pith. sign in

arxiv: 2512.05742 · v3 · pith:U4HD66UInew · submitted 2025-12-05 · 💻 cs.CY

Internal Deployment in the AI Act

Matteo Pistillo This is my paper

Pith reviewed 2026-05-21 17:48 UTC · model grok-4.3

classification 💻 cs.CY
keywords AI Actinternal deploymentscope of applicationEU regulationAI governancelegal interpretationscientific research exceptioncompliance obligations
0
0 comments X

The pith

The AI Act can be interpreted to cover internally deployed AI systems via its scope and exception rules.

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

This paper stress-tests arguments about whether the EU AI Act applies to AI models and systems that organizations use only internally rather than placing them on the market. It walks through how the general scope rules in Article 2(1)(a)-(c) could bring such systems inside the Act, then weighs possible objections from the scientific research exception in Article 2(6) and related clauses in Article 2(8). A sympathetic reader cares because the outcome decides whether companies must perform risk assessments, ensure transparency, or meet other obligations even when their AI stays in-house, directly affecting internal innovation practices and compliance costs across the EU.

Core claim

The memorandum establishes that Articles 2(1)(a)-(c), 2(6), and 2(8) of the AI Act can be interpreted to support application to internally deployed AI models and systems. These provisions become complementary once broken down to their most plausible meanings and read together with the definitions in Articles 3(1), 3(3), 3(4), 3(9), 3(10), 3(11), 3(12), 3(63) and Recitals 12, 13, 21, 25, 97, 109, and 110.

What carries the argument

The complementary interpretation of scope provisions in Article 2(1) alongside the scientific R&D exception in Article 2(6) and further rules in Article 2(8), clarified through the Act's definitions and recitals.

If this is right

  • Providers and deployers of internal AI systems may need to meet transparency, risk management, and documentation requirements.
  • The research exception is limited and does not automatically exempt all internal AI uses from the Act.
  • Legal interpretations can treat the listed articles as mutually reinforcing rather than in tension.
  • Compliance planning for companies becomes clearer once internal deployment is addressed through these pathways.

Where Pith is reading between the lines

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

  • The Commission could issue targeted guidance to reduce uncertainty for organizations running AI only inside their own operations.
  • Companies may need to classify and document internal AI uses more systematically to determine if exceptions apply.
  • This scope reading could create overlap points with data protection rules when internal AI processes personal data.

Load-bearing premise

The scientific R&D exception under Article 2(6) can be delineated without creating irresolvable conflicts with the general scope provisions in Article 2(1) when applied to internal deployment scenarios.

What would settle it

A guidance document or court ruling from EU institutions that explicitly states whether internally deployed non-research AI systems fall inside or outside the AI Act's obligations.

read the original abstract

This memorandum analyzes and stress-tests arguments in favor and against the inclusion of internal deployment within the scope of the European Union Artificial Intelligence Act (AI Act). In doing so, it aims to offer several possible interpretative pathways to the European Commission, AI providers and deployers, courts, and the legal and policy community at large based on Articles 2(1), 2(6), 2(8) of the AI Act. Specifically, this memorandum first analyzes interpretative pathways based on Article 2(1)(a)-(c) supporting the application of the AI Act to internally deployed AI models and systems. Then, it examines possible objections and exceptions based on Articles 2(6) and 2(8), with particular attention to the complexity of the scientific R&D exception under Article 2(6). Finally, it illustrates how Articles 2(1), 2(6), and 2(8) can be viewed as complementary to each other, once broken down to their most plausible meaning and interpreted in conjunction with Articles 3(1), 3(3), 3(4), 3(9), 3(10), 3(11), 3(12), 3(63), and Recitals 12, 13, 21, 25, 97, 109, and 110.

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

1 major / 2 minor

Summary. The memorandum analyzes interpretative pathways under the EU AI Act for internal deployment of AI models and systems. It first examines how Articles 2(1)(a)-(c) support application to internal deployments, then addresses potential objections and exceptions under Articles 2(6) (scientific R&D) and 2(8), and concludes that these provisions, when parsed with Article 3 definitions and Recitals 12, 13, 21, 25, 97, 109, and 110, can be read as complementary rather than in conflict.

