Pith. sign in

REVIEW 5 major objections 5 minor 48 references

The main barrier to AI adoption in the public sector is lack of training, not technology, according to two auditable Brazilian government cases.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 12:33 UTC pith:XVH5PCCG

load-bearing objection Real case data and a practical training method, but the central claim that training is the determining barrier to AI adoption is not supported by the two-unit before/after design, and the paper's own admitted confounds undermine the attribution. the 5 major comments →

arxiv 2606.01517 v4 pith:XVH5PCCG submitted 2026-06-01 cs.CY

The Main Barrier to AI Adoption in the Public Sector Is Lack of Training: How a Structured Method Accompanied Productivity Gains in Two Brazilian Government Cases

Vinicius Santana Gomes This is my paper
classification cs.CY
keywords Generative artificial intelligencePublic sectorAI governanceTraining methodologyInstitutional replicabilityData protectionProductivity measurementHuman-in-the-loop
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that the deciding obstacle to generative-AI value in the public sector is not the capability of the models but the absence of a structured way to train public servants to use them safely and methodically. It describes a four-layer training method—foundation, prompt technique, applied finishing, and governance roof—installed in two internal-control units of the Brazilian Federal District government. In the first unit, average processing time fell 18.2%; in the second, it fell 50%, while technical-report output rose 92% and cases totaling US$104.3 million were reviewed, with no information-security incident recorded. The author is explicit that this is observational evidence: the gains coincided with new unit leadership and workflow reorganization, so the method is not isolated from those factors. A sympathetic reader would care because the claim, if right, means public agencies can obtain large productivity gains from free AI simply by investing in structured staff training rather than buying technology.

Core claim

On the paper's own terms, the central discovery is that in two Brazilian public-sector internal-control units the determining barrier between AI availability and value delivery was training-related, not technological. The paper reports that after its four-layer AI House method was applied, official system indicators showed average processing time fell from 17.92 to 14.66 days (18.2%) at the health-sector control office and from 34 to 17 days (50%) at the economic-development internal-control unit. The second unit also raised technical-report output from 66 to 127 (92%), issued 288 formal recommendations, and reviewed US$104.3 million in case value, with no information-security incident ident

What carries the argument

The AI House is a four-layer training architecture for adopting generative AI in public administration: Foundation (what the model is and its limits), Walls (prompt technique, including the PACTO pattern—Persona, Action, Context, Tone, Observations, Example), Finishing (AI functions tied to the civil servant's actual documents), and Roof (data-protection law, procedural secrecy, administrative principles, and international AI governance). The method's core operational mechanism is a custom AI loaded with the unit's normative base, queried only with de-identified data, and always overseen by a human who signs the final document. This structure carries the argument: it separates what the tool

Load-bearing premise

The claim that the gains came from the training method rests on the assumption that the measured improvements are not instead explained by the new unit leadership and the workflow reorganization that happened at the same time, since the study has no control group or formal causal design.

What would settle it

A concrete falsifying observation would be a matched comparison in which one unit receives the four-layer training and a comparable unit does not, under the same leadership and workflow, with both measured on the same official processing-time indicator; if the trained unit does not show a meaningfully larger reduction, the central claim fails. A simpler check: apply the method in a third agency while keeping the same supervisor and unchanged workflows, and see whether the time-reduction pattern replicates.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If training is the binding constraint, then deploying more capable models without a structured method will not yield comparable gains; the bottleneck is operator repertoire.
  • Public agencies can pursue AI productivity gains without dedicated AI budgets, because the method uses free-tier versions and custom AIs built on free functionality.
  • The four-layer order matters: skipping a layer produces predictable failures—hallucinations, vague prompts, abstract knowledge, or compliance incidents.
  • The governance roof with de-identified queries and integral human review can allow sensitive public-sector work to use commercial AI without security incidents.
  • The observed gains concentrated in the document types the method targets, suggesting the mechanism is the method, not general tool improvement.

