Text Simplification for Citizens with GenAI

This information was sourced from "Text Simplification for Citizens with GenAI" by Observatory of Public Sector Innovation (OECD), https://oecd-opsi.org/innovations/text-simplification-for-citizens-with-genai/. Licensed under Creative Commons Attribution-ShareAlike 3.0 IGO (CC BY-SA 3.0 IGO).

Innovation Overview

Clear and accessible communication is essential for ensuring that citizens can effectively navigate public services. This work examines the use of generative AI models to simplify administrative texts on the INPS (Italy’s National Institute of Social Security) web portal, which provides nearly 500 digital services related to pensions and welfare. As Europe’s largest welfare institution, INPS serves approximately 45 million citizens, making the clarity of its service descriptions crucial for accessibility and user experience.

INPS has been running a structured simplification program aimed at making its content understandable to users with a middle school education level. To date, over 700 texts have been manually simplified, using the Gulpease index to measure readability. The average score improved from 57 to 64, reflecting a substantial increase in accessibility.

Currently, INPS relies on a manual process to simplify the language of its Service Pages—a time-consuming and resource-intensive approach that requires specialized expertise. The need for clearer bureaucratic language has been recognized in Italy for decades. A 2002 directive, still in force, states that “all texts produced by administrations must be designed and written to be understood by those who receive them.” Despite this, institutional communication often remains complex, limiting citizens' ability to fully understand and access essential services. While simplification enhances transparency and public trust, a major challenge is maintaining the legal precision required for official documentation.

This work explores how Large Language Models (LLMs) can support INPS in transforming its manual simplification process into a semi-automated system that is both efficient and scalable. The goal is to develop a solution that can process large volumes of text while ensuring readability, accuracy, and compliance with legal standards. To achieve this, the research establishes an evaluation framework based on text metrics and conducts user surveys to assess the accessibility and the quality of AI-generated text compared to original and manually simplified versions.

The findings of this study highlight the effectiveness of AI-driven text simplification, demonstrating that AI-generated versions not only match but, in some aspects, surpass human expert work. AI-simplified texts showed significant improvements in readability, fluency, and user engagement, often leading to a preference rate comparable to or even higher than that of human-simplified versions. Most notably, the results of the metric-based evaluation were validated through large-scale user testing with over 1,200 participants—an innovative step rarely seen in similar research.
Looking ahead, the potential for refining and expanding this approach is substantial; among the possibilities, the generation of pre-simplified texts directly from legal norms, the introduction of a compliance-evaluation framework, and the development of collaborative workflows, where AI and human experts iteratively refine simplifications, combining automation with expert oversight.

What Makes Your Project Innovative?

Most of the literature on AI-assisted text simplification focuses on the English language, making the simplification of Italian texts a relatively unexplored field, despite its importance in the context of public administration. This study aims to provide an innovative contribution to this area, focusing on the simplification of administrative texts in Italian.

Existing metrics often fail to capture all aspects of text simplification or correlate with human judgment. This study incorporates user evaluation to assess whether analytical metrics actually correlate with improved user perception and introduces GULBERT, a novel simplification metric that balances readability and semantic preservation, specifically tailored for Italian administrative texts.

By incorporating real user assessments on factors such as clarity, engagement, and confidence in the content, the study ensures that AI-generated simplifications align with user needs.

What is the current status of your innovation?

The experimentation on text simplification has been developed as a prototype and has successfully passed the evaluation stage; it is now transitioning into an operational business process by INPS and its technology partners, integrating a graphical interface to support the human-in-the-loop approach, in the form of a collaborative workflow.

Alternative promising developments are in the evaluation stage, such as the generation of Service Pages from scratch, using exclusively technical documents and regulations via RAG (Retrieval Augmented generation), while ensuring compliance with completeness and accessibility criteria, or the introduction of a source-referencing system. Moreover, the currents results are being prepared for publication as a comparative analysis.