Significance. If the interpretations are accepted, the work offers practical guidance to the European Commission, providers, deployers, and courts on a high-volume compliance question: whether and how the AI Act reaches AI used internally within organizations. By stress-testing arguments for and against inclusion and mapping multiple pathways, it contributes to regulatory clarity in an area where internal deployment is widespread but textual guidance is limited.

major comments (1)
  1. [Article 2(6) analysis] Section on the scientific R&D exception (Article 2(6)): the claim that the exception can be delineated without irresolvable conflict with Article 2(1) for internal deployment rests on the 'solely' or 'primarily' language and Recital 21, yet the text does not supply a concrete test or example for mixed-use cases in which the same internal system serves both research and operational/commercial functions within one entity. This leaves the complementarity conclusion open to the precise objection raised in the stress-test note.
minor comments (2)
  1. The abstract and conclusion list specific recitals but the main text would benefit from more granular, numbered cross-references (e.g., 'as required by Recital 21, ...') each time a recital is invoked to support a particular reading.
  2. Article 2(8) receives comparatively brief treatment; a short dedicated subsection clarifying its interaction with internal deployment would improve readability.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive review and for recognizing the memorandum's contribution to clarifying the scope of the AI Act for internal deployments. The suggestion to strengthen the Article 2(6) discussion with more concrete guidance on mixed-use cases is well taken, and we address it directly below.

read point-by-point responses

  1. Referee: [Article 2(6) analysis] Section on the scientific R&D exception (Article 2(6)): the claim that the exception can be delineated without irresolvable conflict with Article 2(1) for internal deployment rests on the 'solely' or 'primarily' language and Recital 21, yet the text does not supply a concrete test or example for mixed-use cases in which the same internal system serves both research and operational/commercial functions within one entity. This leaves the complementarity conclusion open to the precise objection raised in the stress-test note.

    Authors: We agree that the original text would benefit from greater specificity on mixed-use scenarios to reinforce the complementarity reading. The analysis in the manuscript relies on the 'solely' or 'primarily' qualifier in Article 2(6), read together with Recital 21 and the definitions in Article 3, to argue that the R&D exception need not conflict with the broad scope in Article 2(1)(a)-(c). To address the referee's point, the revised version adds a dedicated illustrative example of a mixed-use internal system (e.g., an AI tool initially developed for internal research but subsequently used for both R&D and operational decision-making within the same entity). The example applies the 'primarily' criterion by reference to documented purpose, resource allocation, and the recitals' emphasis on the exception's narrow character, showing how the provisions can operate without irresolvable overlap. This addition directly engages the stress-test objection while remaining within the memorandum's scope as an interpretative analysis rather than a prescriptive compliance manual. revision: yes

Circularity Check

0 steps flagged

No circularity: legal interpretation grounded in external statute text

full rationale

The memorandum conducts a textual analysis of the AI Act's scope provisions (Articles 2(1), 2(6), 2(8)) and related definitions/recitals, offering interpretative pathways without any self-referential definitions, fitted inputs renamed as predictions, or load-bearing self-citations. All reasoning draws from the fixed external text of the EU legislation itself rather than reducing claims to the paper's own prior outputs or author-specific theorems. This constitutes a standard legal exegesis that remains self-contained against the statute as benchmark.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The analysis relies on standard principles of EU statutory interpretation and the text of the AI Act itself; no free parameters, new entities, or ad-hoc inventions are introduced.

axioms (1)
  • domain assumption Standard methods of statutory interpretation apply to the AI Act, including reading provisions in context with definitions and recitals.
    Invoked throughout the analysis of Articles 2(1), 2(6), and 2(8) to support complementary readings.

pith-pipeline@v0.9.0 · 5763 in / 1223 out tokens · 56367 ms · 2026-05-21T17:48:08.646654+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    This memorandum analyzes ... interpretative pathways based on Article 2(1)(a)-(c) supporting the application of the AI Act to internally deployed AI models and systems ... Articles 2(1), 2(6), and 2(8) can be viewed as complementary ... interpreted in conjunction with Articles 3(1), 3(3), ... and Recitals 12, 13, 21, 25, 97, 109, and 110.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Risk Reporting for Developers' Internal AI Model Use

    cs.CY 2026-04 unverdicted novelty 4.0

    A harmonized risk reporting standard for internal frontier AI model use, structured around autonomous misbehavior and insider threats using means, motive, and opportunity factors.