Where Pith is reading between the lines

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

  • Editorial extension: a testable next step would compare units trained with the same four-layer method against matched untrained units; the current design cannot separate training from concurrent management changes.
  • Editorial extension: if the training-based account is right, the same method should replicate in other government functions, such as procurement audits or policy analysis; a failed replication in a distinct agency would weigh against the claim.
  • Editorial extension: the US$104.3 million figure is 'value reviewed,' not savings; the paper's own modeled mitigation range (US$1.2M to $5.7M, central $3M) is based on literature-calibrated probabilities and non-binding recommendations, so the larger figure should not be treated as avoided cost.
  • Editorial extension: the 'no incidents' result rests on internal control mechanisms; independent external audit of logs and de-identification practices would strengthen this claim.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper argues that the main barrier to generative-AI adoption in the public sector is not technological but training-related, and describes a four-layer pedagogical method (the “AI House”). Evidence comes from two Brazilian government internal-control units: SES/CONT (Health, 2023→2024) and UCI/SEDET (Economic Development, 2024→2025). Official SEI-GDF statistics show reductions in average processing time of 18.2% and 50%, respectively, and at UCI a 92% increase in technical reports and US$104.3M in financial volume analyzed. The author designed the method, taught the course, led both units, and signed one of the two source reports; the paper transparently acknowledges this double role. The manuscript also explicitly states that the method was accompanied by new management and workflow reorganization, and that the study does not isolate the method from these co-occurring factors.

Significance. If the central attribution were valid, the paper would make a useful applied contribution: it provides a concrete, low-cost training method, demonstrates descriptively plausible portability across two distinct agencies, and documents official statistics with SEI document identifiers, reproducible data-extraction procedures, and explicit modeling assumptions. The paper is transparent about several limitations, including the non-binding nature of UCI recommendations and the modeled character of the savings estimate. However, the central claim—that lack of training was the determining barrier—is causal/attributive, and the design (uncontrolled two-unit before/after comparison, interventionist as evaluator, no control group, no mediator measurement) cannot support that claim. The contribution is therefore best read as a detailed case report, not as an empirical demonstration of the training-attribution thesis.

major comments (5)
  1. [Abstract and Section XI] The abstract states that the determining barrier to adoption was training-related, and Section XI concludes that the training-based account is the most parsimonious reading. Yet Section XI explicitly says: “We do not claim to isolate the method from these co-occurring factors.” With only two units, no control period, no control group, no pre-registered specification, and no measurement of the hypothesized mediator (e.g., AI usage, prompt quality, skill acquisition), the before/after SEI-GDF statistics cannot distinguish the training method from the concurrent leadership change, workflow reorganization, or intensified management attention. The paper's headline causal claim is therefore not established by its evidence.
  2. [Section VI / Table A.2.3] The normalized adjudication analysis excludes August 2023 as “atypical” in order to report a monthly average of 31.00 in 2024 versus 30.36 in 2023 (+2.1%). If August is included, the 2023 average is 33.58, which is higher than the 2024 value. The exclusion is documented, but it is post hoc and flatters the comparison; no pre-specified criterion or sensitivity analysis including August is provided. The stability claim should be re-run with and without August, and the results reported either way.
  3. [Section VI, “Estimated released capacity”] The arithmetic treats the drop in average processing time as though it were uniformly applicable released capacity: 1,581 cases × 17.92 days minus 1,581 × 14.66 days = 5,155 days ≈ 352 additional cases. But the case mix changed (cases processed fell from 1,752 to 1,581) and the largest time reductions concentrate in specific case types (Table A.2.2). Without case-mix adjustment or evidence that the 2023 average processing time applies to the 2024 caseload, the “released capacity” estimate is not identified and may simply reflect composition effects.
  4. [Sections I, VI, X and Appendix A.0] The author is simultaneously method designer, course instructor, unit manager, and signatory of the primary SES/CONT report; the UCI report was signed by the author's hierarchical superior. The statistical pipeline is “maintained by the author” and is only “available upon request.” While this entanglement is disclosed, it means the key productivity claims are not independently verified. Official SEI-GDF statistics are valuable descriptive evidence, but the absence of an independent audit or third-party analysis is a load-bearing weakness for a paper whose stated contribution rests on the empirical demonstration of a causal training effect.
  5. [Section X / Appendix A.4] The potential-mitigation estimate (US$1.2–5.7M) is clearly labeled as modeled, and the appendix honestly states that the probabilities are “judgment-based assumptions” rather than estimates from the cited literature. Still, the conclusion presents this modeled range as one of the paper's “four verifiable fronts.” The GAO and ACFE references document aggregate improper-payment and fraud-loss patterns; they do not provide an empirical basis for the probability that a specific UCI recommendation prevents a specific loss. The estimate should be moved from the conclusions to the limitations discussion, or substantially de-emphasized.
minor comments (5)
  1. [Abstract (submitted file vs. full text)] The abstract included in the submission materials states an 85% increase in technical reports, 286 recommendations, and US$94.8M; the full-text abstract states 92%, 288, and US$104.3M. Reconcile these numbers.
  2. [Section IV] The claim that applying the course outside the unit gives results “internal validity that ad-hoc training built within the unit could never have” is overstated. The external course setting does not address selection or confounding; it only reduces one specific training-delivery confound.
  3. [Appendix A.0] The currency conversion states the reference rate is R$5.00 = US$1.00, “corresponding to the commercial dollar quotation of May 14, 2026.” Please provide the source for this quotation and state whether the rate is a period average or a point-in-time value, as the model estimates are sensitive to this choice.
  4. [Tables A.2.4/A.3.4] Consider adding a 2022 baseline for UCI/SEDET, as is done for SES/CONT, to strengthen the descriptive before/method/baseline comparison. Also, the coefficient-of-variation table would benefit from confidence intervals or at least a statement that CVs are descriptive only.
  5. [General] The full Portuguese translation is included in the same document. If this is intentional, add a note on the title page explaining the dual-language format; if not, remove it to avoid confusion.

