In early June 2026, the Italian data protection authority halted a company that had built a system to infer employees’ psychological stress levels by automatically analyzing their conversations on Slack and Microsoft Teams. The case, reported by Red Hot Cyber, raises a question that goes beyond Italy. How far can AI emotion recognition at work go before it becomes an unlawful intrusion on the person?
The answer is starting to take definite shape, and any security leader evaluating tools that measure the human factor should understand it before signing a contract.
What exactly did the Italian data protection authority halt?
The product under review promised something that sounds reasonable at first glance. An artificial intelligence engine read the semantic content of internal chats and returned aggregated reports to the employer about the team’s stress level, without handing over individual conversations.
The Italian data protection authority did not stop at that promise of aggregation. It observed that, to produce those reports, the system had to process the emotional content of each person’s private communications, and that this processing opened indirect access to highly sensitive information. On that basis, it invoked privacy law, the Workers’ Statute and the European artificial intelligence regulation, and required the company to put adequate measures in place from the design of the service to prevent any access to emotional information.
Three specific concerns underpinned the measure:
- The opacity of the semantic analysis model, which prevents the worker from understanding how what they write is interpreted.
- The risk of making discriminatory decisions based on that interpretation.
- The limit that the Workers’ Statute places on remote monitoring of staff activity.
It is worth being precise about what we mean. Emotion recognition is the use of an automated system to deduce or identify a person’s affective state (stress, anxiety, anger, enthusiasm) from their data, without that person having declared it. It does not measure what someone does, but what is inferred about how they feel when they speak or write in a certain way.
Why is inferring emotions at work a category of its own?
When an organization measures its people, it helps to distinguish two planes that are often confused:
- Observable conduct. Whether a person reported a suspicious email, completed training or clicked on a simulation. These are verifiable facts, tied to a concrete task.
- Inferred internal state. Whether that person is stressed, distracted or emotionally vulnerable. It is a conjecture about their psyche, not a fact about their work.
The first plane is the legitimate ground of a security program. The second enters territory that the European legislator has already expressly delimited. Article 5 of the Artificial Intelligence Regulation prohibits the use of AI systems to infer the emotions of a natural person in the workplace and in education, with the sole exception of medical or safety purposes. That prohibition has been in force since 2 February 2025; it is neither a future recommendation nor a best practice.
The aggregated-report format does not solve the underlying problem either. As the Italian authority itself observed, to deliver a team average the system first infers each member’s emotional state, and it is that prior individual processing that the rule reaches. What the employer ultimately sees is secondary to what the machine deduced along the way.
The logic behind the rule is understandable. In an employment relationship there is a structural power imbalance, and a worker can hardly refuse with real freedom to have their employer read their mood. Recital 44 of the Artificial Intelligence Regulation sets it out plainly:
Among the key shortcomings of such systems are the limited reliability, the lack of specificity and the limited generalisability. […] AI systems identifying or inferring emotions or intentions of natural persons on the basis of their biometric data may lead to discriminatory outcomes and can be intrusive to the rights and freedoms of the concerned persons. Considering the imbalance of power in the context of work or education, combined with the intrusive nature of these systems, such systems could lead to detrimental or unfavourable treatment of certain natural persons or whole groups thereof.
That risk of discriminatory decisions, combined with the opacity of the models that infer emotions, is what led to classifying this practice among the unacceptable ones, not among those that are merely regulated.
What does the legal framework say about monitoring employees with AI in Latin America?
No country in the region yet has a direct equivalent to the European regulation. It would be a mistake, however, to conclude that in Latin America this practice falls outside legal control.
A person’s psychological state constitutes sensitive data, and in several of the region’s legal systems it is treated as data relating to health, subject to a reinforced protection regime. In Argentina, for example, Law 25.326 on the Protection of Personal Data subjects sensitive data to strict processing conditions, since its improper use enables the very discrimination the law seeks to avoid.
Consent, moreover, does not work as a key that opens everything. Within a relationship of dependency, its validity is called into question precisely because of the asymmetry noted above. Added to this is that several countries in the region recognize data protection at the constitutional level through the habeas data action, which gives these principles a standing that is hard to bypass through a contract.
On top of that comes a principle that runs across the region and that labor case law has long recognized. Corporate monitoring must be limited to what is connected with the task and must respect the worker’s dignity, and inferring emotions with AI strains that principle at its core.
The European experience, in this sense, works as a directional signal. It marks where the global regulatory conversation is heading, and anticipates the standard the region will end up adopting. We already analyzed the purely regulatory side of this discussion in psychological profiles on awareness platforms.
Where does well-designed awareness fit?
The news could be read as a bad sign for any program that measures human behavior, but that reading confuses two different things. What the Italian case questions is the emotional profiling of the person, and the risk that an interpretation error poses to their rights and guarantees. Measuring observable conduct is on another plane.
A mature awareness program is built on the legitimate plane described above. It measures observable conduct in controlled scenarios (a phishing simulation, the completion of a module, the reporting of an incident) and works with aggregated, anonymized data that describe the collective, not an individual’s private life. It is worth clarifying that this is not only a compliance best practice, but the difference between a defensible program and one that exposes the organization to a claim.
That boundary is what sustains the legitimacy of monitoring. We developed it some time ago when discussing monitoring, awareness and the expectation of privacy, and it connects with an idea that runs through our work. Human risk in cybersecurity is reduced by understanding how a person decides in front of the screen, not by auditing a mood that, as in anyone, changes from one day to the next.
SMARTFENSE is a cybersecurity awareness platform oriented to Latin America and Spain, designed to measure and change behavior without psychologically profiling users. That architectural decision, far from being a detail, is what allows an organization to run the program with legal backing.
The limit an awareness program should not cross
The Italian case leaves a practical conclusion for whoever selects tools that measure staff behavior. The right question to ask a vendor offering to measure the human factor is not only how accurate its model is, but what data it processes to reach the result.
If the answer involves inferring emotions, psychological traits or moods of each person, the product does not solve a security problem. It adds a compliance one, and in Europe it has already crossed a prohibited line. Effective awareness never needed to read anyone’s mind. It is enough to observe what people do and give them the chance to do it better.
Leave a Reply