Transparency obligations under Article 50 of the EU AI Act came into force on 2^nd August.^[1] These obligations apply to relevant providers and deployers of AI systems across four specific contexts. Given how widely these systems are used in daily life, transparency in these contexts is crucial.

In this regard, obligations for providers of the first category which the accompanying guidelines on transparency obligations define as ‘interactive AI systems’ (‘IAS’) are worth looking at more closely. Addressing the risks posed by IAS is a pressing concern, given the growing risks of emotional dependency, and other forms of psychological harm^[2] that these systems have been linked to. The EU itself has expressed concern regarding the growing rise of AI companions^[3], and excessive use, especially among children and more vulnerable users. As lawsuits in the USA mount, alleging that chatbot design contributed to psychological harm^[4], and other evidence globally suggests that more people are relying on AI chatbots for affective interactions^[5], including children, the policy stakes are high.

The transparency obligations for IAS apply if four conditions are cumulatively satisfied. First, the system must qualify as an AI system under the AI Act. Second, it must be designed for a genuine two-way exchange with people, rather than merely collecting data or providing automated responses. The interactions could occur in various forms (for example, audio, visual or text-based). It could also be a one-time exchange, or a multi-turn conversation. Third, the interaction must be direct, meaning the AI itself communicates with the person rather than through a human intermediary. Fourth, the interaction must be with natural persons, whether consumers, professionals, or other users.  AI systems operating solely in the background, through machine-to-machine communication, or without direct contact with people, fall outside the scope of this obligation.

The scope of this definition is a welcome development, clearly bringing in the various types of IAS, from general-purpose models like OpenAI’s ChatGPT or Anthropic’s Claude, to systems which are specific to serving a ‘companionship’ function, like Character.Ai or Replika. The lack of any specific exclusion would also mean that platforms where users can create their own AI companions would also fall within these transparency requirements. The obligations also suggest a contextual approach to transparency, accounting for the user or user group interacting with the IAS in question, which is another welcome move. Disclosure techniques should account for user vulnerabilities, and be implemented in a proportionate way, calibrated according to the audience, ensuring users are “explicitly informed about the artificial, non-human nature of the interacting counterpart”.^[6]

While this contextual approach to transparency —which has been previously advocated for by scholars in the context of AI systems^[7]— is desirable considering the dynamic and personalized nature of these IAS and the risks of emotional dependency, the exceptions created for these obligations create some worrying loopholes. There are two separate problems here: first, the ‘obviousness’ exception may be too easy for open-ended systems to invoke; and second, even where a disclosure is present, it may not address the actual source of harm. Moreover, the very fact that IAS is only addressed in the AI Act through the lens of transparency misses the larger driver of potential harm through these systems.

An exception is created for ‘obvious interactions with an AI system’. To rely on this exception, a provider of IAS will need to demonstrate “(i) the obvious artificial nature of the interaction to (ii) a natural person who must be reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use.”^[8] This standard draws on EU consumer protection law, and the notion of an average consumer under the Unfair Commercial Practices Directive. The guidance on applying this exception is that it should be applied restrictively, considering the audience with whom the AI system is ‘intended and reasonably likely to interact’.

Upon establishing the “hypothetical reasonably well-informed, observant and circumspect natural person, providers should evaluate whether it will be obvious to such a person that they are directly interacting with an AI system”. The factors to assess this would be the anticipated nature of the interaction, composition of the target and foreseeable audience, and the level of difficulty in knowing whether an interaction with an AI system or an actual human being is taking place. The guidance states that the “obviousness” exception should be limited to cases where there is almost no doubt left about the nature of the interaction for an average person from the targeted and reasonably foreseeable audience.”