Circularity Check

0 steps flagged

No significant circularity: the paper's causal attribution is weak/observational, but no result is derived from its own inputs by construction.

full rationale

The paper's central claim—that the determining barrier was training-related and that the structured method accompanied productivity gains—is an interpretive, observational conclusion, not a derivation that reduces to its own inputs. The quantitative indicators come from SEI-GDF, an external official system, and the paper repeatedly labels the financial estimate as a modeled range based on judgment-based probability assumptions, explicitly stating: 'These values are not directly estimated from the literature; rather, they represent prudential, judgment-based assumptions.' The main weakness is causal identification: the author acknowledges co-occurring leadership change and workflow reorganization and writes 'We do not claim to isolate the method from these co-occurring factors.' That is an honest limitation of an uncontrolled before/after design, not a circular step. There is no fitted parameter renamed as a prediction, no self-citation chain used to force the conclusion, and no equation in which an outcome is definitionally equal to an input. The exclusion of August 2023 from the normalized adjudication comparison is a post-hoc analytical choice that could be criticized, but it does not make the reported average equal to the method's assumptions by construction. Overall, the paper's evidentiary weakness belongs to validity and generalizability, not to circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

No new physical or conceptual entities are introduced. The free parameters are the judgment-based probability and materiality values used only in the hypothetical-savings model (Appendix A.4), which is explicitly labeled as a modeled estimate. The central productivity claims do not use fitted parameters, but they rely heavily on unfalsifiable assumptions about the underlying administrative data and implementation fidelity.

