pith. sign in

arxiv: 2605.22650 · v1 · pith:5SRDIWKYnew · submitted 2026-05-21 · 💻 cs.CL

Whose Voice Counts? Mapping Stakeholder Perspectives on AI Through Public Submissions to the U.S. Government

Alina Karakanta , Alex Christiansen , Tom\'as Dodds , Bissie Anderson , Matteo Fuoli , Marcus Perlman , Aletta G. Dorst This is my paper

Pith reviewed 2026-05-22 05:42 UTC · model grok-4.3

classification 💻 cs.CL
keywords AI policypublic submissionsstakeholder perspectivestopic modelingUS AI Action Planindividual concernsprivate sectorpolicy alignment
0
0 comments X

The pith

Public submissions to the US AI Action Plan show individuals focus on AI's effects on daily life while businesses and academics emphasize development and security, with the plan aligning more with private sector views.

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

The paper examines letters submitted during the public consultation for the Trump Administration's US AI Action Plan to map what different groups say about AI. It applies topic modeling and frequency analysis to a cleaned corpus of submissions from individuals, academia, and the private sector, then compares those topics to the content of the final plan. Results indicate that individuals raise strong concerns about AI's impact on life, while other stakeholders prioritize development, security, and policies. This matters because it reveals gaps in whose input shapes national AI strategy. Readers care since AI policies affect everyone yet may underweight everyday perspectives.

Core claim

The analysis of public comments on the US AI Action Plan shows that individuals voice strong concerns related to the impact of AI on life, while other stakeholders are more concerned with AI development. The comparison of topics suggests that the AI Action Plan reflects predominantly the concerns of the private sector on security, policies, and development, with individuals' concerns less represented.

What carries the argument

Topic modeling on submissions grouped by stakeholder category (individuals, academia, private sector) compared against topics extracted from the AI Action Plan document itself.

If this is right

  • The AI Action Plan incorporates more private sector priorities on security and development than individual concerns about life impacts.
  • Different stakeholder groups raise distinct topics, with individuals less focused on technical development.
  • Public consultations may underrepresent citizen views on AI's societal effects in final policy documents.
  • Private sector input appears to have stronger influence on the plan's emphasis areas.

Where Pith is reading between the lines

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

  • Consultation processes could add targeted outreach to amplify individual voices in future AI policy.
  • Applying similar topic comparisons to other national AI strategies might show consistent patterns of stakeholder weighting.
  • If citizen concerns remain underrepresented, it could reduce public trust in how AI rules are set.
  • Releasing cleaned submission data allows others to test alternative groupings or modeling choices.

Load-bearing premise

The classification of submitters into subgroups and the topics identified from the modeling accurately capture distinct stakeholder perspectives without significant bias or mislabeling.

What would settle it

A manual coding of a sample of submissions that finds individuals discuss development and security topics at rates similar to the private sector, or a revised plan text that incorporates more life-impact themes than development ones.

Figures

Figures reproduced from arXiv: 2605.22650 by Aletta G. Dorst, Alex Christiansen, Alina Karakanta, Bissie Anderson, Marcus Perlman, Matteo Fuoli, Tom\'as Dodds.

Figure 1
Figure 1. Figure 1: Topics per class for the REST of the respondent types. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The frequency by which the word representation of the top 10 topics per subcorpora group (IND [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
read the original abstract

As artificial intelligence (AI) systems become more common in our daily lives, it is important to understand how different stakeholders comprehend and envisage the role that these technologies play in shaping social, political, and economic realities. In this paper, we investigate public perceptions of AI based on a corpus of letters submitted during the public consultation for the Trump Administration's US AI Action Plan. To this aim, we release a corpus cleaning pipeline and perform topic modelling and frequency analysis to explore predominant topics discussed by different subgroups (e.g., academia, individuals, private sector) and those appearing in the AI Action Plan. Our results show that individuals voice strong concerns related to the impact of AI on life, while other stakeholders are more concerned with AI development. Our comparison of topics suggests that the AI Action Plan reflects predominantly the concerns of the private sector on security, policies, and development, with individuals' concerns less represented.