Transparency and disclosing that users are interacting with AI is important, particularly for children. However, conceiving of a hypothetical average, circumspect natural person seems to run contradictory to contextual transparency, in principle. For an IAS that is limited to a very controlled function, for example a chatbot that answers customer support queries, or a specialist system meant to be used by a particular type of professional, the ‘obviousness’ criterion might be one that makes sense. When we expand this to uncontrolled systems (general-purpose LLMs and AI companion chatbots) this category of an average, rational user, may be illusory since these systems are precisely designed to be personalised to the specific user in question. Regardless, given how the exception is worded, providers of these open-ended systems could also potentially claim this exception. For example, it might be seemingly ‘obvious’ that a chatbot on Character.ai or Replika is AI, since the platform is clearly intended for the purpose of creating an AI companion with a digital avatar/voice, etc. A reasonably informed person, then, could be assumed to know this given the anticipated nature of the interaction.

Going a step further, the problem is not so much in relation to the applicability of the obviousness exception, but the idea that transparency will lead a reasonable user of IAS to take informed decisions regarding their interaction with the system and calibrate their trust in the content accordingly. Even assuming that all users of IAS are such rational people, this idea of transparency continues to frame the policy problem to be solved as one of deception — that if people are simply informed of the nature of the system, it will mitigate risk.

It is crucial for regulators to recognise that potential psychological harms from affective interactions with AI do not arise because users believe they are interacting with a human. The interaction between users and AI is voluntary and continually co-constructed, and such bonds can form even without a user’s original intent to seek one out. Emotional dependence on AI often develops incidentally within routine, task-oriented interactions.^[9] Muldoon terms this as the ‘synthetic social’ — with synthetic personas, he says, “a part of us may know they are not real, but our connection with them activates a more primitive behavioural response pattern, based on their perceived feelings for us.”^[10]

Even where transparency-driven disclaimers are built into the system, the language chatbots generate is such that it signals some form of feeling on the system’s part.^[11] Users therefore choose to continue interacting with the system despite knowing they are speaking to AI, because it offers frictionless emotional support. Framing this, from a policy or design perspective, primarily as a problem of deception therefore misses the point: the more pressing question is why users continue to find value in an interaction they know to be artificial, and what that reveals about the support they are, or are not, otherwise able to access.

I have argued elsewhere that the relationship between users and Social AI tools, and the tendency to interpret machine interactions in personal ways, blurs the line between persuasion and addictive design and challenges traditional notions of harm driven by causation.^[12] Natale highlights the issues which arise from this form of ‘banal deception’, which is often not overtly malicious and may often contribute some value to the user.^[13] However, the effects of banal deception are argued to be deeper and wider than straightforward delusion since AI developers have the potential to affect the deeper structures of social life and experience.^[14]

Therefore, more than transparency, we need ex ante governance approaches which are informed by an understanding of design features^[15] and how users perceive them — examining, for instance, language/features that could create dependency, age-appropriate defaults, friction around prolonged use, escalation pathways in crisis situations, and testing for relational harms before deployment — to develop obligations around harm-prevention by regulating platform design.  China’s recent guidance on human-like interactive AI systems provides useful guidance on design obligations.[16]

Further, policymakers need to be adopting a wider, societal view of potential psychological harm as arising from the design of Social AI, moving beyond the focus on individuals to effects on collective societal resilience.  Social media has now reached the point where courts and regulators are increasingly being asked to consider the effects of platform design as a societal-level harm^[17] requiring structural changes. Before we reach a higher level of harm, proactive regulation on IAS should be next, requiring deeper thinking about governance which addresses how these systems are used in a relational capacity, rather than as assistive, functional tools.

Footnotes

[1] European Commission, ‘Guidelines on the Implementation of the Transparency Obligations for Certain AI Systems under Article 50 of the AI Act’ (2026) <https://digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems> accessed 28 August 2026.

[2] Rhiannon Williams, ‘It’s Surprisingly Easy to Stumble into a Relationship with an AI Chatbot’ MIT Technology Review (24 September 2025) <https://www.technologyreview.com/2025/09/24/1123915/relationship-ai-without-seeking-it/> accessed 28 August 2026. See also Robert Mahari and Pat Pataranutaporn, ‘Addictive Intelligence: Understanding Psychological, Legal, and Technical Dimensions of AI Companionship’ MIT Case Studies in Social and Ethical Responsibilities of Computing (Winter 2025) <https://mit-serc.pubpub.org/pub/iopjyxcx/release/2> accessed 28 August 2026.