free parameters (3)
  • Probability of risk materialization (Nature II, central scenario) = 5%
    Chosen by judgment, not estimated from data (Appendix A.4.5).
  • Probability of risk materialization (Nature III, central scenario) = 25%
    Judgment-based, consistent with literature ranges but not directly estimated (Appendix A.4.5).
  • Materiality modifiers = high=1.00, medium=0.75, low=0.50, null=0.00
    Arbitrary scaling factors used to adjust probabilities (Appendix A.4.4).
axioms (4)
  • domain assumption SEI-GDF statistics accurately reflect unit productivity and processing time.
    All quantitative claims rest on these official statistics; no independent verification is provided.
  • domain assumption The free-tier AI platforms and de-identification protocols were applied consistently as described.
    The paper asserts this, but there is no audit trail to confirm every interaction followed the protocol.
  • domain assumption Absence of reported incidents indicates absence of security or data-protection incidents.
    The paper treats 'no incidents identified by internal control mechanisms' as 'no incidents occurred' (Sections VI, X, XI).
  • domain assumption The same team composition and case mix between base year and intervention year allow a valid pre-post comparison.
    Stated in Section XI; but changes in leadership and workflow are acknowledged as co-occurring, so this assumption is fragile.

pith-pipeline@v1.3.0-alltime-deepseek · 47609 in / 8655 out tokens · 83563 ms · 2026-08-02T12:33:02.145530+00:00 · methodology

0 comments
read the original abstract

The adoption of generative AI in the public sector has been treated predominantly as a technological problem, with the expectation that productivity gains would follow from the availability of increasingly capable models. This paper argues, drawing on two auditable cases in the Brazilian Public Service, that the determining barrier to adoption observed in these units was not technological but training-related, and describes the four-layer structured pedagogical methodology developed by the author. The method was applied in two units with distinct institutional profiles: the Sectoral Internal Control Office of the Federal District Department of Health throughout 2024, and the Internal Control Unit of the Federal District Department of Economic Development, Labor and Income throughout 2025. In both cases, the official indicators from the Electronic Information System of the Federal District Government (SEI-GDF), verifiable by third parties, recorded gains that accompanied the method's rollout: average processing time fell by 18.2% at SES/CONT and by 50% at UCI/SEDET, with UCI also recording an 85% increase in technical-report production, the issuance of 286 formal recommendations to public managers, and the analysis of cases and matters whose total value, per the unit's own signed statistics, was US$ 94.8 million, the value of the matters submitted to technical analysis. The analysis is consistent with the hypothesis that the method is portable across agencies with distinct mandates, operates within protocols designed to comply with international and national data-protection law and with the principles of public administration, and is accessible to public entities under budget constraints, since it used free AI models.

Review history (2 revisions) →

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

48 extracted references · 2 linked inside Pith

  1. [1]

    Department of Health

    FEDERAL DISTRICT. Department of Health. Sectoral Internal Control Office (SES/CONT). Executive Management Report: Impact of Artificial Intelligence on Operational Efficiency. SEI Doc. No. 197403428, SEI case 00060-00582291/2024-11. Brasília, DF, January 2, 2025

    2024

  2. [2]

    Department of Economic Development, Labor and Income

    FEDERAL DISTRICT. Department of Economic Development, Labor and Income. Internal Control Unit (UCI/SEDET-DF). Report No. 1/2026: Management Results Report 2025 — UCI/GAB/SEDET. SEI Doc. No. 193384267, SEI case 04035-00000853/2026-24. Brasília, DF, January 28, 2026

    2026

  3. [3]

    Department of Economic Development, Labor and Income

    FEDERAL DISTRICT. Department of Economic Development, Labor and Income. Internal Control Unit (UCI/SEDET-DF). AI Governance Framework — UCI. SEI Doc. No. 194251158. Brasília, DF

  4. [4]

    Artificial Intelligence in the public sector: techniques, risks, and applications

    FEDERAL DISTRICT. Department of Health. Sectoral Internal Control Office — Advisory Office for the Adjudication of Administrative Proceedings (SES/CONT/ASJULG). Memorandum No. 4/2024 — Proposal for the creation of the course “Artificial Intelligence in the public sector: techniques, risks, and applications”. SEI Doc. No. 146621014, SEI case 00060-00314546...

    2024

  5. [5]

    Artificial Intelligence in the public sector: techniques, risks, and applications

    FEDERAL DISTRICT. Executive Secretariat of the Civil House — Sub-secretariat of Administrative Management — School of Government (SEEC/SEGEA/EGOV). Memorandum No. 166/2024 — Institutional approval of the course “Artificial Intelligence in the public sector: techniques, risks, and applications”. SEI Doc. No. 146526272, SEI case 04044-00021535/2024-26. Sign...