Reference graph

Works this paper leans on

66 extracted references · 66 canonical work pages · cited by 1 Pith paper · 3 internal anchors

  1. [1]

    arXiv preprint arXiv:2504.12170 , year =

    Stix, Charlotte and Carlsmith, Joe and Schiavone, Stephen and Christiano, Paul and others , title =. arXiv preprint arXiv:2504.12170 , year =

    work page arXiv

  2. [2]

    arXiv preprint arXiv:2511.15846 , year =

    Stix, Charlotte and Sharkey, Lee and others , title =. arXiv preprint arXiv:2511.15846 , year =

    work page arXiv

  3. [3]

    Model evaluation for extreme risks

    Shevlane, Toby and Farquhar, Sebastian and Garfinkel, Ben and Phuong, Mary and Whittlestone, Jess and Leung, Jade and Kokotajlo, Daniel and Marchal, Nahema and Anderljung, Markus and Kolt, Noam and Ho, Lewis and Siddarth, Divya and Avin, Shahar and Hawkins, Will and Kim, Been and Gabriel, Iason and Bolina, Vijay and Clark, Jack and Bengio, Yoshua and Chri...

    work page arXiv

  4. [4]

    2023 , howpublished =

    Hubinger, Evan and Denison, Carson and Mu, Jesse and Lambert, Mike and Tong, Meg and MacDiarmid, Monte , title =. 2023 , howpublished =

    work page 2023

  5. [5]

    Hubinger, Evan and Denison, Carson and Mu, Jesse and Lambert, Mike and Tong, Meg and MacDiarmid, Monte and Lanham, Tamera and Ziegler, Daniel M. and Maxwell, Tim and Cheng, Newton and Jermyn, Adam and Askell, Amanda and Radhakrishnan, Ansh and Anil, Cem and Duvenaud, David and Ganguli, Deep and Barez, Fazl and Clark, Jack and Ndousse, Kamal and Sachan, Ks...

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    Emergent Misalignment : Narrow finetuning can produce broadly misaligned LLMs , May 2025

    Betley, Jan and Tan, Daniel and Warncke, Niels and Sztyber-Betley, Anna and Bao, Xuchan and Soto, Mart. Emergent Misalignment: Narrow Finetuning Can Produce Broadly Misaligned. arXiv preprint arXiv:2502.17424 , year =

    work page arXiv

  7. [7]

    arXiv preprint arXiv:2506.11613 , year =

    Turner, Alex and others , title =. arXiv preprint arXiv:2506.11613 , year =

    work page arXiv

  8. [8]

    arXiv preprint arXiv:2506.19823 , year =

    Wang, Miles and others , title =. arXiv preprint arXiv:2506.19823 , year =

    work page arXiv

  9. [9]

    arXiv preprint arXiv:2510.20487 , year =

    Hua, Wenjia and others , title =. arXiv preprint arXiv:2510.20487 , year =

    work page arXiv

  10. [10]

    arXiv preprint arXiv:2503.10965 , year =

    Marks, Samuel and Treutlein, Johannes and Bricken, Trenton and Lindsey, Jack and Marcus, Jonathan and Mishra-Sharma, Siddharth and Ziegler, Daniel and Perez, Ethan and Sharma, Mrinank and Denison, Carson and Xu, Fabien Roger and others , title =. arXiv preprint arXiv:2503.10965 , year =

    work page arXiv

  11. [11]

    Open Problems in Mechanistic Interpretability

    Sharkey, Lee and Chughtai, Bilal and Batson, Joshua and Lindsey, Jack and Wu, Jeff and Bushnaq, Lucius and Goldowsky-Dill, Nicholas and Heimersheim, Stefan and Ortega, Alejandro and Bloom, Joseph and Biderman, Stella and Conmy, Arthur , title =. arXiv preprint arXiv:2501.16496 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    A Causal Framework for

