pith. sign in

arxiv: 2503.05571 · v2 · submitted 2025-03-07 · 💻 cs.CY · cs.AI· cs.ET

Compliance of AI Systems

Julius Sch\"oning , Niklas Kruse This is my paper

Pith reviewed 2026-05-23 00:57 UTC · model grok-4.3

classification 💻 cs.CY cs.AIcs.ET
keywords AI complianceEU AI Actdata set complianceedge devicesbest practicestrustworthinesstransparencyexplainability
0
0 comments X

The pith

Data set compliance is the cornerstone for AI systems to meet EU AI Act standards, especially on edge devices.

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

The paper examines how AI systems can achieve compliance with legislation such as the EU AI Act, with particular attention to data sets. It identifies challenges that arise when AI is deployed on edge devices, which are decentralized and have limited computing resources. The authors analyze existing AI implementations to spot these issues and put forward the first set of best practices for developing, deploying, and operating compliant AI. Emphasis is placed on data set compliance as essential for building trustworthiness, transparency, and explainability. A sympathetic reader would care because these steps help align practical AI use with emerging legal and ethical requirements.

Core claim

By analyzing AI implementations, the paper identifies challenges and proposes the first best practices for legal compliance when developing, deploying, and running AI. The importance of data set compliance is highlighted as a cornerstone for ensuring the trustworthiness, transparency, and explainability of AI systems, which must be aligned with ethical standards set forth in regulatory frameworks such as the AI Act.

What carries the argument

Best practices for legal compliance in AI development, deployment, and operation, centered on data set compliance to support edge device use.

If this is right

  • Following the best practices will support trustworthiness, transparency, and explainability in AI outputs.
  • Data set compliance will serve as the foundation for meeting ethical standards in regulatory frameworks like the AI Act.
  • Edge device deployments will require attention to compliance mechanisms suited to limited resources.
  • Responsible development of embedded AI systems will advance through these initial practices.
  • Ongoing discourse on AI compliance will benefit from the identified challenges and suggested approaches.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Developers may need to audit existing data collection pipelines before applying the practices to new projects.
  • The absence of technical implementation details leaves room for follow-up work on lightweight compliance tools for resource-constrained hardware.
  • Similar compliance needs could arise in other regulatory environments beyond the EU AI Act, such as data protection laws.
  • Testing the practices on real-world edge hardware in varied domains could reveal domain-specific adjustments.

Load-bearing premise

Best practices for legal compliance can be feasibly implemented on edge devices despite their decentralized nature and limited computing resources, without providing concrete technical details on how this is achieved.

What would settle it

Documenting cases where edge-deployed AI systems follow the proposed best practices yet still violate AI Act requirements on transparency or data handling would falsify the central claim.

Figures

Figures reproduced from arXiv: 2503.05571 by Julius Sch\"oning, Niklas Kruse.

Figure 1
Figure 1. Figure 1: The pipeline of applying AI [17] and the steps where compliance is needed are highlighted with a gray background. a plethora of options available [18], it is important to carefully examine several architectures. • 5. Training the AI Architecture: The most resource￾intensive step involves using 80% of the data set to train the model. The remaining 10% is used to evaluate training performance and detect over… view at source ↗
Figure 2
Figure 2. Figure 2: Different XAI techniques concerning the training process [ [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Proposed platform structure based on the respective stakeholder resources [ [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
read the original abstract

The increasing integration of artificial intelligence (AI) systems in various fields requires solid concepts to ensure compliance with upcoming legislation. This paper systematically examines the compliance of AI systems with relevant legislation, focusing on the EU's AI Act and the compliance of data sets. The analysis highlighted many challenges associated with edge devices, which are increasingly being used to deploy AI applications closer and closer to the data sources. Such devices often face unique issues due to their decentralized nature and limited computing resources for implementing sophisticated compliance mechanisms. By analyzing AI implementations, the paper identifies challenges and proposes the first best practices for legal compliance when developing, deploying, and running AI. The importance of data set compliance is highlighted as a cornerstone for ensuring the trustworthiness, transparency, and explainability of AI systems, which must be aligned with ethical standards set forth in regulatory frameworks such as the AI Act. The insights gained should contribute to the ongoing discourse on the responsible development and deployment of embedded AI systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper examines compliance of AI systems with the EU AI Act, with emphasis on data-set compliance as a cornerstone for trustworthiness, transparency, and explainability. It identifies challenges specific to edge devices arising from decentralization and limited compute resources, analyzes AI implementations, and proposes what it describes as the first best practices for legal compliance during development, deployment, and operation of AI systems aligned with regulatory and ethical standards.