    2024

  6. [6]

    Department of Health

    FEDERAL DISTRICT. Department of Health. Sectoral Internal Control Office — Advisory Office for the Adjudication of Administrative Proceedings (SES/CONT/ASJULG). Report No. 12/2024 — Institutional record of the clearance of the unit’s case backlog. Brasília, DF, December 14, 2024. Brazilian Legislation

    2024

  7. [7]

    BRAZIL. Law No. 13.709, of August 14, 2018. Brazilian General Data Protection Law (LGPD). Official Gazette of the Union, Brasília, DF. Public Auditing

    2018

  8. [8]

    GOVERNMENT ACCOUNTABILITY OFFICE (GAO)

    U.S. GOVERNMENT ACCOUNTABILITY OFFICE (GAO). Improper Payments: Information on Agencies’ Fiscal Year 2024 Estimates. GAO-25-107753. Washington, DC: GAO, 2025. Available at: https://www.gao.gov/products/gao-25-107753

    2024

  9. [9]

    GOVERNMENT ACCOUNTABILITY OFFICE (GAO)

    U.S. GOVERNMENT ACCOUNTABILITY OFFICE (GAO). Payment Integrity: Agencies’ Estimated Improper Payments Increased to $186 Billion in Fiscal Year 2025. GAO-26-108694. Washington, DC: GAO,

    2025

  10. [10]

    Occupational Fraud 2024: A Report to the Nations

    ASSOCIATION OF CERTIFIED FRAUD EXAMINERS (ACFE). Occupational Fraud 2024: A Report to the Nations. Austin, TX: ACFE, 2024. Available at: https://www.acfe.com/-/media/files/acfe/pdfs/rtt n/2024/2024-report-to-the-nations.pdf. International AI Governance Frameworks

    2024

  11. [11]

    Artificial Intelligence Risk Management Framework (AI RMF 1.0)

    NATIONAL INSTITUTE OF STANDARDS AND TECHNOLOGY (NIST). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. Gaithersburg, MD: U.S. Department of Commerce,

  12. [12]

    GOVERNMENT ACCOUNTABILITY OFFICE (GAO)

    U.S. GOVERNMENT ACCOUNTABILITY OFFICE (GAO). Artificial Intelligence: An Accountability Framework for Federal Agencies and Other Entities. GAO-21-519SP. Washington, DC: GAO, 2021. Available at: https://www.gao.gov/products/gao-21-519sp

    2021

  13. [13]

    Regulation (EU) 2024/1689 of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act)

    EUROPEAN PARLIAMENT AND COUNCIL OF THE EUROPEAN UNION. Regulation (EU) 2024/1689 of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union, Brussels: European Union, 2024. Available at: https://eur- lex.europa.eu/eli/reg/2024/1689/oj

    2024

  14. [14]

    Recommen- dation of the Council on Artificial Intelligence

    ORGANISATION FOR ECONOMIC CO-OPERATION AND DEVELOPMENT (OECD). Recommen- dation of the Council on Artificial Intelligence. OECD/LEGAL/0449. Paris: OECD, adopted 22 May 2019, amended 8 November 2023 and 3 May 2024. Available at: https://oecd.ai/en/ai-principles. Methodological References

    2019

  15. [19]

    Most eGovernment-for-development projects fail: How can risks be reduced? Manch- ester: Institute for Development Policy and Management, University of Manchester, 2003

    HEEKS, Richard. Most eGovernment-for-development projects fail: How can risks be reduced? Manch- ester: Institute for Development Policy and Management, University of Manchester, 2003. Available at: https://ssrn.com/abstract=3540052. 29

    2003

  16. [24]

    Valor explícitomencionado na recomendação (glosa específica, devolução, ressarcimento quantificado): usa esse valor

  17. [25]

