Humanities & Social Sciences Communications · 2026

Generative AI Guidelines in Industry

A text-mining analysis of corporate guidelines and policy statements across 14 industrial sectors (IGGA)

Junfeng Jiao, Saleh Afroogh, Kevin Chen, David Atkinson, Amit Dhurandhar

Paper Code Data

160corporate guidelines and policy statements

14industrial sectors

k = 8governance themes, chosen by three diagnostics

≥ 75%top terms stable under bootstrap (ARI ≳ 0.7)

Overview

Companies are writing their own rules for generative AI faster than regulation can keep up, yet little is known about how corporate governance compares across sectors. We built IGGA (Industrial Guidelines and Policy Statements for Generative AIs), a dataset of 160 official guidelines and policy statements from top-tier companies in 14 industrial sectors, and analyzed it with TF-IDF, K-Means clustering, keyword frequency maps and Sankey co-occurrence diagrams. The analysis surfaces eight governance themes and shows clear divergence by sector: transparency in technology, consent in healthcare and IP protection in publishing. We translate the findings into a modular governance architecture: universal baseline duties plus sector-specific risk modules.

Highlights

Eight themes emerge: data governance and privacy; safety and human oversight; security and abuse prevention; IP and content integrity; transparency and explainability; risk and compliance; workforce and change management; innovation and sandboxes.
Privacy dominates the discourse (227 mentions in legal tech alone), while disclosure appears only 14 times and human-centric just twice across all documents: clear blind spots for future guidelines.
A convergent baseline is forming across frameworks from OECD to ISO and NIST: privacy-by-design, explainability and continuous risk management.
We propose universal duties (audit trails, bias logs, disclosure triggers) plus sector “risk modules”, so lower-resource sectors adopt only what they need.

Results

Fig. 1 of the paper: frequency of the ten key concepts in each of the 14 sectors.

Fig. 2 of the paper: industry-wise Sankey diagrams of keyword co-occurrence in AI integration guidelines.

All sectors

Collaborative governance

Modular “living” guidelines co-created by developers, regulators, users and ethicists, refined through AI-assisted audits and sandboxes

Adaptive oversight

Healthcare & pharma

Low-risk tools iterate quickly; high-risk diagnostics stay under real-time safety monitoring, with consent traceability

Augmentation-first

Creative & language

Human review of AI-generated content and full disclosure of AI use

Tiered explainability

Social media

High-level summaries for users, detailed logic for regulators, within GDPR/CCPA limits

Human–AI collaboration

Design & entertainment

Label AI-generated output and secure creator IP

Proactive governance

Legal tech

Pre-deployment bias simulations before tools reach clients

Consumer control

Advertising & marketing

User-controlled data-use boundaries that balance personalization and exploitation

Model-risk control

Finance

Model-risk approvals and sandbox stress-tests for fraud detection

Sector-specific governance takeaways from the paper’s synthesis (Table 2).

TF-IDF cosine similarity between industry guideline texts (from the IGGA notebook).

How it works

Selected 160 top-tier companies with official GAI/LLM guidelines across 14 sectors and diverse regions.
Cleaned, tokenized and lemmatized texts with synonym merging; TF-IDF vectorization.
K-Means for k = 2–12: elbow, silhouette and Calinski–Harabasz all favor k = 8; bootstrap stability on 80% resamples, 100 runs.
Key-concept frequency maps, TF-IDF heatmaps and Sankey co-occurrence diagrams per sector.

Resources

Paper (HSSC)Paper ↗Preprint (arXiv)arXiv ↗IGGA datasetDataverse ↗IGGA notebookGitHub ↗

Citation

Jiao, J., Afroogh, S., Chen, K., Atkinson, D., & Dhurandhar, A. (2026). Generative AI and LLMs in industry: a text-mining analysis and critical evaluation of guidelines and policy statements across 14 industrial sectors. Humanities and Social Sciences Communications, 13, 410.