Significance. If the proposed best practices were shown to be actionable and technically feasible on resource-constrained edge devices, the work could usefully contribute to ongoing discussions on responsible embedded-AI deployment under the AI Act. The explicit focus on data-set compliance as foundational is a clear framing that aligns with regulatory priorities.

major comments (2)
  1. [Abstract (edge-device challenges paragraph)] The central claim that the paper proposes the first best practices for AI Act compliance rests on the identification of edge-device challenges, yet the manuscript supplies only high-level issue statements without concrete mechanisms (e.g., lightweight provenance logging, on-device auditing protocols, or memory-bounded explainability methods) or resource estimates. This absence directly undermines the feasibility assertion for decentralized, resource-limited devices.
  2. [Abstract (analysis paragraph)] The analysis states that best practices were derived 'by analyzing AI implementations,' but provides neither the scope of the implementations examined, the methodology used, nor any empirical or case-study evidence supporting the resulting practices. This evidentiary gap is load-bearing for the claim of having identified actionable compliance solutions.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback, which identifies key areas where the presentation of our contributions can be clarified and strengthened. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Abstract (edge-device challenges paragraph)] The central claim that the paper proposes the first best practices for AI Act compliance rests on the identification of edge-device challenges, yet the manuscript supplies only high-level issue statements without concrete mechanisms (e.g., lightweight provenance logging, on-device auditing protocols, or memory-bounded explainability methods) or resource estimates. This absence directly undermines the feasibility assertion for decentralized, resource-limited devices.

    Authors: The manuscript presents a conceptual analysis and high-level best practices rather than detailed technical mechanisms or resource estimates. This scope reflects the paper's focus on framing regulatory compliance challenges and foundational guidelines aligned with the AI Act, particularly data-set compliance. We will revise the abstract and relevant sections to qualify the claim as proposing 'initial' rather than 'the first' best practices and to explicitly note that the proposals are high-level guidelines requiring further technical development for implementation on resource-constrained devices. revision: partial

  2. Referee: [Abstract (analysis paragraph)] The analysis states that best practices were derived 'by analyzing AI implementations,' but provides neither the scope of the implementations examined, the methodology used, nor any empirical or case-study evidence supporting the resulting practices. This evidentiary gap is load-bearing for the claim of having identified actionable compliance solutions.

    Authors: The best practices were synthesized from regulatory requirements and general knowledge of edge AI deployments in the literature, without a formal systematic review or empirical validation. We will revise the manuscript to add a dedicated subsection describing the approach, including examples of AI implementations considered (such as on-device inference models) and the reasoning linking challenges to practices. This will remain a qualitative analysis without new empirical evidence. revision: yes

Circularity Check

0 steps flagged

No circularity: qualitative legal analysis with external references only

full rationale

The paper is a discussion of AI Act compliance challenges for edge devices and proposes high-level best practices, with data-set compliance as a cornerstone. It contains no equations, fitted parameters, predictions, or derivations. All claims reference external legislation and standards rather than reducing to self-defined quantities or self-citations. The central argument rests on identification of issues and general recommendations, not on any load-bearing internal logic that could be circular by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is a qualitative policy discussion paper with no mathematical derivations, free parameters, axioms, or invented entities; the central claims rest on interpretation of external legislation rather than internal formal structure.

pith-pipeline@v0.9.0 · 5685 in / 1144 out tokens · 56724 ms · 2026-05-23T00:57:36.740982+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages

  1. [1]

    Artificial intelligence (ai) in agriculture,

    S. Y . Liu, “Artificial intelligence (ai) in agriculture,” IT Professional , vol. 22, no. 3, pp. 14–15, May 2020 . doi: 10.1109/mitp.2020.2986121

    work page doi:10.1109/mitp.2020.2986121 2020

  2. [2]

    Ai-based crop rotation for sus- tainable agriculture worldwide,

    J. Schoning and M. L. Richter, “Ai-based crop rotation for sus- tainable agriculture worldwide,” in 2021 IEEE Global Humanitar- ian Technology Conference (GHTC) . IEEE, Oct. 2021 . doi: 10.1109/ghtc53159.2021.9612460 pp. 142–146

    work page doi:10.1109/ghtc53159.2021.9612460 2021

  3. [3]

    Safe and trustful ai for closed-loop control systems,

    J. Sch ¨oning and H.-J. Pfisterer, “Safe and trustful ai for closed-loop control systems,” Electronics, vol. 12, no. 16, p. 3489, Aug. 2023 . doi: 10.3390/electronics12163489

    work page doi:10.3390/electronics12163489 2023

  4. [4]

    Overview of closed-loop control systems and artificial intelligence utilization in greenhouse farming,