    Vício que afeta a validade do ato(certidão expirada, ausência de empenho, divergência atesto×NF, ausência de designação formal de fiscal por ato publicado, cláusula contratual em desacordo com edital, ausência de cláusula obrigatória, descumprimento de exigência legal de habilitação): usa o valor principal do pagamento/contrato em análise

  18. [26]

    Divergência quantitativa sem valor especificado:usa o valor principal como teto máximo de exposição, com justificativa explícita de que o risco real pode ser inferior

  19. [27]

    vício acessório à recomendação X.Y, evitando dupla contagem

    Recomendação procedimental pura(padronização, detalhamento, melhoria de relatório, designação preventiva para futuras contratações): valor zero. 5.Recomendação de melhoria futura, vigilância ou manutençãode boas práticas: valor zero. 6.Observância genérica de legalidadesem apontamento de falha específica: valor zero. Adicionalmente, quando múltiplas recom...

    2024

  20. [28]

    A estimativa quantifica potencial de mitigação, não realização documentada

    Caráter não-vinculante das recomendações UCI.A UCI emite controle prévio orientativo; não há registro sistemático de quais recomendações foram efetivamente acatadas pelos gestores. A estimativa quantifica potencial de mitigação, não realização documentada

  21. [29]

    As probabilidades aplicadas são derivadas de literatura internacional, não calibradas com dados específicos do GDF

    Counterfactual não observável.Não é possível estabelecer com certeza o cenário alternativo sem controle UCI. As probabilidades aplicadas são derivadas de literatura internacional, não calibradas com dados específicos do GDF

  22. [30]

    Subestimação por classificação conservadora.Recomendações classificadas como Natureza I (procedimental) recebem probabilidade zero, mesmo quando podem prevenir perdas indiretas, opção deliberada por rigor metodológico

  23. [31]

    Benefícios não-monetários não capturados.Ganhos em transparência, conformidade processual, padronização documental e capacidade técnica da equipe não são quantificados pela estimativa financeira

  24. [32]

    Referências Documentos Institucionais SEI-GDF

    Critério moderado é interpretativo.A atribuição do valor inteiro do pagamento/contrato a vícios de validade é defensável, mas representa uma escolha metodológica que poderia ser apresentada de forma mais ou menos generosa em interpretações alternativas. Referências Documentos Institucionais SEI-GDF

  25. [33]

    Secretaria de Estado de Saúde

    DISTRITO FEDERAL. Secretaria de Estado de Saúde. Controladoria Setorial da Saúde (SES/CONT). Relatório Gerencial Executivo: Impacto da Inteligência Artificial na Eficiência Operacional. Doc. SEI nº197403428, Processo SEI 00060-00582291/2024-11. Brasília, DF, 2 de janeiro de 2025

    2024

  26. [34]

    Secretaria de Estado de Desenvolvimento Econômico, Trabalho e Renda

    DISTRITO FEDERAL. Secretaria de Estado de Desenvolvimento Econômico, Trabalho e Renda. Unidade de Controle Interno (UCI/SEDET-DF).Relatório Nº1/2026: Relatório Gerencial de Resultados 2025 — UCI/GAB/SEDET.Doc. SEI nº193384267, Processo SEI 04035-00000853/2026-24. Brasília, DF, 28 de janeiro de 2026

    2026

  27. [35]

    Secretaria de Estado de Desenvolvimento Econômico, Trabalho e Renda

    DISTRITO FEDERAL. Secretaria de Estado de Desenvolvimento Econômico, Trabalho e Renda. Unidade de Controle Interno (UCI/SEDET-DF).Framework de Governança de IA — UCI.Doc. SEI nº 194251158. Brasília, DF

  28. [36]

    Inteligência Artificial no setor público: técnicas, riscos e aplicações

    DISTRITO FEDERAL. Secretaria de Estado de Saúde. Controladoria Setorial da Saúde — Assessoria de Apoio aos Julgamentos de Processos Administrativos (SES/CONT/ASJULG).Memorando Nº4/2024 — Proposta de criação do curso "Inteligência Artificial no setor público: técnicas, riscos e aplicações".Doc. SEI nº146621014, Processo SEI 00060-00314546/2024-15. Signatár...

