What the EU AI Act requires of a conversational wellbeing system from August 2026

Since 2 August 2026, Article 50 of the European Union’s Regulation (EU) 2024/1689 on artificial intelligence requires any conversational system placed on the European market to inform the person, in clear language and no later than the first interaction, that the reply comes from a machine. The obligation also covers the marking of synthetic content in a machine-readable format.

That date arrived surrounded by confusion. The Digital Omnibus on AI, in force since 27 July 2026, moved the high-risk regime to December 2027 and August 2028, and much of the coverage read it as a general delay to the regulation. Article 50 transparency has applied since August on its original terms.

What Article 50 actually requires

Article 50 governs the transparency of certain systems according to how they relate to people, whatever sector they operate in. It imposes four distinct duties.

• Interaction with people. Providers must ensure the user knows they are dealing with an AI system, unless this is obvious from the context to a reasonably observant person.
• Synthetic content. Artificially generated audio, image, video or text outputs must be marked in a machine-readable format and be detectable as such.
• Deepfakes. Deployers of content depicting real people, places or events must disclose that it has been generated or manipulated.
• Emotion recognition and biometric categorization. Deployers of these systems inform the people exposed to them and process their data in line with data protection law.

The legal alert Goodwin published in August 2026 confirms that these obligations are in force and puts the penalties at up to 15 million euros or 3% of total worldwide annual turnover, whichever is higher. A Ropes & Gray analysis from the same month, focused on putting the article into practice, locates the problem in the interface: the notice has to be perceptible where the conversation happens.

There is a single reprieve. Systems already on the market before 2 August 2026 have until 2 December 2026 to implement the technical marking of synthetic content.

Why transparency carries different weight in emotional wellbeing

In a search engine or an online shop, an automated system notice works as an administrative formality. In a conversation about distress it performs a technical function, and that is where the mechanism sits: someone who does not know what they are talking to cannot calibrate how much weight to give what they receive.

A person going through a hard night who receives an encouraging sentence reads it one way if it comes from a professional and another way if it is produced by a language model trained to sustain fluent conversation. Attribution of authority shifts with the sender. That calibration is part of the safety of the system, because it determines what the person does next with what they have read.

The analysis of conversational AI, ethical challenges and responsible development that yeshcube published in 2025 framed the criterion as operational transparency: the interface must identify the automated nature of the interaction, its limits, and the routes for review or complaint. Article 50 covers the first of the three requirements. The known limits of the system and the complaint route remain a design decision for whoever deploys it.

There is one point where the regulation does enter emotional territory. Since February 2025 it has prohibited inferring emotions in the workplace and in educational settings, with medical and safety exceptions, and Article 50 requires informing anyone exposed to an emotion recognition system operating outside those prohibited contexts.

What the Digital Omnibus postponed and what it left untouched

The Council gave its final green light on 29 June 2026 and the act was signed on 8 July. It was published in the Official Journal on 24 July and entered into force three days later, on the 27th, as Regulation (EU) 2026/1744, amending Regulations (EU) 2024/1689, (EU) 2018/1139 on civil aviation and (EU) 2023/1230 on machinery. Its effects on the calendar are four.

• Stand-alone high-risk systems under Annex III, covering employment, education, creditworthiness, essential services and law enforcement, move from 2 August 2026 to 2 December 2027.
• High-risk systems embedded in regulated products under Annex I move to 2 August 2028.
• Two practices are prohibited from 2 December 2026: systems intended to generate or manipulate non-consensual intimate imagery, and those intended to produce child sexual abuse material.
• Machinery moves from Section A to Section B of Annex I, so the high-risk obligations no longer apply to it directly and the requirement is channeled through its sectoral regulation. Medical devices and toys remain inside the regime.

The Jones Walker analysis of the real scope of the delay sums it up precisely: the Omnibus left most of the transparency obligations exactly where they were.

The intervening stretch, August 2026 to December 2027

For some sixteen months, a conversational system accompanying a person through a difficult moment can comply with the entire regulation by means of one line at the start of the conversation.

The postponement has a stated and understandable cause. The harmonized standards meant to serve as the technical reference for the high-risk regime were not finished, and demanding compliance with a regime whose standard was still being drafted would have produced an obligation that was hard to verify. The consequence follows from the calendar and the standardization machinery, rather than from a decision about sensitive uses.

Throughout that stretch the GDPR, the prohibitions in force since February 2025 and Article 50 itself continue to apply. The gap that stays open is the one between declaring and demonstrating. A system can say what it is without any authority yet asking it for the assessment documentation, the risk management, the effective human oversight or the independent review that the high-risk regime would have required. Anyone deploying in delicate contexts decides during this period whether those practices form part of their development or wait until December 2027.

The evidence threshold and the date in the journal

The algorithmic empathy that defines Somia consists of interpreting context signals without simulating emotions and without hiding that the reply comes from an AI. The obligation that entered into force in August coincides with that earlier design decision.

On the ERL scale, regulatory and ethical compliance is one of the six dimensions assessed before a transfer is authorized, alongside scientific evidence, user acceptance, technical feasibility, economic sustainability and social impact. Nothing is transferred below ERL-3. That threshold is set by the evidence gathered and stays the same when an application date shifts by sixteen months.

The European calendar kept in force the duty to declare what a system is and postponed the duty to demonstrate what it does. For a conversational wellbeing solution, the second question is the one that decides whether it can accompany someone through a bad moment, and the answer depends on the evidence assembled before deployment.

Research on assessment and responsible deployment

yeshcube develops this line within Allies, its scientific collaboration system, with four partner types and three principles: value for value, traceability and independence. No partner can veto a publication.

Work on assessing conversational systems in delicate contexts is of interest to scientific teams researching auditing and the measurement of effects, to organizations deploying these systems that need to document their limits, and to transfer partners seeking to establish evidence before the high-risk regime applies in December 2027.

Discover our Allies program →

Fill in our contact form →