automation. Scientific Reports 13(1), 8091 (2023) https://doi.org/10.1038/s41598-023-34874-6

    work page doi:10.1038/s41598-023-34874-6 2023

Show all 51 references
  1. [9]

    Scientific Reports 12(1), 16286 (2022) https://doi.org/10.1038/s41598-022-19637-z

    Sun, L., Shawe-Taylor, J., D’Ayala, D.: Artificial intelligence-informed planning for the rapid response of hazard-impacted road networks. Scientific Reports 12(1), 16286 (2022) https://doi.org/10.1038/s41598-022-19637-z

    2022 doi

  2. [10]

    Scientific Reports 14(1), 17606 (2024) https://doi.org/10.1038/s41598-024-67886-x

    Fan, Z., Lin, Y., Ai, Y., Xu, H.: Adaptive task migration strategy with delay risk control and reinforcement learning for emergency monitoring. Scientific Reports 14(1), 17606 (2024) https://doi.org/10.1038/s41598-024-67886-x

    2024 doi

  3. [11]

    IGARSS 2020 - 2020 IEEE International Geoscience and Remote Sensing Symposium, 1699–1702 (2020) 36

    Gupta, T., Roy, S.: A hybrid model based on fused features for detection of natural disasters from satellite images. IGARSS 2020 - 2020 IEEE International Geoscience and Remote Sensing Symposium, 1699–1702 (2020) 36

    2020

  4. [12]

    CoRR abs/2004.11838 (2020) 2004.11838

    Ofli, F., Alam, F., Imran, M.: Analysis of social media data using multimodal deep learning for disaster response. CoRR abs/2004.11838 (2020) 2004.11838

    2004 arXiv

  5. [13]

    Soft Computing 23(24), 13393–13408 (2019) https://doi.org/10.1007/s00500-019-03878-8

    Yang, L., Cervone, G.: Analysis of remote sensing imagery for disaster assessment using deep learning: a case study of flooding event. Soft Computing 23(24), 13393–13408 (2019) https://doi.org/10.1007/s00500-019-03878-8

    2019 doi

  6. [14]

    In: Proceedings of the 2020 2nd International Conference on Image, Video and Signal Processing

    Dotel, S., Shrestha, A., Bhusal, A., Pathak, R., Shakya, A., Panday, S.P.: Disaster assessment from satellite imagery by analysing topographical features using deep learning. In: Proceedings of the 2020 2nd International Conference on Image, Video and Signal Processing. IVSP ’...

    2020

  7. [15]

    CoRR abs/1805.00713 (2018) 1805.00713

    Alam, F., Ofli, F., Imran, M.: Crisismmd: Multimodal twitter datasets from natural disasters. CoRR abs/1805.00713 (2018) 1805.00713

    2018 arXiv

  8. [16]

    IEEE Access 11, 72037– 72049 (2023) https://doi.org/10.1109/ACCESS.2023.3294474

    Rezk, M., Elmadany, N., Hamad, R.K., Badran, E.F.: Categorizing crises from social media feeds via multimodal channel attention. IEEE Access 11, 72037– 72049 (2023) https://doi.org/10.1109/ACCESS.2023.3294474

    2023

  9. [17]

    CoRR abs/1911.09296 (2019) 1911.09296

    Gupta, R., Hosfelt, R., Sajeev, S., Patel, N., Goodman, B., Doshi, J., Heim, E.T., Choset, H., Gaston, M.E.: xbd: A dataset for assessing building damage from satellite imagery. CoRR abs/1911.09296 (2019) 1911.09296

    1911 arXiv

  10. [18]

    Scientific Data 10(1) (2023) https://doi.org/10.1038/s41597-023-02799-4

    Rahnemoonfar, M., Chowdhury, T., Murphy, R.: Rescuenet: A high resolu- tion uav semantic segmentation dataset for natural disaster damage assessment. Scientific Data 10(1) (2023) https://doi.org/10.1038/s41597-023-02799-4

    2023 doi

  11. [19]

    arXiv preprint arXiv:2208.02205 (2022)

    Kaur, N., Lee, C.-C., Mostafavi, A., Mahdavi-Amiri, A.: Large-scale building damage assessment using a novel hierarchical transformer architecture on satellite images. arXiv preprint arXiv:2208.02205 (2022)

    2022 arXiv

  12. [20]

    Information Systems Frontiers 12, 49–65 (2010) https://doi.org/ 10.1007/s10796-009-9174-z

    Bharosa, N., Lee, J., Janssen, M.: Challenges and obstacles in sharing and coor- dinating information during multi-agency disaster response: Propositions from field exercises. Information Systems Frontiers 12, 49–65 (2010) https://doi.org/ 10.1007/s10796-009-9174-z

    2010 doi

  13. [21]

    International Journal of Disaster Risk Reduction51, 101762 (2020) https://doi.org/10.1016/j.ijdrr.2020.101762

    Misra, S., Roberts, P., Rhodes, M.: Information overload, stress, and emergency managerial thinking. International Journal of Disaster Risk Reduction51, 101762 (2020) https://doi.org/10.1016/j.ijdrr.2020.101762

