From Today, AI Documentation Is Also Evidence
Legislative Decree 160/2026 enters into force. The same documents a month-long series described as voluntary standard requirements become the material a company uses to support its position in court.
In short. Legislative Decree 160/2026 is in force in Italy from today, and in a damages claim involving an AI system a court can order disclosure of logs, risk management documentation, technical documentation and human oversight arrangements. The same documents a month-long series described as requirements of a voluntary standard become, in that forum, the material a company uses to support its position. This article revisits the series from that angle.
For a month we wrote about inventories, impact assessments, logging, human oversight, competence and certification, with the argument compliance usually makes: it's needed for the audit, better done early, more expensive to reconstruct later. From today that argument has a second dimension, worth setting out now that the decree is operative.
What changes today
Legislative Decree of 9 September 2026, No. 160, published in the Official Gazette on 15 September, enters into force on the 30th. Articles 16 to 20 govern civil damages claims for harm caused through the use of an AI system, described in detail in our article on evidence and causation.
Three mechanisms work together. A court can order disclosure of evidence about how the system operated, and a party failing to comply without justified reason risks having the facts alleged against it treated as established. Where the harm stems from a breach of AI Act obligations, causation is presumed, subject to proof to the contrary. Conformity with the regulation, even when certified, does not in itself exclude liability.
Two limits are worth keeping in mind, because the effective perimeter is narrower than headlines suggest. The causation presumption operates only where harm follows from an AI Act breach, and several high-risk obligations apply from December 2027 for Annex III and August 2028 for Annex I. Access to evidence, by contrast, applies immediately to any damages claim tied to the use of an AI system, whatever its risk class.
The same documents, two readers
The four document types Article 17 lists as disclosable are those the AI Act already requires: automatic logs, the risk management system, technical documentation, and human oversight parameters and arrangements. They are also, almost one for one, what a management system aligned with ISO/IEC 42001 produces across the life cycle.
The difference lies in who reads them and with what question. An auditor checks that the management system works and samples to establish it, working from conformity to a model. A court looks at a single episode and asks what happened that day, on that input, with what human intervention. The first can accept that a procedure describes a control; the second wants the trace of that control in the specific case.
This distance explains why some documents entirely adequate for an audit turn weak in litigation. An approved and current human oversight procedure passes an audit; if nobody records the actual interventions, though, it doesn't show that anyone looked that day.
Where the series shifts weight
The AI system inventory served to know what needed governing. From today it also serves, when a challenge lands on one specific system, to know who owns it, what data it handles and where its logs sit. A company unable to answer those questions within a reasonable time starts at a disadvantage.
The choices described in designing an ISO-ready AI system matter more than they appeared to. A log recording only the output won't let anyone reconstruct a decision, and that moves from technical limitation to evidentiary difficulty. The same applies to oversight designed without traceable means of intervention.
On competence, clause 7 and the roles-competence matrix connect to human oversight. Demonstrating that oversight existed also means showing that whoever exercised it had the means to, which leads back to the individual evidence for critical roles covered in AI literacy and ISO 42001 competence.
On certification, Article 19 settles the question explicitly. The articles on the map of accredited bodies and the three certification paths remain valid for what they say, and should be read knowing that a certificate attests a system existed on a given date, while the defence in a specific case is built from operational evidence.
Which framework does your company actually need?
AI Rating measures maturity across the four areas of the model and shows where to start, with priorities and estimated effort.
Start your AI RatingThe question that shifts the perspective
Anyone wanting to gauge their exposure in half an hour can try a simple exercise. Pick one AI system in use, pick any date in recent weeks, then try to reconstruct a single decision taken that day: what input the system received, which model version was live, what outcome it produced, whether anyone intervened and who.
If the reconstruction requires requesting files from a vendor, if the logs have been rotated, if the human intervention appears nowhere, that is the position the company would be in facing a disclosure order. The difference is that today the exercise can be done calmly, and the result doesn't go on the record.
Three checks worth running now
The three most pressing checks concern contracts, retention periods and the traceability of interventions.
On vendor contracts, verify whether a clause guarantees access to logs and technical documentation in the event of litigation, within timeframes compatible with a court order. A vendor keeping logs for thirty days with no release obligation leaves the company without its principal material.
On retention, Article 26 of the AI Act sets at least six months for high-risk systems, which is a floor rather than necessarily the right duration relative to how long it takes for harm to surface and be contested.
On traceability of human interventions, check whether the act of oversight leaves a trace, or whether it remains a practice described in procedures and invisible in the data.
Why the method stays the same
Anyone who built a management system in order to certify already holds most of what's needed, and mainly has to verify that the evidence can be reconstructed case by case rather than only by sample. Anyone not certifying carries no equivalent documentary obligations, and still has an interest in producing the same material, because a disclosure order doesn't distinguish between certified and uncertified companies.
The inversion sits there: governance documentation stops being a cost borne to obtain a title and becomes how a company accounts, with verifiable elements, for what its system did. Which is why it's worth building when it isn't needed, because by the time it is needed it can no longer be built.
Next step
This article closes the ISO/IEC 42001 series that began on 1 September. The full index sits at the end of the article on the three certification paths, and the wider regulatory picture is in the complete guide to the AI Act.
Through AI Rating we measure a company's maturity across the AI Risk dimension too, covering risk governance, compliance and traceability. To discuss your position you can book an assessment meeting or write to hello@zerofive.ai.
This article describes the legislative text and its organizational implications, and does not constitute legal advice.
Sources. Legislative Decree of 9 September 2026, No. 160, Official Gazette, General Series, No. 214 of 15 September 2026, Articles 16-20; Regulation (EU) 2024/1689 (AI Act) as amended by Regulation (EU) 2026/1744, Articles 9, 11, 12, 14 and 26; ISO/IEC 42001:2023, clauses 7 and 9, Annex A.