Collaborations & Partnerships

INPS (Institutional Partner):

• Defined the strategic goals, use case and evaluation criteria
• Coordinated all phases of the experimentation, enabling access to real-world administrative documents
• Offered domain expertise

Almawave:

• Study of the literature, conceptualization of the experimentation
• Development and industrialization of the prototype

Almaviva:

• Process management support

Unguess:

• Developed and administered the user validation phase

Users, Stakeholders & Beneficiaries

• INPS, the main stakeholder, optimized the document simplification process, improving its communication effectiveness, fostering inclusiveness, transparency and trust
• Citizens, the primary beneficiaries of a more inclusive communication ecosystem, which democratizes access to complex administrative information, especially for individuals with reading difficulties, limited language skills, or lower educational backgrounds.

Results, Outcomes & Impacts

The AI effectively simplifies text, delivering expert-level results in a fraction of the time. Its robust metrics framework accurately evaluates various aspects of simplification and highlights areas for improvement.

A large-scale survey, including A/B testing with 1620 participants, confirmed the AI’s effectiveness. The AI-generated text was preferred in more than 70% of responses vs. the original text, and outperformed the original across all metrics, scoring significantly higher in fluency (+14.1%), readability (+13.5%), clarity (+11.11%), and engagement (+12.3%).

These findings highlight key benefits: scalable AI-driven simplification for public administration, improved accessibility to administrative information, and seamless integration into broader communication strategies.

Challenges and Failures

AI-driven text simplification for administrative documents must navigate challenges like maintaining linguistic consistency, preserving key information, and balancing clarity with depth. Traditional evaluation metrics, such as BLEU and ROUGE, fall short in capturing these complexities, highlighting the need for more advanced assessment frameworks.

To address these gaps, strategies like the novel GULBERT metric, large-scale user testing, and iterative model enhancements have been successfully implemented.

Looking ahead, efforts will focus on strengthening AI-human collaboration, expanding datasets, and refining simplification techniques to ensure continuous improvement and more effective, accessible communication.

Conditions for Success

Success requires a public administration that truly puts users at the center. AI adoption must be guided by a governance framework that prioritizes usefulness and user impact over hype and novelty. Clear policies ensure legal compliance, while strong leadership aligns innovation with public value. Robust infrastructure and adequate funding enable scalability, but it is skilled staff and a shared commitment to accessibility that ensure quality. Only by focusing on citizen needs can AI become a sustainable tool for better services.

Replication

This AI-driven text simplification model demonstrates significant potential for broader application, particularly in sectors such as healthcare and legal services, where complex documents require greater clarity.

Its success in streamlining bureaucratic Italian texts highlights its adaptability. Within public administration, the model could scale across departments—such as social services and taxation—or be implemented in other governmental systems, accommodating diverse languages and regulatory frameworks.

Ensuring a successful expansion requires maintaining a careful balance between automation and human oversight, preserving both accuracy and public trust in official communications.

Lessons Learned

Key insights from this AI-driven text simplification initiative:

• Human-AI collaboration is critical—automation excels at scaling simplification, but human oversight ensures legal accuracy and nuance.
• User validation is not negotiable: metrics alone (e.g., readability scores) don’t capture users’ preference; direct feedback is needed to effectively assess the efficacy of the model.
• Context matters—effectiveness hinges on tailoring prompts/models to institutional language (e.g., Italian bureaucracy).

Future adopters should prioritize hybrid workflows, localized testing, and clear success metrics.

Anything Else?

This case was inspired by a need-driven innovation approach, a strategic pillar in INPS. It started from a well-identified user need—making administrative texts easier to understand—and an organizational pain point: a manual simplification process that was time-consuming and resource-intensive. Generative AI was tested as a targeted solution to address both. The project defined measurable goals, applied a rigorous validation process, and demonstrated concrete value. It shows how innovation can be effective when guided by real needs, not by the technology itself.

Budget: Not disclosed