    D. Walczuch, T. Nitzsche, T. Seidel, and J. Schoning, “Overview of closed-loop control systems and artificial intelligence utilization in greenhouse farming,” in 2022 IEEE International Conference on Omni-layer Intelligent Systems (COINS) . IEEE, Aug. 2022 . doi: 10.1109/coins54846.2022.9854938 pp. 1–6

    work page doi:10.1109/coins54846.2022.9854938 2022

  5. [5]

    Ava: An automated and ai-driven intelligent visual analytics framework,

    J. Wang, X. Li, C. Li, D. Peng, A. Z. Wang, Y . Gu, X. Lai, H. Zhang, X. Xu, X. Dong, Z. Lin, J. Zhou, X. Liu, and W. Chen, “Ava: An automated and ai-driven intelligent visual analytics framework,” Visual Informatics , vol. 8, no. 2, pp. 106–114, Jun. 2024 . doi: 10.1016/j.visinf.2024.06.002

    work page doi:10.1016/j.visinf.2024.06.002 2024

  6. [6]

    Visual analytics for video applications,

    P. Tanisaro, J. Sch ¨oning, K. Kurzhals, G. Heidemann, and D. Weiskopf, “Visual analytics for video applications,” it - Information Technology , vol. 57, no. 1, pp. 30–36, Jan. 2015 . doi: 10.1515/itit-2014-1072

    work page doi:10.1515/itit-2014-1072 2015

  7. [7]

    Explainable ai for safe and trustworthy autonomous driving: A systematic review,

    A. Kuznietsov, B. Gyevnar, C. Wang, S. Peters, and S. V . Al- brecht, “Explainable ai for safe and trustworthy autonomous driving: A systematic review,” IEEE Transactions on Intelligent Transporta- tion Systems , vol. 25, no. 12, pp. 19 342–19 364, Dec. 2024 . doi: 10.1109/tits.2024.3474469

    work page doi:10.1109/tits.2024.3474469 2024

  8. [8]

    A review of artificial intelligence applications in high-speed railway systems,

    X. Li, M. Zhu, B. Zhang, X. Wang, Z. Liu, and L. Han, “A review of artificial intelligence applications in high-speed railway systems,” High-speed Railway , vol. 2, no. 1, pp. 11–16, Mar. 2024 . doi: 10.1016/j.hspr.2024.01.002

    work page doi:10.1016/j.hspr.2024.01.002 2024

  9. [9]

    A framework for the implementa- tion of artificial intelligence in business enterprises: A readiness model,

    M. Nortje and S. Grobbelaar, “A framework for the implementa- tion of artificial intelligence in business enterprises: A readiness model,” in 2020 IEEE International Conference on Engineering, Technology and Innovation (ICE/ITMC) . IEEE, Jun. 2020 . doi: 10.1109/ice/itmc49519.2020.9198436 pp. 1–10

    work page doi:10.1109/ice/itmc49519.2020.9198436 2020

  10. [10]

    Drivers Behind the Public Perception of Artificial Intelligence: Insights from Major Australian Cities

    M. M. Anderson and K. Fort, “From the ground up: developing a practical ethical methodology for integrating ai into industry,” AI & SOCIETY, vol. 38, no. 2, pp. 631–645, Jul. 2022 . doi: 10.1007/s00146- 022-01531-x

    work page doi:10.1007/s00146- 2022

  11. [11]

    Managing ai-first products: Roles, skills, challenges, and strategies of ai product managers,

    N. A. Parikh, “Managing ai-first products: Roles, skills, challenges, and strategies of ai product managers,” IEEE Engineering Management Review, pp. 1–11, 2025 . doi: 10.1109/emr.2025.3530942

    work page doi:10.1109/emr.2025.3530942 2025

  12. [12]

    Artificial intelligence and the ‘good society’: the us, eu, and uk approach,

    C. Cath, S. Wachter, B. Mittelstadt, M. Taddeo, and L. Floridi, “Artificial intelligence and the ‘good society’: the us, eu, and uk approach,”Science and Engineering Ethics , Mar. 2017 . doi: 10.1007/s11948-017-9901-7

    work page doi:10.1007/s11948-017-9901-7 2017

  13. [13]

    In ai we trust incrementally: a multi-layer model of trust to analyze human-artificial intelligence interactions,

    A. Ferrario, M. Loi, and E. Vigan `o, “In ai we trust incrementally: a multi-layer model of trust to analyze human-artificial intelligence interactions,” Philosophy &; Technology , vol. 33, no. 3, pp. 523–539, Oct. 2019 . doi: 10.1007/s13347-019-00378-3

    work page doi:10.1007/s13347-019-00378-3 2019

  14. [14]

