Bottom-Up Perspectives on AI Governance: Insights from User Reviews of AI Products

Stefan Pasch

REVIEW 4 major objections 6 minor 12 references

Bottom-Up Perspectives on AI Governance: Insights from User Reviews of AI Products

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Analyzing 108,998 user reviews of AI products shows governance concerns extend well beyond privacy and fairness into project management, strategy, and customer interaction.

desk verdict A transparent, large-scale topic-modeling study that surfaces plausibly governance-relevant operational themes from user reviews, but the relevance filter has not yet escaped generic management semantics. read the letter →

arxiv 2506.00080 v1 pith:BNSINYKI submitted 2025-05-30 cs.CY cs.AIcs.CLcs.HC

Stefan Pasch This is my paper

classification cs.CYcs.AIcs.CLcs.HC

keywords AIgovernancebottom-upuserreviewsBERTopictopicmodelingsemanticsimilarityG2.comethical

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that AI governance frameworks, built top-down by policymakers and experts, miss what users actually experience with AI systems. To test this, the author analyzes 108,998 G2.com reviews of 249 AI products with BERTopic, then selects topics whose averaged keyword embeddings are semantically close to either the EU's seven 'Trustworthy AI' principles or high-level governance terms. The selected topics cluster into themes that span technical and non-technical, internal and external domains. Some themes, such as privacy, security, and permissions, match institutional frameworks, but others—project management, product strategy, customer support, and communication—are largely absent from formal guidelines. If this holds, user discourse is a valid bottom-up source of governance signals that should complement top-down principle-based frameworks.

What carries the argument

The machinery is a two-stage semantic pipeline: BERTopic extracts latent topics from the reviews; then each topic's averaged top-10 keyword embedding (all-mpnet-base-v2) is compared by cosine similarity against two 'master texts'—one built from the EU Ethics Guidelines' seven principles, the other from the broad terms 'governance', 'ethics', 'trustworthiness'. The largest relative drop in similarity scores sets the cutoff to select governance-relevant topics, and K-means (k=7) organizes them into thematic clusters. The paper also constructs a 2x2 matrix (technical/non-technical by internal/external) as an interpretive lens, treating governance as a distributed sociotechnical process rather than a checklist.

What would settle it

Take a random sample of reviews belonging to the cosine-selected governance topics and have annotators label whether each review actually raises a governance concern; if human raters fail to agree with the similarity-based selection, the semantic-similarity proxy is not tracking governance relevance.

Watch

Extended reading notes

Core claim

The central claim is that governance-relevant themes in user discourse form a wider and more operational set than institutional principles capture. Using topic modeling on user reviews, the paper identifies a bottom-up governance map organized along two axes—technical versus non-technical and internal versus external—that includes deployment infrastructure, permissions and roles, data and analytics, HR, recruitment, project management, strategy and roadmaps, legal compliance, and customer communication. The author argues these themes reflect governance as enacted in practice: users voice accountability concerns through admin rights and access control, and transparency concerns through dashboards and reporting, rather than through the abstract vocabulary of ethics guidelines. Conversely, high-level ideals such as environmental sustainability and societal well-being rarely appear in reviews, not because users dismiss them but because users evaluate tools on task performance and organizational fit.

Load-bearing premise

The load-bearing assumption is that cosine similarity between a topic's averaged keyword embedding and a short seed list (such as the seven EU principles or the words 'governance', 'ethics', 'trustworthiness') is a valid measure of whether that topic is about AI governance.

Editorial extensions

If this is right

Institutional frameworks like the EU's guidelines overlap with user concerns on privacy, transparency, and accountability but miss operational domains like project management, strategy, and customer communication.
Governance-relevant topics form a value-chain-like mapping spanning deployment infrastructure, data handling, analytics, permissions, and downstream reporting.
Absence of abstract ethical ideals such as sustainability and fairness in user reviews should not be read as user indifference; users express governance in immediate, task-oriented language.
A dual master-text approach yields complementary views: principles recover technical backbone themes, while general terms recover relational and communicative themes.
A stability check with the top 50 topics shows the main clusters persist, adding granularity but not fundamentally new governance dimensions.

