Copyright provisions in the new Italian AI-law: reinforcing human authorship and text and data mining (TDM) exceptions

On 17 September, the Italian Senate approved Italy’s new law on AI, after a legislative process of nearly two years. The law was published in the Official Gazette as Law No. 132 of 23 September 2025 and will enter into force on 10 October 2025 (full text in Italian here).

The law includes a provision on copyright (Article 25), which introduces changes to the Italian Copyright Law (ICL) and is divided into two parts:

1. AI-assisted works – Article 25 amends Article 1 ICL (which sets out the general originality clause) to clarify that copyright applies to works created with the aid of AI tools, provided there is a substantial human intellectual contribution.
2. Text and Data Mining (TDM) – Article 25 also inserts a new Article 70-septies ICL, stating that reproductions and extractions from works or other materials lawfully accessible are permitted when carried out through AI systems, including generative AI, subject to the applicable TDM exceptions. While attempting to clarify the legal landscape for AI-generated content and its training, the new law (Article 26(3)) also tightens the framework with new criminal sanctions for TDM violations.

This article delves into these pivotal provisions, exploring their implications for human authorship, the evolving scope of TDM, and the unresolved questions that remain.

Focus on authorship

Article 25 amends Article 1 ICL, which now reads (our translation):

“Works of human intellect of a creative nature belonging to literature, music, figurative arts, architecture, theatre, and cinematography, whatever their mode or form of expression, are protected under this law, even if created with the aid of artificial intelligence tools, provided they constitute the result of the author's intellectual work” (additions in bold)

In other words, it is specified that (i) "works of authorship" must be of "human" origin; (ii) works created with the aid of artificial intelligence tools are protected provided that their creation derives from the intellectual work of the author.

This clarification is largely tautological, as the originality requirement under EU and Italian case law already demands that a work reflect the (human) author's own intellectual creation and his/her personal touch. Nevertheless, the explicit reference to AI as a tool serves as a policy signal, underlining that AI cannot replace human authorship, but AI-assisted works may be protected where a significant human contribution occurs. As recent practice by the US Copyright Office shows, assessing whether there is sufficient human contribution in AI-assisted works will likely prove complex and contested. The new provision doesn’t add much more in that sense.

Focus on TDM

Article 25 introduces a new Article 70-septies ICL, which provides (our translation and emphasis):

"Without prejudice to the provisions of the Berne Convention for the Protection of Literary and Artistic Works, ratified and made enforceable pursuant to Law no. 399 of 20 June 1978, reproductions and extractions from works or other materials contained online or in databases to which legitimate access is granted, for the purpose of text and data mining through artificial intelligence models and systems, including generative AI, are permitted in accordance with the provisions of articles 70-ter and 70-quater.”

This new article complements the two provisions already in the ICL, which almost literally implement the EU Copyright DSM Directive 790/2019:

• Article 70-ter (implementing Article 3 of the Directive) – research TDM permitted where there is lawful access;
• Article 70-quater (implementing Article 4 of the Directive) – commercial TDM permitted unless rightholders have reserved rights through an opt-out.

The novelty is that these rules are expressly extended to reproductions and extractions carried out through AI systems, including generative AI. In this sense, Article 25 aligns with the approach taken by the EU AI Act, which identified TDM exceptions as the key copyright provision for regulation of AI training (see our previous piece on this topic here), reflecting the same spirit at private law level. The real intention of Article 70-septies appears indeed to impose the limitations set out in the TDM exceptions as requirements (e.g., respect of opt-out) for AI training.

However, the scope remains unclear: does the notion of “extraction” in Article 70-septies cover only analytical uses (narrow TDM) or also the full training of AI models? The perimeter of the exception is uncertain. Some may claim that ingestion of data into models for training purposes is covered by TDM (particularly as extended by Article 70-septies), or that without memorization of the training materials there is not even a copyright relevant act, while rightholders may argue that training amounts to a separate act of exploitation requiring authorization. The Italian law leaves this unresolved. The issue is currently under judicial scrutiny in Europe, including before the Court of Justice of the EU, and forthcoming decisions are expected to clarify whether TDM exceptions extend to the training of large language models. In this sense, Article 25 functions more as a policy signal than as a concrete innovation.

As part of the same legislative package, Article 26(3) of the new law amends the copyright criminal provisions under Article 171 ICL by introducing a new letter a-ter, which criminalises unlawful reproductions or extractions of texts and data in breach of TDM provisions, “including when carried out through AI systems.” Under Article 171 ICL, such offences are punishable by fine of up to EUR 2.065 (imprisonment of up to 1 year is provided as an aggravating sanction in cases of work modification or prejudice to the author’s honour).

“Except as provided for in articles 171-bis and 171-ter, whoever, without having the right to do so, for any purpose and in any form, is punished with a fine from EUR 51 ( to EUR 2.065: [...] a-ter) reproduces or extracts text or data from works or other materials available on the network or in databases in violation of articles 70-ter and 70-quater, including through artificial intelligence systems;”

Despite its symbolic rather than effective sanction, the violation of TDM provisions may become relevant from a criminal perspective, with criminal provisions maintaining a more deterrent effect. This provision thus tightens the TDM framework.

Key takeaways on TDM

• The new provision is best read as a policy signal rather than a substantive innovation, apparently pro-rightholders, and reiterates the direction taken by the AI Act on copyright.
• The scope of the AI-TDM exception remains unclear though, in particular whether the training of AI models falls within TDM.
• Article 26(3) further tightens the framework by introducing criminal sanctions for violations of TDM provisions, tightening the TDM framework.

Authored by Francesco Banterle.