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

3 major / 2 minor

Summary. The paper analyzes a corpus of public submissions to the U.S. government's consultation on the Trump Administration's AI Action Plan. It releases a corpus cleaning pipeline and applies topic modeling together with frequency analysis to compare predominant topics across stakeholder subgroups (academia, individuals, private sector) and to contrast those topics with the content of the AI Action Plan itself. The central claim is that individuals emphasize AI's impact on daily life while other stakeholders focus on development; the Action Plan is argued to align predominantly with private-sector concerns on security, policies, and development, leaving individual perspectives underrepresented.

Significance. If the subgroup classifications and topic interpretations prove robust, the work supplies an empirical mapping of stakeholder voices in a high-stakes AI policy consultation and quantifies potential gaps between public input and final policy language. The release of the cleaning pipeline constitutes a concrete contribution to reproducibility in computational social-science analyses of government consultations.

major comments (3)
  1. [Methods] Methods / Data Processing: The assignment of submissions to stakeholder categories (academia, individuals, private sector) is described without the explicit rules, keyword heuristics, or self-report fields employed, and without any validation metrics such as inter-rater agreement, manual audit of a labeled sample, or error-rate estimates. Because the headline contrast between individuals' life-impact concerns and private-sector development concerns rests directly on these labels, even modest misclassification rates would alter the frequency comparisons and the claim that the Action Plan under-represents individuals.
  2. [Results] Topic Modeling and Results: No coherence scores, held-out perplexity, or human topic-quality evaluation are reported for the chosen number of topics, nor are statistical tests (chi-square, permutation, or regression) provided for differences in topic prevalence across subgroups. Without these, the assertions that individuals voice 'strong concerns' on life impact and that the Action Plan 'reflects predominantly' private-sector topics remain descriptive rather than inferentially supported.
  3. [Discussion] Comparison to AI Action Plan: The procedure for extracting or aligning topics from the Action Plan document itself is not detailed (e.g., whether the same topic model was applied, how document-level topic proportions were computed, or what similarity metric was used). This step is load-bearing for the conclusion that the Plan aligns more with private-sector than individual concerns.
minor comments (2)
  1. [Abstract] The abstract states that a 'corpus cleaning pipeline' is released but does not indicate the repository or license; this information should appear in the abstract or a prominent footnote.
  2. [Figures] Figure captions and axis labels for any topic-frequency plots should explicitly state the topic-model algorithm, number of topics, and whether frequencies are normalized or raw counts.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed comments, which help clarify how to strengthen the transparency and rigor of our analysis of public submissions to the U.S. AI Action Plan consultation. We address each major comment below and indicate the revisions we will make.

read point-by-point responses

  1. Referee: [Methods] Methods / Data Processing: The assignment of submissions to stakeholder categories (academia, individuals, private sector) is described without the explicit rules, keyword heuristics, or self-report fields employed, and without any validation metrics such as inter-rater agreement, manual audit of a labeled sample, or error-rate estimates. Because the headline contrast between individuals' life-impact concerns and private-sector development concerns rests directly on these labels, even modest misclassification rates would alter the frequency comparisons and the claim that the Action Plan under-represents individuals.

    Authors: We agree that the stakeholder classification procedure requires more explicit documentation to support the validity of our subgroup comparisons. In the revised manuscript, we will provide the complete set of keyword heuristics, self-report fields from the submissions, and any additional rules used to assign categories (academia, individuals, private sector). We will also report the results of a manual audit performed on a random sample of 200 submissions, including inter-rater agreement metrics such as Cohen's kappa and estimated error rates. These additions will allow readers to evaluate potential misclassification effects on the reported frequency differences. revision: yes

  2. Referee: [Results] Topic Modeling and Results: No coherence scores, held-out perplexity, or human topic-quality evaluation are reported for the chosen number of topics, nor are statistical tests (chi-square, permutation, or regression) provided for differences in topic prevalence across subgroups. Without these, the assertions that individuals voice 'strong concerns' on life impact and that the Action Plan 'reflects predominantly' private-sector topics remain descriptive rather than inferentially supported.