    2024

  29. [37]

    Inteligência Artificial no setor público: técnicas, riscos e aplicações

    DISTRITO FEDERAL. Secretaria Executiva da Casa Civil — Subsecretaria de Gestão Administrativa — Escola de Governo (SEEC/SEGEA/EGOV).Memorando Nº166/2024 — Aprovação institucional do curso "Inteligência Artificial no setor público: técnicas, riscos e aplicações".Doc. SEI nº 146526272, Processo SEI 04044-00021535/2024-26. Signatária: Raquel Aben Athar de So...

    2024

  30. [38]

    Secretaria de Estado de Saúde

    DISTRITO FEDERAL. Secretaria de Estado de Saúde. Controladoria Setorial da Saúde — Assessoria de Apoio aos Julgamentos de Processos Administrativos (SES/CONT/ASJULG).Relatório Nº12/2024 — Registro institucional de zeração do passivo de processos da unidade.Brasília, DF, 14 de dezembro de 2024. Normativas Brasileiras

    2024

  31. [39]

    Diário Oficial da União, Brasília, DF

    BRASIL.Lei nº13.709, de 14 de agosto de 2018.Lei Geral de Proteção de Dados Pessoais (LGPD). Diário Oficial da União, Brasília, DF. 57 Auditoria Pública

    2018

  32. [40]

    GOVERNMENT ACCOUNTABILITY OFFICE (GAO).Improper Payments: Information on Agencies’ Fiscal Year 2024 Estimates.GAO-25-107753

    U.S. GOVERNMENT ACCOUNTABILITY OFFICE (GAO).Improper Payments: Information on Agencies’ Fiscal Year 2024 Estimates.GAO-25-107753. Washington, DC: GAO, 2025. Disponível em: https://www.gao.gov/products/gao-25-107753

    2024

  33. [41]

    GOVERNMENT ACCOUNTABILITY OFFICE (GAO).Payment Integrity: Agencies’ Es- timated Improper Payments Increased to $186 Billion in Fiscal Year 2025.GAO-26-108694

    U.S. GOVERNMENT ACCOUNTABILITY OFFICE (GAO).Payment Integrity: Agencies’ Es- timated Improper Payments Increased to $186 Billion in Fiscal Year 2025.GAO-26-108694. Washington, DC: GAO, 2026. Disponível em: https://www.gao.gov/products/gao-26-108694

    2025

  34. [42]

    Disponível em: https://www.acfe.com/- /media/files/acfe/pdfs/rttn/2024/2024-report-to-the-nations.pdf

    ASSOCIATION OF CERTIFIED FRAUD EXAMINERS (ACFE).Occupational Fraud 2024: A Report to the Nations.Austin, TX: ACFE, 2024. Disponível em: https://www.acfe.com/- /media/files/acfe/pdfs/rttn/2024/2024-report-to-the-nations.pdf. Frameworks Internacionais de Governança de Inteligência Artificial

    2024

  35. [43]

    Gaithersburg, MD: U.S

    NATIONAL INSTITUTE OF STANDARDS AND TECHNOLOGY (NIST).Artificial Intelligence Risk Management Framework (AI RMF 1.0).NIST AI 100-1. Gaithersburg, MD: U.S. Department of Commerce, 2023. Disponível em: https://www.nist.gov/itl/ai-risk-management-framework

    2023

  36. [44]

    GOVERNMENT ACCOUNTABILITY OFFICE (GAO).Artificial Intelligence: An Account- ability Framework for Federal Agencies and Other Entities.GAO-21-519SP

    U.S. GOVERNMENT ACCOUNTABILITY OFFICE (GAO).Artificial Intelligence: An Account- ability Framework for Federal Agencies and Other Entities.GAO-21-519SP. Washington, DC: GAO,

  37. [46]