    Bayesian neural networks uncertainty quantification with cubature rules

    K. Crockett, M. Garratt, A. Latham, E. Colyer, and S. Goltz, “Risk and trust perceptions of the public of artifical intelligence applications,” in 2020 International Joint Conference on Neural Networks (IJCNN) . IEEE, Jul. 2020 . doi: 10.1109/ijcnn48605.2020.9207654 pp. 1–8

    work page doi:10.1109/ijcnn48605.2020.9207654 2020

  15. [15]

    (2020) White paper on artificial intelligence: A european approach to excellence and trust

    European Commission. (2020) White paper on artificial intelligence: A european approach to excellence and trust. [Online]. Available: https://commission.europa.eu/publications/ white-paper-artificial-intelligence-european-approach-excellence-and-trust en

    work page 2020

  16. [16]

    Argumentative xai: A survey,

    K. ˇCyras, A. Rago, E. Albini, P. Baroni, and F. Toni, “Argumentative xai: A survey,” 2021

    work page 2021

  17. [17]

    Ai-in-the-loop – the impact of hmi in ai-based application,

    J. Sch ¨oning and C. Westerkamp, “Ai-in-the-loop – the impact of hmi in ai-based application,” 2023

    work page 2023

  18. [18]

    van Veen and S

    F. van Veen and S. Leijnen. (2019, Apr.) The neural network zoo. [On- line]. Available: https://www.asimovinstitute.org/neural-network-zoo/

    work page 2019

  19. [19]

    A survey of safety and trustwor- thiness of deep neural networks: V erification, testing, adv ersarial attack and defence, and interpretability

    X. Huang, D. Kroening, W. Ruan, J. Sharp, Y . Sun, E. Thamo, M. Wu, and X. Yi, “A survey of safety and trustworthiness of deep neural networks: Verification, testing, adversarial attack and defence, and interpretability,” Computer Science Review , vol. 37, p. 100270, 2020 . doi: 10.1016/j.cosrev.2020.100270

    work page doi:10.1016/j.cosrev.2020.100270 2020

  20. [20]

    Adversarial attacks on deep neural networks for time series classifica- tion,

    H. I. Fawaz, G. Forestier, J. Weber, L. Idoumghar, and P.-A. Muller, “Adversarial attacks on deep neural networks for time series classifica- tion,” in International Joint Conference on Neural Networks (IJCNN) . IEEE, 2019 . doi: 10.1109/ijcnn.2019.8851936

    work page doi:10.1109/ijcnn.2019.8851936 2019

  21. [21]

    Explainable and trustworthy ai compli- ance for farms,

    N. Kruse and J. Sch ¨oning, “Explainable and trustworthy ai compli- ance for farms,” in 45. Jahrestagung der Gesellschaft f ¨ur Informatik in der Land-, F orst- und Ern ¨ahrungswirtschaft (GIL) , 2025 . doi: 10.18420/giljt2025 32

    work page doi:10.18420/giljt2025 2025

  22. [22]

    Explainable artificial intelligence (xai): Concepts, taxonomies, opportunities and challenges toward responsi- ble ai,

    A. B. Arrieta, N. D ´ıaz-Rodr´ıguez, J. D. Ser, A. Bennetot, S. Tabik, A. Barbado, S. Garcia, S. Gil-Lopez, D. Molina, R. Benjamins, R. Chatila, and F. Herrera, “Explainable artificial intelligence (xai): Concepts, taxonomies, opportunities and challenges toward responsi- ble ai,” Information Fusion , vol. 58, pp. 82–115, Jun. 2020 . doi: 10.1016/j.inffus...

    work page doi:10.1016/j.inffus.2019.12.012 2020

  23. [23]

    Hardin, Trust and trustworthiness , ser

    R. Hardin, Trust and trustworthiness , ser. The Russell Sage Foundation series on trust. New York: Russell Sage Foundation, 2002, no. volume

    work page 2002

  24. [24]

    ISBN 9781610442718 Includes bibliographical references and index

    work page

  25. [25]

    Legal conform data sets for yard tractors and robots,

    N. Kruse and J. Sch ¨oning, “Legal conform data sets for yard tractors and robots,” Computers and Electronics in Agriculture , vol. 223, p. 109106, Aug. 2024 . doi: 10.1016/j.compag.2024.109106

    work page doi:10.1016/j.compag.2024.109106 2024

  26. [26]

    Compliance of agricultural ai systems – app-based legal verification throughout the development,

    N. Kruse, P. Wachter, and J. Sch ¨oning, “Compliance of agricultural ai systems – app-based legal verification throughout the development,” in

    work page

  27. [27]

    doi: 10.18420/giljt2024 01

    Jahrestagung der Gesellschaft f ¨ur Informatik in der Land-, F orst- und Ern ¨ahrungswirtschaft (GIL) , 02 2024 . doi: 10.18420/giljt2024 01

    work page doi:10.18420/giljt2024 2024