Annex VI to Part Five of the Environmental Code (Legislative Decree No 152 of 3 April 2006) already provides that the continuous monitoring system validates data automatically: point 3.7.2 states that the validation system must validate the elementary values and the hourly averages “automatically, on the basis of predefined verification procedures”, and that the validation procedures “must be established by the competent control authority, after consulting the operator” (Normattiva). The automation of environmental data therefore predates artificial intelligence and comes with a precise rule on who decides. When an AI model estimates missing concentrations, discards values it judges to be anomalous or determines when an alarm should be triggered, the question arises whether that rule is respected and who is answerable for the data that reach the authority.
Monitoring obligations: from the Directive to Legislative Decree No 152 of 2006
Directive 2010/75/EU requires the permit to include emission monitoring requirements specifying “measurement methodology, frequency and evaluation procedure”, together with an obligation to supply the authority, at least annually, with the information needed to verify compliance (Article 14(1), points (c) and (d)). For large combustion plants, Annex V, Part 3, requires the quality assurance of automated measuring systems to follow CEN standards, with parallel measurements using the reference methods at least once a year, and treats as invalid any day in which more than three hourly average values are invalid due to malfunction or maintenance (EUR-Lex).
In Italy, Article 29-sexies(6) of Legislative Decree No 152 of 2006 transposes these contents for the integrated environmental authorisation (Normattiva); Article 29-decies(2) places on the operator the transmission of monitoring data in accordance with the arrangements and frequencies laid down in the authorisation, and paragraph 3 entrusts ISPRA, for installations falling within State competence, or the competent authority, which relies on the regional and provincial agencies, with verifying the “regularity of the controls incumbent on the operator” (Normattiva). For establishments authorised under Part Five, Article 271(17) refers to Annex VI and allows, under the conditions set out in that paragraph, exceedances to be established by means of a continuous monitoring system that complies with the quality assurance procedures of standard UNI EN 14181 (Normattiva); for installations subject to integrated environmental authorisation, Article 267(3) preserves the rules of Title III-bis (Normattiva).
Technical standards and reference documents
UNI EN 14181:2015, the official English-language version of EN 14181, concerns the quality assurance of automated measuring systems for emissions from stationary sources; according to its public record, it specifies the procedures for calibration and determination of variability (QAL2), for ongoing quality assurance during operation (QAL3) and for the annual surveillance test (AST) (UNI). Annex VI makes large combustion plants, cement works, glassworks and steelworks subject to it and requires the suitability of analysers to be certified in accordance with UNI EN 15267.
In July 2018 the Joint Research Centre published the Reference Report on Monitoring (ROM), which offers practical guidance on applying the BAT conclusions on monitoring and helps authorities define monitoring requirements in permits (JRC). The BAT conclusions for large combustion plants, adopted by Implementing Decision (EU) 2021/2326, define the predictive emissions monitoring system (PEMS) as a “System used to determine the emissions concentration of a pollutant from an emission source on a continuous basis, based on its relationship with a number of characteristic continuously monitored process parameters”, and allow its use in certain cases (EUR-Lex). EU law already recognises emission data estimated by a model, within the limits set by the BAT conclusions. In our view, a machine learning model that derives concentrations from process parameters falls, by virtue of its function, within that definition and is subject to the same conditions.
What artificial intelligence does with the data: estimating, filtering, flagging
The functions offered by vendors can be reduced to three. The first is estimation: reconstructing concentrations when the analyser is out of service, or replacing it with a model. For establishments under Part Five, Annex VI, point 2.5, provides that, where continuous measurement is not possible, the operator shall carry out alternative controls based, among other things, on “correlations with operating parameters”, but it is the competent control authority that establishes the estimation procedures, after consulting the operator; under point 2.6, data estimated in this way contribute to the verification of the limit values. The second is filtering: discarding implausible values, an activity that point 3.7.2 entrusts to validation procedures established by that same authority and, for large combustion plants, to thresholds set by it. The third is flagging: point 3.7 requires the system to manage alarms and anomalies, and point 3.7.1 requires data entered by the operator to be archived according to the same criteria as the other parameters (Normattiva).
