evidoria

← Back to browse

Good practice Imported

INPS's AI-Rewritten Plain-Language Content — Italy's Social Security Institute Matches Expert Writers in a 1,600-User Blind Test

Italy · Rome · See the Italy profile · See the Rome profile

Evidence: Quasi-experimental Top 25% 73/100 · Ask Evidence Copilot about this practice

Italy's INPS used AI to rewrite 2,100+ pieces of citizen-facing content in plain language; in a blind test with 1,600 users, AI-simplified text matched or beat expert versions ~70% of the time, while 59 AI systems now handle most service requests.

~70%
User preference for AI-simplified text vs original (blind test, 2024–2026 AI plan)
just over 70%
User preference for expert-simplified text vs original (same test)
2,100+
Pages/text items rewritten by AI (2024–present)
59
AI systems in operation at INPS (2026)
~80%
Incoming service requests handled by AI systems (2026)
87%
Certified email (PEC) traffic handled by AI (2026)
INPS's AI-Rewritten Plain-Language Content — Italy's Social Security Institute Matches Expert Writers in a 1,600-User Blind Test

Details

Maturity
Established
Promoter
Istituto Nazionale della Previdenza Sociale (INPS)
Period
2024–present
Keywords
social security, digital government, citizen communication, accessibility

Context

INPS, Italy's national social security institute, serves roughly 45 million residents, receives about 800 million site visits a year and offers over 450 services across more than 2,500 forms and applications.

Objectives

In April 2024 INPS issued its first official directive on AI use, followed by a 2024–2026 AI development plan. As part of that plan, INPS used generative AI to rewrite institutional web content — replacing dense administrative Italian with clearer, more direct language.

Activities

The process pairs AI-generated drafts with human review for semantic accuracy before publication, applied across more than 2,100 pages and text items to date. To test quality, INPS and researchers ran a blind comparison with roughly 1,600 members of the public: participants were shown original text alongside both an expert-simplified version and an AI-simplified version, without being told which was which.

Results

Expert versions were preferred in just over 70% of cases, and the AI versions achieved essentially the same ~70% preference rate, with some users describing the AI output as clearer and more scannable than the human original. By INPS's 2026 annual report, 59 distinct AI systems were in operation across the institute, together handling about 80% of incoming service requests and 87% of certified email (PEC) traffic.

Conclusions

Independent Italian press and a workers' union bulletin corroborate the rollout, though none names the specific underlying AI model or publishes an error-rate/hallucination audit for the rewriting tool.

Implementation

Indicative cost
High (€500k–€5M) — No specific budget disclosed; investment spans a formal 2024–2026 AI development plan across 59 operational AI systems.
Time to results
Long (> 3 years) — First AI directive issued April 2024; 2024–2026 AI development plan; content rewriting and system rollout ongoing through 2026.
Staffing & skills
INPS internal AI/content teams, human reviewers pairing with AI-generated drafts for semantic accuracy

Conditions for success

  • human review paired with AI-generated drafts before publication
  • a formal AI directive and multi-year AI development plan
  • blind user testing to validate quality against an expert baseline

Where it fits

Governance type
national social security institute
Scale
national, 45 million residents, 800 million visits/year
Income level
high-income (Italy, EU)

Commonly funded by

Digital Europe Programme National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful