Who checks AI-written texts

In my work, I encounter companies that use artificial intelligence without yet having adopted organisational procedures and governance processes.

On 5 October 2026, OpenAI announced the gradual introduction of textGrain in eligible text outputs from ChatGPT and Codex in the European Union. The technology embeds a statistical signal in the model’s linguistic choices, which may remain detectable even after metadata has been removed.

When AI use must be disclosed

Article 50 of Regulation (EU) 2024/1689, the AI Act, distinguishes the obligations of the provider from those of the person or organisation using the system in a professional capacity, the deployer. The provider is responsible for technically marking outputs in the cases covered by the Regulation, subject to the exceptions it sets out.

For the deployer, the obligation concerning text applies to AI-generated or AI-manipulated texts published to inform the public on matters of public interest. Corporate communications, resolutions and accounting documents must therefore be assessed according to their content and the purpose of publication.

The disclosure obligation does not apply where both conditions are met: the text has undergone human review or editorial control, and a natural or legal person holds editorial responsibility for its publication.

For professional services involving intellectual work, Italian law also imposes a specific national obligation: information about the AI systems used by the professional must be communicated to the recipient of the service in clear, simple and comprehensive language, under Article 13(2) of Italian Law No 132 of 23 September 2025. Human review of the text does not remove this separate duty to inform.

For breaches of transparency obligations, Article 99 of the AI Act provides for administrative fines of up to EUR 15 million. Specific rules applying the lower limit apply to SMEs, including start-ups, and to small mid-cap enterprises. The actual amount depends on the circumstances of the infringement.

A client or competitor who has grounds to consider that the AI Act has been infringed may also lodge a complaint with the competent market surveillance authority. Legal proceedings remain possible where the relevant conditions are met. As regards compensation, Italian Legislative Decree No 160 of 9 September 2026 provides procedural tools for damage caused by the use of AI systems: the mere presence of a watermark proves neither damage nor liability.

For professionals and businesses, the practical response is to organise the work. Retain sources and significant versions, clarify permitted uses of AI and identify who checks and approves each publication. An inventory of the tools used, their purposes, the data processed and the people responsible helps identify the applicable obligations. Documentation should then be tailored to the organisation’s role, the system and its particular use.