    Authors: We acknowledge the benefit of including quantitative validation and statistical support. In the revision, we will report topic coherence using Normalized Pointwise Mutual Information (NPMI) and held-out perplexity for the selected number of topics, along with a brief human evaluation of topic interpretability on a sample of top words and documents. Additionally, we will apply and report chi-square tests to evaluate the statistical significance of differences in topic prevalence across stakeholder subgroups. These changes will move the analysis from purely descriptive to including inferential evidence for the observed contrasts. revision: yes

  3. Referee: [Discussion] Comparison to AI Action Plan: The procedure for extracting or aligning topics from the Action Plan document itself is not detailed (e.g., whether the same topic model was applied, how document-level topic proportions were computed, or what similarity metric was used). This step is load-bearing for the conclusion that the Plan aligns more with private-sector than individual concerns.

    Authors: We will provide a detailed account of the alignment procedure in the revised manuscript. The same LDA topic model trained on the public submissions was applied to the full text of the AI Action Plan to derive document-level topic proportions via the model's inference mechanism. We will specify the exact method for computing these proportions and the cosine similarity metric used to compare the resulting topic distribution against the average distributions from each stakeholder subgroup. This will make the comparison fully transparent and reproducible. revision: yes

Circularity Check

0 steps flagged

No circularity: observational analysis of external public submissions with independent methodological steps

full rationale

The paper applies standard corpus cleaning, topic modeling, and frequency analysis to an external corpus of public letters submitted to the U.S. government. No equations, fitted parameters, or predictions are defined in terms of the target results. Submitter classification into stakeholder groups and topic interpretation are one-time methodological choices applied to independent data rather than self-referential definitions or fits that reduce the headline claims to inputs by construction. Self-citations, if present, are not load-bearing for the central comparison of stakeholder concerns versus the AI Action Plan. The work is self-contained against external benchmarks and exhibits no enumerated circularity patterns.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central comparison depends on accurate stakeholder classification and topic interpretability; these are standard domain assumptions in NLP policy analysis rather than new inventions.

free parameters (1)
  • number of topics in model
    Hyperparameter in topic modeling that influences which predominant topics are extracted for subgroup comparison.
axioms (2)
  • domain assumption Submitters can be reliably grouped into categories such as individuals, academia, and private sector based on available metadata or text.
    Used to segment the corpus for differential topic analysis.
  • domain assumption Topic modeling outputs correspond to meaningful and distinct stakeholder concerns.
    Required to interpret frequency differences and compare to the Action Plan.

pith-pipeline@v0.9.0 · 5708 in / 1283 out tokens · 37407 ms · 2026-05-22T05:42:22.630250+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.

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.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages · 1 internal anchor

  1. [1]

    Catalan Speecon database

    Speecon Consortium. Catalan Speecon database. 2011

    work page 2011

  2. [2]

    The EMILLE/CIIL Corpus

    Anthony McEnery and others. The EMILLE/CIIL Corpus. 2004

    work page 2004

  3. [3]

    The OrienTel Moroccan MCA (Modern Colloquial Arabic) database

    Khalid Choukri and Niklas Paullson. The OrienTel Moroccan MCA (Modern Colloquial Arabic) database. 2004

    work page 2004

  4. [4]

    ItalWordNet v.2

    Roventini, Adriana and Marinelli, Rita and Bertagna, Francesca. ItalWordNet v.2

    work page

  5. [5]

    Comments Received in Response to: Request for Information on the Development of an Artificial Intelligence (AI) Action Plan (“Plan”) , year =

    work page

  6. [6]

    Removing Barriers to American Leadership in Artificial Intelligence , year =

    work page

  7. [7]

    Available at SSRN 3312874 , year=

    Artificial intelligence: American attitudes and trends , author=. Available at SSRN 3312874 , year=

    work page

  8. [8]

    Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society , pages=

    US public opinion on the governance of artificial intelligence , author=. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society , pages=

    work page

  9. [9]

    Pew Research Center: Washington, DC, USA , year=

    How Americans think about artificial intelligence , author=. Pew Research Center: Washington, DC, USA , year=

    work page

  10. [10]

    Public First , year=

    What does the public think about AI , author=. Public First , year=

    work page

  11. [11]

    International Journal of Human--Computer Interaction , volume=

    The General Attitudes towards Artificial Intelligence Scale (GAAIS): Confirmatory validation and associations with personality, corporate distrust, and general trust , author=. International Journal of Human--Computer Interaction , volume=. 2023 , publisher=

    work page 2023

  12. [12]