    2020

  14. [22]

    International Journal of Disaster Risk Reduction 95, 103831 (2023) https://doi.org/10.1016/ j.ijdrr.2023.103831 37

    Ingham, V., Wuersch, L., Islam, M.R., Hicks, J.: Indicators of community disas- ter fatigue: A case study in the new south wales blue mountains. International Journal of Disaster Risk Reduction 95, 103831 (2023) https://doi.org/10.1016/ j.ijdrr.2023.103831 37

    2023

  15. [23]

    IEEE Access 7, 54595–54614 (2019) https://doi.org/10.1109/ACCESS.2019.29 13340

    Shah, S.A., Seker, D.Z., Hameed, S., Draheim, D.: The rising role of big data ana- lytics and iot in disaster management: Recent advances, taxonomy and prospects. IEEE Access 7, 54595–54614 (2019) https://doi.org/10.1109/ACCESS.2019.29 13340

    2019 doi

  16. [24]

    Information Systems Journal 33(1), 34–55 (2023) https://doi.org/10.1111/isj.12350 https://onlinelibrary.wiley.com/doi/pdf/10.1111/isj.12350

    Nussbaumer, A., Pope, A., Neville, K.: A framework for applying ethics-by- design to decision support systems for emergency management. Information Systems Journal 33(1), 34–55 (2023) https://doi.org/10.1111/isj.12350 https://onlinelibrary.wiley.com/doi/pdf/10.1111/isj.12350

    2023 doi

  17. [25]

    SAE International Journal of Transportation Safety 4(1), 15–24 (2016) https://doi.org/10.4271/2016-01-0128

    Koopman, Philip, Wagner, Michael: Challenges in autonomous vehicle testing and validation. SAE International Journal of Transportation Safety 4(1), 15–24 (2016) https://doi.org/10.4271/2016-01-0128

    2016 doi

  18. [26]

    Rand Corporation, ??? (2017)

    Osoba, O.A., Welser IV, W., Welser, W.: An Intelligence in Our Image: The Risks of Bias and Errors in Artificial Intelligence. Rand Corporation, ??? (2017)

    2017

  19. [27]

    Transport Reviews 39(1), 103–128 (2019) https://doi.org/10.1080/01441647.2018.1494640 https://doi.org/10.1080/01441647.2018.1494640

    Taeihagh, A., Lim, H.S.M.: Governing autonomous vehicles: emerging responses for safety, liability, privacy, cybersecurity, and industry risks. Transport Reviews 39(1), 103–128 (2019) https://doi.org/10.1080/01441647.2018.1494640 https://doi.org/10.1080/01441647.2018.1494640

    2019

  20. [28]

    Ethics and information technology 6, 175–183 (2004)

    Matthias, A.: The responsibility gap: Ascribing responsibility for the actions of learning automata. Ethics and information technology 6, 175–183 (2004)

    2004

  21. [29]

    Law, Innovation and Technology 6(2), 193–220 (2014)

    Leenes, R., Lucivero, F.: Laws on robots, laws by robots, laws in robots: Regu- lating robot behaviour by design. Law, Innovation and Technology 6(2), 193–220 (2014)

    2014

  22. [30]

    Butcher, J., Beridze, I.: What is the state of artificial intelligence governance globally? The RUSI Journal 164(5-6), 88–96 (2019) https://doi.org/10.1080/03 071847.2019.1694260 https://doi.org/10.1080/03071847.2019.1694260

    2019

  23. [31]

    Kim, S.: Crashed software: Assessing product liability for software defects in automated vehicles. Duke L. & Tech. Rev. 16, 300 (2017)

    2017

  24. [32]

    Law, Innovation and Technology 9(1), 1–44 (2017)

    Leenes, R., Palmerini, E., Koops, B.-J., Bertolini, A., Salvini, P., Lucivero, F.: Regulatory challenges of robotics: some guidelines for addressing legal and ethical issues. Law, Innovation and Technology 9(1), 1–44 (2017)

    2017

  25. [33]

    Technological Forecasting and Social Change 167, 120686 (2021) https://doi.or g/10.1016/j.techfore.2021.120686

    Tan, S.Y., Taeihagh, A., Tripathi, A.: Tensions and antagonistic interactions of risks and ethics of using robotics and autonomous systems in long-term care. Technological Forecasting and Social Change 167, 120686 (2021) https://doi.or g/10.1016/j.techfore.2021.120686

    2021

  26. [34]

    Lele, A.: Disarmament, Arms Control and Arms Race, pp. 217–229. Springer, Singapore (2019) 38

    2019

  27. [35]

    Journal of Military Ethics 13(3), 211–227 (2014) https://doi.org/10.1080/1502 7570.2014.975010 https://doi.org/10.1080/15027570.2014.975010

    Roff, H.M.: The strategic robot problem: Lethal autonomous weapons in war. Journal of Military Ethics 13(3), 211–227 (2014) https://doi.org/10.1080/1502 7570.2014.975010 https://doi.org/10.1080/15027570.2014.975010

