Insights

RAG and Law: Why Reliable AI Cites Its Sources

3 min read · Editorial oversight: Avv. Federico Papa

AI adoption in the legal sector has reached a turning point: according to the IX 2024 CENSIS and Cassa Forense Report on the Legal Profession, 55.3% of Italian lawyers express a positive inclination or openness toward its use. However, simple text generation does not guarantee the accuracy required for a legal document or professional opinion.

In brief

The IX 2024 CENSIS and Cassa Forense Report indicates that 55.3% of lawyers are open to AI. RAG (Retrieval-Augmented Generation) technology mitigates hallucination risks by consulting specific legal documents before responding. Negligent use of non-existent citations creates liability for bad-faith litigation under Art. 96 c.p.c. and breaches the duty of diligence. EU Regulation 2024/1689 (AI Act) mandates transparency obligations under Art. 50 effective August 2026. Current legislation prohibits the complete delegation of document drafting, requiring human oversight as reaffirmed by the National Bar Council and Art. 27 of the Code of Conduct for Lawyers.

1.
Beyond AI Memory: RAG Technology
Generalist models like ChatGPT or Claude rely solely on their training data, which can result in hallucinations. RAG (Retrieval-Augmented Generation) technology enables the system to consult specific documents before responding. This approach extracts relevant passages from trusted sources to formulate answers based on actual data rather than mere statistical probabilities.
2.
The edit.legal Standard: Post-Generation Verification
Mere document retrieval is insufficient to guarantee professional safety. The edit.legal standard includes a citation verification phase that cross-references the generated output against over one million official documents, including legislation and Supreme Court precedents. This verification ensures that every legal reference cited is accurate and exists within official databases.
3.
The Legal Risk of Hallucinations
Relying on unverified systems can lead to severe financial and ethical consequences. International cases of sanctions for frivolous litigation are already well known, while in Italy the negligent use of non-existent AI-generated citations exposes the practitioner to liability for bad-faith litigation under Art. 96 c.p.c. and breach of the duty of diligence.
Apply this research directly with edit.legal
Legal research and drafting with citations checked against official databases. Try edit.legal for free, no credit card.
Try edit.legal
4.
AI Act and New Transparency Obligations
While rules on GPAI models and governance will apply from August 2, 2025, transparency obligations under Art. 50 of EU Regulation 2024/1689 (AI Act) will take effect on August 2, 2026. Those using AI systems must clearly label synthetic content and inform the user. AI systems intended for use by judicial authorities are classified as high-risk.
5.
The Human Role in Legal Assistance
Current legislation establishes that AI can only be used for support functions, prohibiting the complete delegation of legal document drafting to automated systems. Human oversight must remain central, a principle reaffirmed by the National Bar Council regarding professional liability and the necessity of the lawyer's supervision.
6.
Operational Tools for the Modern Lawyer
Qualitative analysis of thousands of real queries demonstrates that efficiency comes from tool integration. edit.legal offers a native editor with variable fields and automated legal basis (Base Giuridica), along with a Microsoft Word add-in and an MCP server for Claude or Cursor. These tools allow working across 21 practice areas with multiple agents deliberating on the specific case.
7.
Data Security and European Infrastructure
Protecting sensitive client data is a priority guaranteed by an entirely European infrastructure compliant with the GDPR. Unlike many foreign services, edit.legal offers transparent pricing and a free trial requiring no credit card. This allows professionals to test the quality of source verification without mandatory sales demos or hidden costs.

Editorial oversight: Avv. Federico Papa·ICAM

Frequently asked questions

What is the difference between ChatGPT and a legal RAG system?

ChatGPT uses its internal memory, while a legal RAG system reads specific legal documents before responding. edit.legal goes further, verifying every citation against official databases such as the Official Gazette to eliminate the risk of invented citations.

Can AI be sanctioned if it makes a citation error?

The sanction applies to the lawyer and not to the software, for breach of the duty of diligence and technological competence. Case law has already begun to apply sanctions for bad-faith litigation regarding pleadings based on non-existent precedents generated by AI.

Do I need to inform my client if I use AI?

Yes, based on an extensive interpretation of the duty of information under Art. 27 of the Code of Conduct for Lawyers. While the rule does not explicitly mention AI, the duty to inform the client about the execution of the engagement requires transparency regarding the use of automated tools under human supervision.