Reading between the lines

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

If this bottom-up method transfers, regulators could monitor platform reviews or support tickets as a real-time early-warning signal for emergent governance failures before formal audits catch them.
The same pipeline could be run on reviews of specific AI categories, such as hiring tools or generative chatbots, to reveal domain-specific governance concerns that could be tested against risk tiers in the EU AI Act.
The 2x2 matrix suggests user-centered governance metrics: internal-technical (access control), external-technical (data reporting), internal-nontechnical (workflow coordination), and external-nontechnical (customer trust).
Because the method depends on seed words, a different seed set (for example, 'safety', 'misuse', or 'liability') might surface different clusters, so the claimed bottom-up discovery is bounded by the seeds' semantic neighborhood.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Punchline: this is a credible, clearly written exploratory study with a real novel contribution—applying BERTopic to 108k G2 reviews to map user-expressed themes around AI governance. The pipeline is described in enough detail for a reader to reconstruct it, and the robustness check with a wider cutoff is a good-faith sanity check. The finding that users talk about operational and managerial themes—project management, roadmaps, internal communication—that are absent from the EU Ethics Guidelines is the kind of gap that deserves empirical attention.

The soft spots are real and concentrate on construct validity. The G-condition master text is just 'governance', 'ethics', 'trustworthiness'; the E-condition master text is the seven EU principles themselves. Cosine similarity between topic-keyword embeddings and these seeds is the entire selection mechanism. The stress-test note is fair: without a control condition, you can't tell whether topics like 'Project Management' get selected because users are voicing governance concerns or because 'governance' is semantically adjacent to 'management'. Appendix C makes that worry vivid—topics are full of product names (bitrix24, hive, notion, aha, jira) and generic task words. The paper's own Section 4.4 comparison to the EU principles is partly circular on the E-condition, since the same principles were used to build the master text. The fixed-cutoff robustness check adds granularity but doesn't test the relevance filter itself. On the positive side, the paper is explicit about these limits (Section 5.4) and doesn't overclaim normatively; it's descriptive.

So: the central claim is plausible but not yet demonstrated beyond reasonable doubt. The paper deserves a serious peer review, because the data and method are concrete and the question matters, but reviewers should request a control condition (e.g., a master text built from management terms like 'administration', 'organization', 'efficiency') and ideally release the anonymized topic lists or code. If the selection survives that test, it would be a solid empirical complement to the governance literature. If not, the 'bottom-up' framing would need to be dialed back substantially.

Referee Report

4 major / 6 minor

Summary. The paper proposes a bottom-up, empirical approach to AI governance by analyzing 108,998 user reviews of AI products from G2.com. It applies BERTopic to extract latent topics, then filters these topics by cosine similarity to two 'master texts': one based on the seven EU Ethics Guidelines principles and one based on the general terms 'Governance', 'Ethics', and 'Trustworthiness'. The selected topics are clustered with K-means and organized into a 2x2 matrix of technical/non-technical and internal/external dimensions. The authors report that user discourse overlaps with institutional frameworks on topics such as privacy, transparency, and security, but also surfaces areas largely absent from those frameworks, such as project management, strategy, and customer interaction, and they argue for more user-centered, operationally grounded AI governance.

Significance. If the methodology is valid, the paper makes a useful contribution: it provides a large-scale, reproducible empirical map of topics that may reflect user governance concerns, and it offers a transparent pipeline with an explicit robustness check. The appendices with full topic-keyword lists support transparency. However, the central claim—that the resulting map represents 'bottom-up' governance perspectives rather than a projection of institutional vocabulary—depends on the construct validity of the governance-relevance filter, which is currently not established. The significance of the contribution therefore hinges on the methodological concerns detailed below; if those are addressed, the paper could be a valuable complement to principle-based AI governance scholarship.

major comments (4)

[3.3 and 4.4] The E-condition master text is constructed from the seven principles of the EU Ethics Guidelines for Trustworthy AI, and Section 4.4 then reports 'a strong degree of overlap' between the topics selected using this master text and the EU principles themselves. This overlap is partly tautological: the filter was built from the same seven principles, so observing privacy, transparency, and accountability topics in the selected set is not independent evidence that user discourse aligns with the EU framework. Please either repeat the comparison using a principle set not used for selection (for example, the OECD AI Principles or the values in Jobin et al., 2019) or explicitly frame the overlap as a consequence of the filtering procedure rather than as a substantive discovery.
[3.3–3.5 and Appendix C] The governance-relevance filter has no control condition. The G-condition master text ('Governance', 'Ethics', 'Trustworthiness') is semantically broad, and the selected topics include many that appear to be generic software-review topics driven by product names or management/operations vocabulary: for example, E2.1 ('bitrix24', 'hive', 'task'), E1.1 ('sql', 'database'), G7.5 ('sales', 'marketing', 'leads'), and G1.1 ('glider', 'candidates', 'assessment'). Without a control condition—such as comparing against a seed list of generic management terms (e.g., 'management', 'administration', 'operations') or against random word lists—the analysis cannot distinguish topics that reflect user governance concerns from topics that are merely semantically adjacent to management discourse. This directly affects the central claim in the Abstract and in Section 4.1 that the map reflects 'how governance is perceived and experienced' by users. Please add a control analysis or a human-annotation study of a sample of selected topics to validate the construct.
[3.5] The drop-off-based cut-off includes an explicit free parameter: 'we required a minimum inclusion of 10 topics.' This choice, together with the use of the largest relative drop in cosine similarity, determines which topics enter Tables 1–2 and Figure 2. The robustness check in Section 4.5 (top-50 fixed cut-off) shows that broader inclusion enriches existing clusters rather than changing the structure, which is reassuring. However, the sensitivity of the main results to the minimum-inclusion threshold (e.g., 5, 10, 15) and to the number of top keywords used for topic embeddings (currently fixed at 10) is not reported. Please add this sensitivity analysis to demonstrate that the central map is not an artifact of these choices.
[5.4] The limitations section acknowledges that 'the construction of governance master texts involves subjective design decisions that can influence which topics are highlighted,' but it does not address the more fundamental issue that the E-condition is definitionally tied to the EU principles and that the G-condition may select for generic management vocabulary. Since the paper's central claim is that the analysis surfaces user-centered governance concerns beyond institutional frameworks, the limitations discussion should explicitly state that the filtering step reintroduces a top-down component and that the 'bottom-up' claim is therefore limited to the topic extraction stage, not the full pipeline. This would be an honest and important caveat.

minor comments (6)

[Table 1] The cluster label 'E6.1. GDRP & Compliance' contains a typo; it should read 'GDPR & Compliance'.
[Section 2.2] The reference to 'Mittelsatdt, 2019' is a misspelling; it should be 'Mittelstadt, 2019'.
[Appendix C] Many topic keywords are dominated by product names (e.g., 'bitrix24', 'hive', 'domo', 'aha', 'truefoundry'). Since the paper aims to identify governance-related themes rather than product-specific feedback, consider masking or aggregating product names in the topic representation, or at least discussing how product-specific topics were interpreted.
[References] The reference 'Kevans, J. (2024). 10 Best Sooftware Review Sites' contains a typo; it should be 'Software'.
[Appendix A] The titles of Figures AF1 and AF2 use 'Similiarity' instead of 'Similarity'.
[Figure 4] The mapping between bottom-up clusters and EU principles is interpretive and not data-driven; the caption should state this clearly so that readers do not infer a quantitative alignment.

Circularity Check

1 steps flagged · score 6.0 of 10

EU-principle seed list makes the reported EU overlap partly a selection artifact; G-condition and non-seed operational topics preserve some independent content.

self definitional [Section 3.3 (master-text construction) and Section 4.4 (comparison with EU principles)]
"we drew on the seven core principles outlined in the European Commission’s Ethics Guidelines for Trustworthy AI (2019)—a widely cited and influential framework in both policy and academic domains—and use the words for the seven core principles for our governance master text. [...] The visual mapping shows a strong degree of overlap in technical domains. EU principles such as Technical Robustness & Safety and Accountability correspond closely with user-identified clusters including Permission & Roles, Privacy & Regulation."

The E-condition topic set is selected as the topics with the highest cosine similarity to a master text composed of the EU principles' label words (accountability, transparency, fairness, human agency and oversight, privacy and data governance, technical robustness and safety, societal and environmental well-being). Reporting that the resulting clusters show strong overlap with those same EU principles is therefore a restatement of the selection criterion rather than an independent empirical finding. The alignment between E-clusters and Technical Robustness & Safety, Accountability, and Privacy is built into the ranking step.

full rationale

The main derivation chain is BERTopic topics -> cosine similarity to two master texts -> drop-off cutoff -> K-means clustering -> comparison with EU principles. The only load-bearing reduction to inputs is the E condition: because the EU master text is literally the seven EU principles, the E-derived clusters' overlap with those principles is entailed by the selection step. The G condition is a genuinely different seed set, and the recurrence of privacy, permissions/roles, HR, and communication themes under both conditions gives those themes independent support. The operational topics (project management, strategy, customer interaction) are not in the seed lists, so the central 'overlooked areas' contribution is not purely definitional. No load-bearing self-citation or imported uniqueness theorem is present: Pasch & Ha (2025) only supports the data source's relevance, and the empirical pipeline is self-contained against the corpus. The top-50 robustness check is a reasonable stability test rather than a circular maneuver. The paper's self-acknowledged limitation in Section 5.4, namely that master-text construction involves subjective design decisions, further signals that the EU-overlap finding should be read as partly method-driven. Overall, the circularity is partial and concentrated in the EU-overlap comparison, while the G-condition results and non-seed operational topics keep the central claim from collapsing entirely into the inputs.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper's central mapping rests on the choice of master texts, the minimum topic inclusion, the cluster count, and the unvalidated premise that embedding similarity equals governance relevance. These are reasonable exploratory choices, but they are not derived from data or theory.

free parameters (4)

minimum topic inclusion for governance-relevant subset = 10
Section 3.5 imposes a floor of 10 topics to avoid overly narrow cut-offs; the value is hand-selected, not estimated from data.
number of K-means clusters (k) = 7
Section 3.6 selects k=7 based on elbow and silhouette heuristics; these are suggestive visual guidelines, not formal model selection.
top keywords per topic for embedding representation = 10
Section 3.4 averages embeddings of the top 10 keywords per topic; this count is arbitrary and affects all subsequent similarity scores.
BERTopic hyperparameters = not reported
UMAP, HDBSCAN, and BERTopic settings such as min_topic_size and random_state are not given, so the exact topic solution is not reproducible.

assumptions (4)

domain assumption Semantic similarity to a governance seed list is a valid proxy for whether a topic is governance-relevant.
This is the core operationalization in Sections 3.4 and 3.5; it is not validated against human annotation or external criteria.
domain assumption The seven EU Trustworthy AI principles adequately summarize the space of established AI governance and ethics frameworks.
Section 3.3 relies on the EU Ethics Guidelines as the 'established principles' master text, citing Papagiannidis et al. (2023) for coverage; this is a contestable but defensible choice.
domain assumption G2.com reviews by business and IT professionals are a suitable window into how AI governance is enacted in organizations.
Section 3.1 justifies the data source, and the limitations section acknowledges exclusion of public sector and civil society voices.
standard math BERTopic's unsupervised topics are coherent semantic units that can be compared via averaged keyword embeddings.
The paper relies on the algorithm's outputs without reporting topic coherence or stability checks, a common but unverified assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Bottom-Up Perspectives on AI Governance: Insights from User Reviews of AI Products." pith.science (2026). https://pith.science/paper/BNSINYKI

@misc{pith2026250600080,
  author       = {Pith},
  title        = {Pith review of: Bottom-Up Perspectives on AI Governance: Insights from User Reviews of AI Products},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BNSINYKI}},
  note         = {Machine review of arXiv:2506.00080}
}

read the original abstract

With the growing importance of AI governance, numerous high-level frameworks and principles have been articulated by policymakers, institutions, and expert communities to guide the development and application of AI. While such frameworks offer valuable normative orientation, they may not fully capture the practical concerns of those who interact with AI systems in organizational and operational contexts. To address this gap, this study adopts a bottom-up approach to explore how governance-relevant themes are expressed in user discourse. Drawing on over 100,000 user reviews of AI products from G2.com, we apply BERTopic to extract latent themes and identify those most semantically related to AI governance. The analysis reveals a diverse set of governance-relevant topics spanning both technical and non-technical domains. These include concerns across organizational processes-such as planning, coordination, and communication-as well as stages of the AI value chain, including deployment infrastructure, data handling, and analytics. The findings show considerable overlap with institutional AI governance and ethics frameworks on issues like privacy and transparency, but also surface overlooked areas such as project management, strategy development, and customer interaction. This highlights the need for more empirically grounded, user-centered approaches to AI governance-approaches that complement normative models by capturing how governance unfolds in applied settings. By foregrounding how governance is enacted in practice, this study contributes to more inclusive and operationally grounded approaches to AI governance and digital policy.

Figures

Figures reproduced from arXiv: 2506.00080 by the authors.