    Sharkey, Lee and Ghuidhir, Cl. A Causal Framework for. 2023 , url =

    work page 2023

  13. [13]

    arXiv preprint arXiv:2511.13653 , year =

    Gao, Leo and others , title =. arXiv preprint arXiv:2511.13653 , year =

    work page arXiv

  14. [14]

    arXiv preprint arXiv:2505.13787 , year =

    Cundy, Chris and Gleave, Adam , title =. arXiv preprint arXiv:2505.13787 , year =

    work page arXiv

  15. [15]

    and Morrow, Carmen and Novati, Guido and Hasanbegovic, Zinovia and Hu, Shanqing and Banerjee, Suraj and Sanders, Tessa M

    Vaxenburg, Roman and Siwanowicz, Igor and Merel, Josh and Robie, Alice A. and Morrow, Carmen and Novati, Guido and Hasanbegovic, Zinovia and Hu, Shanqing and Banerjee, Suraj and Sanders, Tessa M. and Lauer, Jessica and others , title =. Nature , year =

    work page

  16. [16]

    Physics Letters B , volume =

    Observation of a New Particle in the Search for the. Physics Letters B , volume =. 2012 , url =

    work page 2012

  17. [17]

    Highly Accurate Protein Structure Prediction with

    Jumper, John and Evans, Richard and Pritzel, Alexander and Green, Tim and Figurnov, Michael and Ronneberger, Olaf and Tunyasuvunakool, Kathryn and Bates, Russ and. Highly Accurate Protein Structure Prediction with. Nature , volume =. 2021 , url =

    work page 2021

  18. [18]

    Early Science Acceleration Experiments with

    Bubeck, S. Early Science Acceleration Experiments with. 2025 , url =

    work page 2025

  19. [19]

    The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

    Lu, Chris and Lu, Cong and Lange, Robert Tjarko and Foerster, Jakob and Clune, Jeff and Ha, David , title =. arXiv preprint arXiv:2408.06292 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    Robert Tjarko Lange, Yuki Imajuku, and Edoardo Cetin

    Kwa, Thomas and West, Ben and Becker, Joel and Deng, Amy and Garcia, Katharyn and Hasin, Max and Jawhar, Sami and Kinniment, Megan and Rush, Nate and Arx, Sydney Von and Bloom, Ryan and Broadley, Thomas and Du, Haoxing and Goodrich, Brian and Jurkovic, Nikola and Miles, Luke Harold and Nix, Sea and Lin, Tao and Parikh, Neev and Rein, David and Sato, Lucas...

    work page arXiv

  21. [21]

    2024 , url =

    Wijk, Hjalmar and Lin, Tao and Becker, Joel and Jawhar, Sami and Parikh, Neev and Broadley, Thomas and Chan, Lawrence and Chen, Michael and Clymer, Joshua and Dhyani, Jai and Ericheva, Elena and Garcia, Katharyn and Goodrich, Brian and Jurkovic, Nikola and Kinniment, Megan and Lajko, Aron and Nix, Seraphina and Sato, Lucas and Saunders, William and Taran,...

    work page 2024

  22. [22]

    2025 , url =

    Behrouz, Ali and others , title =. 2025 , url =

    work page 2025

  23. [23]

    arXiv preprint arXiv:2302.02337 , year =

    Hacker, Philipp and Engel, Andreas and Mauer, Marco , title =. arXiv preprint arXiv:2302.02337 , year =

    work page arXiv

  24. [24]

    GRUR International , volume =

    Finck, Michele , title =. GRUR International , volume =. 2025 , url =

    work page 2025

  25. [25]

    Van Eecke, Patrick and Regenhardt, Janina , title =. The. 2024 , url =

    work page 2024

  26. [26]

    German Law Journal , volume =

    Almada, Marco and Radu, Anca , title =. German Law Journal , volume =. 2024 , url =

    work page 2024

  27. [27]

    Computer Law & Security Review , year =

    Czerniawski, Mariusz , title =. Computer Law & Security Review , year =

    work page

  28. [28]

    Minnesota Journal of Law, Science & Technology , year =

    Boine, Claire and Rolnick, David , title =. Minnesota Journal of Law, Science & Technology , year =

    work page

  29. [29]

    Lanam. The. 2024 , url =

    work page 2024

  30. [30]

    2025 , url =

    Acharya, Ashwin and Delaney, Oscar , title =. 2025 , url =

    work page 2025

  31. [31]

    2025 , url =

    Preparedness Framework, Version 2 , institution =. 2025 , url =

    work page 2025

  32. [32]

    2025 , url =

    Responsible Scaling Policy , institution =. 2025 , url =

    work page 2025

  33. [33]

    2025 , url =

    Transparency Model Report , institution =. 2025 , url =

    work page 2025

  34. [34]

    2026 , howpublished =

    Auditing an. 2026 , howpublished =

    work page 2026

  35. [35]

    Understanding Neural Networks through Sparse Circuits , year =

    work page

  36. [36]

    2025 , howpublished =

    work page 2025

  37. [37]

    2025 , url =

    Frontier Safety Framework, Version 3 , institution =. 2025 , url =

    work page 2025

  38. [38]

    2026 , url =

    Advanced. 2026 , url =

    work page 2026

  39. [39]

    2025 , url =

    Common Elements of Frontier. 2025 , url =

    work page 2025

  40. [40]

    2026 , howpublished =

    Measuring. 2026 , howpublished =

    work page 2026

  41. [41]

    2024 , howpublished =

    Amodei, Dario , title =. 2024 , howpublished =

    work page 2024

  42. [42]

    2025 , howpublished =

    Adler, Steven , title =. 2025 , howpublished =

    work page 2025

  43. [43]

    2025 , url =

    Kokotajlo, Daniel and Alexander, Scott and Larsen, Thomas and Lifland, Eli and Dean, Romeo , title =. 2025 , url =

    work page 2025

  44. [44]

    2025 , howpublished =

    Davidson, Tom , title =. 2025 , howpublished =

    work page 2025

  45. [45]

    2025 , howpublished =

    Davidson, Tom and Hadshar, Rose and MacAskill, Will , title =. 2025 , howpublished =

    work page 2025

  46. [46]

    2025 , howpublished =

    Eth, Daniel and Davidson, Tom , title =. 2025 , howpublished =

    work page 2025

  47. [47]

    2025 , howpublished =

    Altman, Sam , title =. 2025 , howpublished =

    work page 2025

  48. [48]

    2025 , howpublished =

    Interview with. 2025 , howpublished =

    work page 2025

  49. [49]

    2025 , url =

    Bengio, Yoshua and others , title =. 2025 , url =

    work page 2025

  50. [50]

    2017 , url =

    Guidance on Scientific Research and Development (. 2017 , url =

    work page 2017

  51. [51]

    2025 , url =

    Commission Guidelines on the Definition of an. 2025 , url =

    work page 2025

  52. [52]

    2025 , url =

    Commission Guidelines on Prohibited. 2025 , url =

    work page 2025

  53. [53]

    2025 , url =

    Guidelines on the Scope of the Obligations for General-Purpose. 2025 , url =

    work page 2025

  54. [54]

    2022 , url =

    The `Blue Guide' on the Implementation of. 2022 , url =

    work page 2022

  55. [55]

    2025 , url =

    Code of Practice for General-Purpose. 2025 , url =

    work page 2025

  56. [56]

    2024 , howpublished =

    Regulation (. 2024 , howpublished =

    work page 2024

  57. [57]

    2016 , howpublished =

    Regulation (. 2016 , howpublished =

    work page 2016

  58. [58]

    2006 , howpublished =

    Regulation (. 2006 , howpublished =

    work page 2006

  59. [59]

    2016 , url =

    Consolidated Version of the Treaty on the Functioning of the. 2016 , url =

    work page 2016

  60. [60]

    , title =

    Trump, Donald J. , title =. 2025 , howpublished =

    work page 2025

  61. [61]

    , title =

    Biden, Joseph R. , title =. 2024 , howpublished =

    work page 2024

  62. [62]

    Senate Bill No. 53:. 2025 , url =

    work page 2025

  63. [63]

    The Nobel Prize in Chemistry 2024 , year =

    work page 2024

  64. [64]

    The Large Hadron Collider , year =

    work page

  65. [65]

    The Human Brain Project Ends: What Has Been Achieved? , year =

    work page

  66. [66]

    2023 , url =

    A Glimpse of the Next Generation of. 2023 , url =

    work page 2023