[3] Mar Negreiro and Öykü Dilara Anaç, ‘The Spread of AI Companions and the Challenges They Generate’ (European Parliamentary Research Service, EPRS_BRI(2026)789299, May 2026) <https://www.europarl.europa.eu/RegData/etudes/BRIE/2026/789299/EPRS_BRI(2026)789299_EN.pdf> accessed 28 August 2026.

[4] Kristie Carrier v OpenAI Foundation, Complaint, Superior Court of the State of California, County of San Francisco <https://techjusticelaw.org/wp-content/uploads/2026/06/2026-06-11-Kristie-Alice-Carrier-v.-OpenAI-Complaint.pdf> accessed 28 August 2026; Garcia v Character Technologies Inc, 2025 WL 1461721 (MD Fla, 21 May 2025); Michael Lines v OpenAI Foundation, Superior Court of the State of California, County of San Francisco <https://techjusticelaw.org/press-releases/openai-california-disability-law-safety-lawsuit-chatgpt-product-liability-negligence/> accessed 28 August 2026; Raine v OpenAI, Superior Court of the State of California, County of San Francisco <https://cdn.sanity.io/files/3tzzh18d/production/5802c13979a6056f86690687a629e771a07932ab.pdf> accessed 28 August 2026.

[5] Tomisin Olanrewaju, ‘How Countries Are Regulating AI Companion Chatbots to Protect Children’ (IAPP, 20 May 2026) <https://iapp.org/news/a/how-countries-are-regulating-ai-companion-chatbots-to-protect-children> accessed 28 August 2026.

[6] European Commission (n 1) 11.

[7] Heike Felzmann, Eduard Fosch Villaronga, Christoph Lutz and Aurelia Tamò-Larrieux, ‘Transparency You Can Trust: Transparency Requirements for Artificial Intelligence between Legal Norms and Contextual Concerns’ (2019) 6(1) Big Data & Society.

[8] European Commission (n 1) 12.

[9] Y Shi, CM Fang, P Maes and A Goldenberg, ‘Stumbling into AI Emotional Dependence: How Routine AI Interactions Reshape Human Connection’ (2026) arXiv:2606.04150; Williams (n 2).

[10] James Muldoon, Love Machines: How Artificial Intelligence Is Transforming Our Relationships (Faber and Faber 2026) 160.

[11] Michal Luria, ‘AI Chatbots Are Emotionally Deceptive by Design’ (Tech Policy Press, 29 August 2025) <https://www.techpolicy.press/ai-chatbots-are-emotionally-deceptive-by-design/> accessed 28 August 2026.

[12] Sriya Sridhar, ‘AI Fought the Law and Did the Law Win? Navigating the Complexity of Regulating Social AI’ (The Pranava Institute, 2025) <https://www.pranavainstitute.com/post/new-blog-ai-fought-the-law-and-did-the-law-win-navigating-the-complexity-of-regulating-social-ai> accessed 28 August 2026.

[13] Simone Natale, Deceitful Media: Artificial Intelligence and Social Life after the Turing Test (Oxford Academic 2021).

[14] ibid.

[15] Andrew McStay, ‘Intimate Functionality: A Phrase Defining the Next Year of AI Regulation’ (Tech Policy Press, 14 July 2026) <https://www.techpolicy.press/intimate-functionality-a-phrase-defining-the-next-year-of-ai-regulation/> accessed 28 August 2026.

[16] Jeremy Daum. The Friends We’ve Made 2: Changes from the Draft’ (China Law Translate 13 March 2026) < https://www.chinalawtranslate.com/en/friends-we-made/> accessed 20 July  2026.

[17] PF (KGM) v Meta Platforms Inc, Case No 22STCV33140 (Superior Court of California, County of Los Angeles). See also State of New Mexico v Meta Platforms Inc, No D-101-CV-2023-02838 (First Judicial District Court, State of New Mexico, County of Santa Fe).

Marcin Wilkowski / https://betterimagesofai.org / https://creativecommons.org/licenses/by/4.0/