**Figure 1.** Overview of Methodological Pipeline Topics Extracted from Online Reviews Extracting Topics with BERTopic Established Governance Principles Master Texts for AI Governance General Governance Terms Topics in Online Reviews of AI Products Ranking Similarity of Topics to AI Governance Measure Textual Similarity Select AI Governance Topics based on Drop-off Online Reviews on AI Products Clusters of AI Governance Topics K-… view at source ↗

**Figure 2.** Summary of Bottom-Up AI Governance Themes Non-Technical Technical External (Downstream) Internal (Upstream) Data & Analytics Reporting Analytics Database Data Privacy & Security Security Privacy Networks Dependencies Permission & Roles Roles Permission Admin Access Deployment & Infrastructure Deployment Orchestration Cloud Infrastructure Human Factors Employees HR Human – AI Recruitment Project Management Strategy P… view at source ↗

**Figure 4.** Comparing Bottom-Up AI Governance Themes and EU Ethics Guidelines for AI Non-Technical Technical External (Downstream) Internal (Upstream) Data & Analytics Reporting Analytics Database Data Privacy & Security Security Privacy Networks Dependencies Permission & Roles Roles Permission Admin Access Deployment & Infrastructure Deployment Orchestration Cloud Infrastructure Human Factors Employees HR Human – AI Recruitmen… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 11 canonical work pages

[1]

on the ground

er increasing scrutiny for their limited applicability in real-world contexts. These documents are typically developed by institutional experts, policymakers, and regulators — rather than by those who use, manage, or are directly affected by AI systems in applied settings. As a result, they often reflect high-level ideals rather than the day-to-day govern...

work page 2024
[2]

The technical/internal quadrant captures foundational system-level operations that support the functioning of AI systems but remain largely invisible to end users

provides a conceptual overview of how users perceive and experience AI governance—not as a fixed institutional checklist, but as a distributed set of concerns shaped by infrastructure, human operations, external communication, and regulatory exposure. The technical/internal quadrant captures foundational system-level operations that support the functionin...

work page 2021
[3]

Artificial Intelligence

Methodology To structure our analysis, Figure 1 provides a high-level overview of our methodological approach. We began by extracting topics from a large corpus of online reviews of AI products using BERTopic. These topics were then ranked according to their semantic similarity to two sets of governance-oriented reference texts: one based on established A...

work page 2024
[4]

governance master text

Overview of Methodological Pipeline Topics ExtractedfromOnline Reviews ExtractingTopics withBERTopic EstablishedGovernancePrinciples Master Texts forAI Governance General GovernanceTerms Topics in Online Reviews ofAI Products Ranking SimilarityofTopics toAI Governance MeasureTextualSimilarity Select AI GovernanceTopics basedon Drop-off Online Reviews on A...

work page 2019
[5]

external

Results Table 1: Topics and Cluster for Established Governance Principles Master Text Cluster Topics E1: Data & Analytics E1.1. Database E1.2. Analytics E1.3. Dashboards & Reporting E1.4. Metrics E1.5. Data Visualization and Business Intelligence E1.6. Data E2: Project Management E2.1. Project Management E2.2. Project Tasks E2.3. Productivity & Organizati...

work page 2024
[7]

These themes reflect how users experience AI governance through support systems, social media, messaging, and public interfaces

Bottom-Up AI Governance Themes by Master Text By contrast, the general governance terms approach (G) more frequently identifies clusters situated in the external, interactional domain, including Customer & Communication, Content & Information Quality, and Human Factors like Human-AI Interaction. These themes reflect how users experience AI governance thro...

work page 2019
[8]

on the ground

Comparing Bottom-Up AI Governance Themes and EU Ethics Guidelines for AI TechnicalNon-Technical External(Downstream) Internal(Upstream) Data & AnalyticsAnalyticsReportingDatabaseData Privacy & SecurityPrivacySecurityNetworksDependencies Permission & RolesPermissionRolesAdminAccess Deployment& InfrastructureOrchestrationDeploymentCloudInfrastructure Human ...

work page 2021
[9]

Latent dirichlet allocation

Discussion 5.1 A Bottom-Up View of AI Governance This study set out to explore how AI governance is discussed and experienced by users in applied, organizational settings — departing from the dominant top-down emphasis in the literature. Rather than starting from established principles or expert taxonomies, we examined user-generated reviews of AI product...

arXiv 2021

Show all 12 references

[18]

Pasch, S., & Ha, S. Y. (2025). Human-AI Interaction and User Satisfaction: Empirical Evidence from Online Reviews of AI Products. arXiv preprint arXiv:2503.17955. Papagiannidis, E., Mikalef, P., & Conboy, K. (2025). Responsible artificial intelligence governance: A review and ...