In all three cases the method is not an internal technical choice. For Part Five, Article 271(18) provides that, where the operator uses monitoring methods or systems that differ from, or do not comply with, the requirements of the authorisation, “the results of their application are not valid”, and the penalty under Article 279(2-bis) applies (Normattiva). For installations subject to integrated environmental authorisation, the methodology is laid down in the authorisation, and departing from it may, in our view, amount to the failure to comply with the permit conditions penalised by Article 29-quattuordecies(2) (Normattiva). Introducing, without going through the authority, a model that estimates or filters data means changing the monitoring method.
AI Act: as a rule, not a high-risk system
Emissions monitoring does not appear among the areas listed in Annex III to Regulation (EU) 2024/1689, the AI Act, point 2 of which concerns safety components in the management and operation of critical digital infrastructure, road traffic, and the supply of water, gas, heating or electricity (EUR-Lex). We therefore consider that an emissions monitoring model is not, as a rule, high-risk. Article 4, as replaced by Regulation (EU) 2026/1744, nevertheless applies and requires deployers to support the development of AI literacy of their staff (EUR-Lex); on this point we refer to our earlier analysis of AI literacy after the Digital Omnibus.
The recording of events (logs) under Article 12 and human oversight under Article 14 concern high-risk systems only, but they offer a useful vocabulary: Article 14(4), point (b), uses the term “automation bias” for the tendency to rely excessively on the output of the system, and this is precisely the risk posed by an operator who accepts without verification the validation proposed by the model. On the environmental side, traceability is already an obligation: Annex VI requires registers of interventions and interruptions and the retention of measured and processed data for at least five years, unless the authorisation provides otherwise (points 2.7, 2.8 and 5.4).
Who is answerable for the transmitted data
The transmitted data belong to the operator. Article 29-quattuordecies(8) penalises the operator who fails to communicate measurement data, with a reduction to one tenth where the communications are late by no more than sixty days or “formally incomplete or inaccurate” but contain the essential elements; paragraph 9 applies the penalty under Article 483 of the Criminal Code, imprisonment of up to two years, to anyone who, in those same communications, “provides falsified or altered data” (Normattiva; Normattiva). For Part Five, Article 271(20) requires any non-conformities found in the operator’s monitoring to be communicated within 24 hours. Finally, Article 452-bis of the Criminal Code punishes anyone who “unlawfully causes a significant and measurable impairment or deterioration” of water or air (Normattiva): the requirement of measurability makes monitoring data, in our view, a central element of criminal proceedings as well.
None of these obligations passes to the provider of the model. The contract may govern the financial consequences of an error; the obligations towards the authority remain with the operator.
Critical aspects: reconstructed data and the delegated alarm threshold
The first aspect concerns the value of reconstructed data. An estimate produced in accordance with a procedure established by the authority contributes to the verification of the limit values; an estimate produced by a model of which the authority is unaware exposes the operator, in our view, to the risk that the results will be declared invalid. Reconstructed data count in inspections only as much as the procedure the authority has approved for producing them.
The second concerns false negatives. If the model classifies a genuine peak as an instrumental anomaly, the breach is not communicated, whereas Article 29-decies(2) requires the authority to be informed immediately. Distinguishing a model error from data tampering will require reconstructing what the system discarded, with which version and with which thresholds: retaining discarded values alongside validated ones is the precondition for demonstrating the correctness of one’s own conduct.
The third concerns the alarm threshold. Annex VI reserves to the competent control authority the validation procedures and, for large combustion plants, certain thresholds, to be set according to the process and the measuring system; a model that adjusts its own thresholds moves that decision outside the intended perimeter. The threshold beyond which a data point becomes an alarm must remain a documented choice attributable to a person. Article 29-quattuordecies(4), point (c), moreover, increases the penalties where exceeding the emission limit values also causes the air quality limit values to be exceeded, which are the subject of the new European standards for 2030. Finally, within the environmental management system, ISO 14001:2026 lists, in clause A.7.1, pollution control systems, databases and software among technological resources.
Conclusions
Artificial intelligence applied to emissions monitoring is not, as a rule, high-risk under the AI Act, but it operates within an environmental framework that already regulates the estimation, validation and retention of data and that assigns methodological choices to the control authority. The provisions referred to do not mention machine learning models, and they apply to them as methods of estimation and validation. The operator remains responsible for the transmitted data, for timely communications and for keeping the registers; the authority remains responsible for approving the procedures. It is advisable to submit to the authority any model that estimates or filters data intended for inspections, to document its versions and thresholds, to retain discarded values and to entrust the validation of anomalies to an identified person.
Author: Valentina Grazia Sapuppo