    2014

  28. [36]

    Regulation & Governance 15(4), 1071–1091 (2021) https://doi.org/10.1111/re go.12344 https://onlinelibrary.wiley.com/doi/pdf/10.1111/rego.12344

    Firlej, M., Taeihagh, A.: Regulating human control over autonomous systems. Regulation & Governance 15(4), 1071–1091 (2021) https://doi.org/10.1111/re go.12344 https://onlinelibrary.wiley.com/doi/pdf/10.1111/rego.12344

    2021 doi

  29. [37]

    Center for a New American Security Washington, DC (2016)

    Scharre, P.: Autonomous weapons and operational risk. Center for a New American Security Washington, DC (2016)

    2016

  30. [38]

    Central European Journal of International & Security Studies 12(3) (2018)

    Solovyeva, A., Hynek, N.: Going beyond the” killer robots” debate: Six dilemmas autonomous weapon systems raise. Central European Journal of International & Security Studies 12(3) (2018)

    2018

  31. [39]

    Information Systems Frontiers 25(1), 123–141 (2023) https://doi.org/10.1007/ s10796-022-10251-y

    Papagiannidis, E., Enholm, I.M., Dremel, C., Mikalef, P., Krogstie, J.: Toward ai governance: Identifying best practices and potential barriers and outcomes. Information Systems Frontiers 25(1), 123–141 (2023) https://doi.org/10.1007/ s10796-022-10251-y

    2023

  32. [40]

    AI Perspectives 1(1), 3 (2019) https://doi.org/10.1186/s42467-019-0003-z

    Ghallab, M.: Responsible ai: requirements and challenges. AI Perspectives 1(1), 3 (2019) https://doi.org/10.1186/s42467-019-0003-z

    2019 doi

  33. [41]

    Information Fusion 99, 101896 (2023) https://doi.org/10.1016/j.in ffus.2023.101896

    D´ ıaz-Rodr´ ıguez, N., Del Ser, J., Coeckelbergh, M., L´ opez de Prado, M., Herrera- Viedma, E., Herrera, F.: Connecting the dots in trustworthy artificial intelligence: From ai principles, ethics, and key requirements to responsible ai systems and regulation. Information Fus...

    2023

  34. [42]

    Quarterly Journal of Computational Technologies for Healthcare 6(8), 1–26 (2021)

    Khanna, S., Khanna, I., Srivastava, S., Pandey, V.: Ai governance framework for oncology: Ethical, legal, and practical considerations. Quarterly Journal of Computational Technologies for Healthcare 6(8), 1–26 (2021)

    2021

  35. [43]

    Neural Computing and Applications 35(2), 1607–1627 (2023) https://doi.org/10.1007/s00521-022 -07790-5

    Koshy, R., Elango, S.: Multimodal tweet classification in disaster response sys- tems using transformer-based bidirectional attention model. Neural Computing and Applications 35(2), 1607–1627 (2023) https://doi.org/10.1007/s00521-022 -07790-5

    2023 doi

  36. [44]

    CoRR abs/1512.03385 (2015) 1512.03385

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. CoRR abs/1512.03385 (2015) 1512.03385

    2015 arXiv

  37. [45]

    https://www.tensor flow.org/api docs/python/tf/keras/preprocessing/text/Tokenizer

    TensorFlow: Tokenizer - TensorFlow Keras Preprocessing. https://www.tensor flow.org/api docs/python/tf/keras/preprocessing/text/Tokenizer. Accessed: 2024-08-12 (2024)

    2024

  38. [46]

    Accessed: August 2024 (2014)

    Pennington, J., Socher, R., Manning, C.: GloVe: Global Vectors for Word Rep- resentation. Accessed: August 2024 (2014). https://nlp.stanford.edu/data/glov 39 e.twitter.27B.zip

    2024

  39. [47]

    In: Moschitti, A., Pang, B., Daelemans, W

    Pennington, J., Socher, R., Manning, C.: GloVe: Global vectors for word represen- tation. In: Moschitti, A., Pang, B., Daelemans, W. (eds.) Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 1532–1543. Association for Computatio...

    2014 doi

  40. [48]

    https: //pytorch.org/vision/stable/index.html

    Contributors, T.: TorchVision: PyTorch’s computer vision library (2016). https: //pytorch.org/vision/stable/index.html

    2016

  41. [49]

    Journal of Machine Learning Research 12, 2825–2830 (2011)

    Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Mach...

    2011

  42. [50]

    CoRR abs/1606.01540 (2016) 1606.01540

    Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., Zaremba, W.: Openai gym. CoRR abs/1606.01540 (2016) 1606.01540

    2016 arXiv

  43. [51]

    CoRR abs/1602.01783 (2016) 1602.01783 40

    Mnih, V., Badia, A.P., Mirza, M., Graves, A., Lillicrap, T.P., Harley, T., Silver, D., Kavukcuoglu, K.: Asynchronous methods for deep reinforcement learning. CoRR abs/1602.01783 (2016) 1602.01783 40

    2016 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.