2025 arXiv
[19]

These inflection points were used as cut-offs to define the final topic sets. In both cases, we conducted robustness checks using a broader inclusion of the top 50 topics (see Section 4.5), which confirmed that additional topics introduced thematic redundancy or weakened conce...

1987
[109]

Jasanoff, S. (2004). The idiom of co-production. In States of knowledge (pp. 1-12). Routledge. Jobin, A., Ienca, M., & Vayena, E. (2019). The global landscape of AI ethics guidelines. Nature machine intelligence, 1(9), 389-399. Mäntymäki, M., Minkkinen, M., Birkstedt, T., & Vi...

2004
[7515]

ethics-washing

Schiff, D., Biddle, J., Borenstein, J., & Laas, K. (2020, February). What's next for ai ethics, policy, and governance? a global overview. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (pp. 153-158). Sigfrids, A., Leikas, J., Salo-Pöntinen, H., & Koskimi...

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

on the ground

er increasing scrutiny for their limited applicability in real-world contexts. These documents are typically developed by institutional experts, policymakers, and regulators — rather than by those who use, manage, or are directly affected by AI systems in applied settings. As a result, they often reflect high-level ideals rather than the day-to-day govern...

work page 2024

[2] [2]

The technical/internal quadrant captures foundational system-level operations that support the functioning of AI systems but remain largely invisible to end users

provides a conceptual overview of how users perceive and experience AI governance—not as a fixed institutional checklist, but as a distributed set of concerns shaped by infrastructure, human operations, external communication, and regulatory exposure. The technical/internal quadrant captures foundational system-level operations that support the functionin...

work page 2021

[3] [3]

Artificial Intelligence

Methodology To structure our analysis, Figure 1 provides a high-level overview of our methodological approach. We began by extracting topics from a large corpus of online reviews of AI products using BERTopic. These topics were then ranked according to their semantic similarity to two sets of governance-oriented reference texts: one based on established A...

work page 2024

[4] [4]

governance master text

Overview of Methodological Pipeline Topics ExtractedfromOnline Reviews ExtractingTopics withBERTopic EstablishedGovernancePrinciples Master Texts forAI Governance General GovernanceTerms Topics in Online Reviews ofAI Products Ranking SimilarityofTopics toAI Governance MeasureTextualSimilarity Select AI GovernanceTopics basedon Drop-off Online Reviews on A...

work page 2019

[5] [5]

external

Results Table 1: Topics and Cluster for Established Governance Principles Master Text Cluster Topics E1: Data & Analytics E1.1. Database E1.2. Analytics E1.3. Dashboards & Reporting E1.4. Metrics E1.5. Data Visualization and Business Intelligence E1.6. Data E2: Project Management E2.1. Project Management E2.2. Project Tasks E2.3. Productivity & Organizati...

work page 2024

[6] [7]

These themes reflect how users experience AI governance through support systems, social media, messaging, and public interfaces

Bottom-Up AI Governance Themes by Master Text By contrast, the general governance terms approach (G) more frequently identifies clusters situated in the external, interactional domain, including Customer & Communication, Content & Information Quality, and Human Factors like Human-AI Interaction. These themes reflect how users experience AI governance thro...

work page 2019

[7] [8]

on the ground

Comparing Bottom-Up AI Governance Themes and EU Ethics Guidelines for AI TechnicalNon-Technical External(Downstream) Internal(Upstream) Data & AnalyticsAnalyticsReportingDatabaseData Privacy & SecurityPrivacySecurityNetworksDependencies Permission & RolesPermissionRolesAdminAccess Deployment& InfrastructureOrchestrationDeploymentCloudInfrastructure Human ...

work page 2021

[8] [9]

Latent dirichlet allocation

Discussion 5.1 A Bottom-Up View of AI Governance This study set out to explore how AI governance is discussed and experienced by users in applied, organizational settings — departing from the dominant top-down emphasis in the literature. Rather than starting from established principles or expert taxonomies, we examined user-generated reviews of AI product...

arXiv 2021