    Disponível em: https://eur-lex.europa.eu/eli/reg/2024/1689/oj

    EUROPEAN PARLIAMENT AND COUNCIL OF THE EUROPEAN UNION.Regulation (EU) 2024/1689 of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act).Official Journal of the European Union, Brussels: European Union, 2024. Disponível em: https://eur-lex.europa.eu/eli/reg/2024/1689/oj

    2024

  38. [47]

    Paris: OECD, adopted 22 May 2019, amended 8 November 2023 and 3 May 2024

    ORGANISATION FOR ECONOMIC CO-OPERATION AND DEVELOPMENT (OECD).Recom- mendation of the Council on Artificial Intelligence.OECD/LEGAL/0449. Paris: OECD, adopted 22 May 2019, amended 8 November 2023 and 3 May 2024. Disponível em: https://oecd.ai/en/ai-principles. Referências Metodológicas

    2019

  39. [48]

    Disponível em: https://arxiv.org/abs/2201.11903 (DOI: https://doi.org/10.48550/arXiv.2201.11903)

    WEI, Jason; WANG, Xuezhi; SCHUURMANS, Dale; BOSMA, Maarten; ICHTER, Brian; XIA, Fei; CHI, Ed H.; LE, Quoc V.; ZHOU, Denny.Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.Advances in Neural Information Processing Systems (NeurIPS), 2022. Disponível em: https://arxiv.org/abs/2201.11903 (DOI: https://doi.org/10.48550/arXiv.2201.11903)

    Pith/arXiv arXiv doi:10.48550/arxiv.2201.11903 2022

  40. [49]

    Disponível em: https://arxiv.org/abs/2005.14165 (DOI: https://doi.org/10.48550/arXiv.2005.14165)

    BROWN, Tom B.; MANN, Benjamin; RYDER, Nick et al.Language Models are Few-Shot Learners.Advances in Neural Information Processing Systems (NeurIPS), 2020. Disponível em: https://arxiv.org/abs/2005.14165 (DOI: https://doi.org/10.48550/arXiv.2005.14165). Adoção Tecnológica e Capacitação em IA

    Pith/arXiv arXiv doi:10.48550/arxiv.2005.14165 2020

  41. [50]

    COHEN, Wesley M.; LEVINTHAL, Daniel A.Absorptive Capacity: A New Perspective on Learning and Innovation.Administrative Science Quarterly, v. 35, n. 1, p. 128–152, 1990. Disponível em: https://doi.org/10.2307/2393553

    doi:10.2307/2393553 1990

  42. [51]

    DUNLEAVY, Patrick; MARGETTS, Helen; BASTOW, Simon; TINKLER, Jane.New Public Management Is Dead — Long Live Digital-Era Governance.Journal of Public Administration Research and Theory, v. 16, n. 3, p. 467–494, 2006. Disponível em: https://doi.org/10.1093/jopart/mui057

    doi:10.1093/jopart/mui057 2006

  43. [52]

    Disponível em: https://ssrn.com/abstract=3540052

    HEEKS, Richard.Most eGovernment-for-development projects fail: How can risks be reduced?Manchester: Institute for Development Policy and Management, University of Manchester, 2003. Disponível em: https://ssrn.com/abstract=3540052

    2003

  44. [53]

    Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, p

    LONG, Duri; MAGERKO, Brian.What is AI literacy? Competencies and design considerations. Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, p. 1–16, 2020. Disponível em: https://doi.org/10.1145/3313831.3376727. 58

    arXiv 2020

  45. [2021]

    Disponível em: https://www.gao.gov/products/gao-21-519sp

  46. [2022]

    IA para gestores

    caracterizam padrões técnicos como formulação de prompts, few-shot prompting e chain-of-thought 32 prompting. Essas contribuições, no entanto, tratam o usuário em abstrato e não abordam o regime jurídico específico em que ele opera. O método aqui descrito integra esses elementos técnicos a uma camada de governança jurídica embutida desde a primeira aula, ...

    2023

  47. [2023]

    Available at: https://www.nist.gov/itl/ai-risk-management-framework

  48. [2026]

    Available at: https://www.gao.gov/products/gao-26-108694