That assurance addresses an important concern about sensitive court information. It does not, however, settle every question about how AI-assisted work is recorded, reviewed or disclosed to the defence. Keeping data out of model training and keeping judicial proceedings accountable are separate challenges.
What the ministry’s response says
Il Sole 24 Ore’s Giovanni Negri reports that the ministry provided a written answer to a parliamentary question from Enrico Costa about the Copilot experiment.
Costa sought details about system instructions, recommended prompts and operational guidance available to magistrates working with judicial documents. He also asked who developed those materials and how they were validated.
According to the newspaper, the ministry’s answer was specifically that no instructions or standardized prompts had been developed or validated to guide:
- Assessment of the merits of a dispute.
- Reconstruction of the facts.
- Evaluation of evidence.
- Adoption of the final decision.
The distinction matters. This is not a claim that participants receive no guidance whatsoever. It is a narrower statement that the ministry’s trial does not provide instructions intended to shape judicial decisions.
The newspaper reports that the ministry characterizes its supporting initiatives as organizational, administrative, documentary and individual assistance—not a mechanism for delegating judicial authority to software.
An earlier official notice adds important context
The Ministry of Justice’s own trial notice confirms that Copilot licenses became available to career magistrates from January 1, 2026, integrated into their existing Office 365 package. It describes participation as voluntary and identifies archive searching as an example of a supporting activity. Magistrates could opt out, after which the ministry would deactivate their Copilot license and restore the previous configuration.
The notice also announced a January 26 webinar, including practical examples and prompts developed through smaller experiments during 2025. That is compatible with the reported parliamentary response: prompts for supporting work are not necessarily prompts for evaluating evidence or deciding a case.
One further detail deserves attention. Although the notice describes use within the institutional Microsoft 365 environment, it also allows external consultation at the user’s informed choice. That does not establish that confidential documents were sent externally, but it means the trial should not be described as categorically disconnected from outside information sources.
“Not used for training” does not mean “not processed”
Microsoft’s commercial Copilot documentation provides corroborating product context: prompts, responses and information accessed through Microsoft Graph are not used to train foundation language models. The service nevertheless uses organizational content to generate answers. Using a document to answer a request is different from using it to train the underlying model.
Microsoft also says it stores interaction data, including prompts and responses, and provides administrative tools for searching that information and applying retention policies through Microsoft Purview. These are general product capabilities, not proof of the Italian ministry’s particular configuration.
For enterprise IT administrators, the practical lesson is to ask separate questions rather than treating “no training” as an all-purpose security certificate:
- Which documents can each user access?
- What interaction records are retained?
- Who can retrieve those records?
- Which external services or agents are permitted?
- Which activities remain outside the approved AI workflow?
Microsoft says Copilot respects existing user access permissions. Reviewing those permissions is therefore an important deployment task: an AI assistant is not a substitute for a sound access-control model.
The defence-transparency question remains open
Costa’s parliamentary question also addressed procedural transparency. According to Il Sole 24 Ore, he asked what safeguards would ensure that defence counsel knew when a judge used AI to analyze case files and could examine the instructions, prompts and generated results.
The ministry’s reported reply emphasized a different institutional boundary: it does not—and cannot—control how individual judges conduct their preparatory research, analysis and deliberation.
That explains the ministry’s position on judicial independence. The reply, as reported, does not establish a specific notification or disclosure procedure. Nor does it demonstrate that such safeguards are absent. The available account simply leaves their operation unresolved.
Il Sole 24 Ore places these concerns within a broader digital-sovereignty debate, citing an open letter signed by more than 50 senior judicial officials, including National Anti-Mafia Prosecutor Giovanni Melillo. The ministry’s reported response stresses that justice data remain within its institutional archive and are unavailable for model training.
The central takeaway is neither that Copilot has become a judge nor that a no-training assurance resolves the controversy. Italy’s ministry describes a support-only experiment with protected institutional data. The remaining governance question is how that boundary is made demonstrable—without confusing technical security, judicial independence and the defence’s ability to scrutinize proceedings.