    Scientific Reports , volume=

    Attitudes towards AI: measurement and associations with personality , author=. Scientific Reports , volume=. 2024 , publisher=

    work page 2024

  13. [13]

    Digital health , volume=

    Attitudes and perception of artificial intelligence in healthcare: a cross-sectional survey among patients , author=. Digital health , volume=. 2022 , publisher=

    work page 2022

  14. [14]

    Computers in Human Behavior , volume=

    Whose AI? How different publics think about AI and its social impacts , author=. Computers in Human Behavior , volume=. 2022 , publisher=

    work page 2022

  15. [15]

    International Journal of Human--Computer Interaction , volume=

    The artificial intelligence paradox: Opportunity or threat for humanity? , author=. International Journal of Human--Computer Interaction , volume=. 2025 , publisher=

    work page 2025

  16. [16]

    Journalism & Mass Communication Quarterly , volume=

    In AI we trust: The interplay of media use, political ideology, and trust in shaping emerging AI attitudes , author=. Journalism & Mass Communication Quarterly , volume=. 2025 , publisher=

    work page 2025

  17. [17]

    Omics: a journal of integrative biology , volume=

    Sentiment analysis of the news media on artificial intelligence does not support claims of negative bias against artificial intelligence , author=. Omics: a journal of integrative biology , volume=. 2020 , publisher=

    work page 2020

  18. [18]

    AI & society , volume=

    Public understanding of artificial intelligence through entertainment media , author=. AI & society , volume=. 2024 , publisher=

    work page 2024

  19. [19]

    Behaviour & Information Technology , volume=

    Effects of country and individual factors on public acceptance of artificial intelligence and robotics technologies: a multilevel SEM analysis of 28-country survey data , author=. Behaviour & Information Technology , volume=. 2022 , publisher=

    work page 2022

  20. [20]

    EPJ Data Science , volume=

    Public perception of generative AI on Twitter: an empirical study based on occupation and usage , author=. EPJ Data Science , volume=. 2024 , publisher=

    work page 2024

  21. [21]

    Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society , pages=

    An AI race for strategic advantage: rhetoric and risks , author=. Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society , pages=

    work page 2018

  22. [22]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Long-term trends in the public perception of artificial intelligence , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

    work page

  23. [23]

    Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society , pages=

    Framing artificial intelligence in American newspapers , author=. Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society , pages=

    work page 2019

  24. [24]

    Telematics and Informatics , volume=

    Global news media coverage of artificial intelligence (AI): A comparative analysis of frames, sentiments, and trends across 12 countries , author=. Telematics and Informatics , volume=. 2025 , publisher=

    work page 2025

  25. [25]

    AI & society , volume=

    AI in the headlines: the portrayal of the ethical issues of artificial intelligence in the media , author=. AI & society , volume=. 2020 , publisher=

    work page 2020

  26. [26]

    Journalism , volume=

    Where exactly between utopia and dystopia? A framing analysis of AI and automation in US newspapers , author=. Journalism , volume=. 2024 , publisher=

    work page 2024

  27. [27]

    AI & SOCIETY , pages=

    Old wine in new bottles: shifting to flexible regulatory approaches for generative AI , author=. AI & SOCIETY , pages=. 2025 , publisher=

    work page 2025

  28. [28]

    , booktitle =

    Qi, Peng and Zhang, Yuhao and Zhang, Yuhui and Bolton, Jason and Manning, Christopher D. , booktitle =. Stanza: A

    work page

  29. [29]

    BERTopic: Neural topic modeling with a class-based TF-IDF procedure

    BERTopic: Neural topic modeling with a class-based TF-IDF procedure , author=. arXiv preprint arXiv:2203.05794 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  30. [30]

    Pacific-Asia conference on knowledge discovery and data mining , pages=

    Density-based clustering based on hierarchical density estimates , author=. Pacific-Asia conference on knowledge discovery and data mining , pages=. 2013 , organization=

    work page 2013

  31. [31]

    Lexicography , volume=

    The sketch engine , author=. Lexicography , volume=. 2014 , publisher=

    work page 2014