Author: Ewa Lejman - WH Partners
On 2 August 2026, Article 50 of the AI Act came into force across the European Union. This marks the next stage in the implementation of this regulation, which is coming into force gradually, section by section: prohibited practices have been in force since February 2025, obligations for general-purpose models since August 2025, whilst the requirements for high-risk systems - following a recent postponement of the deadlines - will, for the most part, only come into force in December 2027 (for autonomous systems under Annex III), and for systems embedded in regulated products, in August 2028 (Annex I).
The media have reported on Article 50 mainly in the context of deepfakes and customer service chatbots, reducing it to yet another labelling requirement. But could this provision also cover generative NPCs in video games - that is, virtual characters that engage in real-time conversations with the player? The answer is: yes.
As a lawyer advising the gaming industry - at both EU and Polish level - I see something more significant in this provision: the obligation to disclose the use of AI depends, amongst other things, on who the actual audience of a given game is. This was one of the topics discussed at Nordic Game 2026, which I attended - on a panel dedicated to child safety in games.
The obviousness test: when AI must be disclosed, and when it does not have to be
Article 50(1) of the AI Act essentially boils down to a single principle: the player must be aware that they are interacting with an AI. Formally, this obligation rests with the AI system provider, but in practice within the games industry it is most often fulfilled by the implementer - a studio or publisher integrating a third-party language model. This distinction matters when negotiating licence agreements with model providers, as this is where it is determined who in the chain of responsibility is actually liable for labelling. The EU legislator has, however, provided for an exception: interactions with AI do not need to be labelled if its artificial nature is ‘obvious’ to a reasonably well-informed, observant and circumspect person.
The European Commission’s final guidelines on Article 50 of the AI Act - Communication C(2026) 5054 - were approved on 20 July 2026, with the various language versions are now awaiting formal adoption. The guidelines explicitly cite an example from our industry: interactions with AI-powered NPCs in a single-player game may meet the ‘obviousness’ threshold and not require additional disclosure. The guidelines themselves are not a source of law and are not binding on the Court of Justice of the European Union. However, national market surveillance authorities will rely on them, making them an important point of reference from now on. . The condition is clear: the nature of the game must unambiguously indicate that no other person is involved in the interaction, and the artificial nature of the conversation must be obvious to all users - including vulnerable individuals.
However, these same guidelines also highlight the other side of this divide. AI-powered companion animal robots that resemble their natural counterparts do not meet the criterion of obviousness. Realistic avatars in immersive VR or AR environments - which children, older people or those with disabilities may not be able to distinguish from a human - also do not meet this criterion. The same applies to chatbots in online customer service centres. In other words: not every AI-NPC is automatically exempt from the disclosure requirement, nor does every such NPC require special labelling. The assessment of ‘obviousness’ depends on the specific context: the game mode (single-player or multiplayer, where an AI-NPC might be mistaken for another player), the realism of the character, the depth of immersion, and, above all, who the game’s audience actually is.
The guidelines explicitly state that the ‘obviousness’ threshold is lowered when the audience includes children or people with lower digital literacy. For a traditional single-player game aimed at adult players, dialogue with NPCs may therefore remain without additional labelling. The same mechanism, in a more immersive or multiplayer format popular with children, may, however, require explicit disclosure. This is the first time that EU AI legislation has differentiated the obligation so precisely depending on who is on the other side of the screen - and this is, in my view, the most practical implication of this provision for our industry. A breach of this obligation is subject to fines of up to €15 million or 3 per cent of global annual turnover - which in itself is a good argument for conducting an audit before the regulator does so.The neighbouring provision, which is discussed even less, is worth keeping in mind too. Article 5 of the AI Act has been in force since February 2025 - and thus for longer than Article 50 - and prohibits an AI system which, intentionally or in effect, significantly distorts a person’s behaviour through the use of manipulative techniques or the exploitation of a specific group’s vulnerabilities, if this causes significant harm. The absence of malicious intent on the part of the studio does not protect against an allegation of infringement if the mechanism in practice produces such an effect. The European Commission has cited AI-generated NPCs designed to steer players towards increased in-game spending as an example of a prohibited practice - with fines of up to €35 million or 7 per cent of global turnover. This is worth bearing in mind when designing systems that personalise purchases based on player profiling: the line between an accurate recommendation and a mechanism that the regulator might deem manipulative is often narrower than it appears at the product design stage - especially when the system learns to optimise conversion on its own.
The consumer, AI, the child: three perspectives on the same issue
As a lawyer who follows this sector closely, I can see the convergence of two previously separate regulatory trends. On the one hand, there is the discussion about ‘safety by design’ for children in games. This is evident in research into the co-design of safety mechanisms with children , as well as in industry panels such as the one at Nordic Game.
On the other hand, we have a compliance regime surrounding artificial intelligence, to be implemented in stages from 2025 and further developed over the coming years - in line with the timetable outlined at the beginning. This framework already affects generative mechanics in games, including conversations with AI-NPCs based on language models. These technologies are so new that they are only now being deployed among real-world player demographics , a significant proportion of whom are children. The European Parliament recognises this too. In its resolution of 26 November 2025 on the protection of minors online - P10_TA(2025)0299, procedure 2025/2060(INI) - it pointed out that randomised mechanisms and engagement systems based on returning to the game deserve particular attention from the legislator, as they exploit psychological mechanisms to which children are particularly susceptible.
It is likely that this issue will soon be the subject of its own separate piece of legislation. The European Commission is planning to present a draft Digital Fairness Act in the fourth quarter of 2026 - a new consumer regulation intended to cover, amongst other things, addictive design and randomised mechanics in digital products, including games. I have written about the DFA public consultation itself in more detail in a separate article, but the most important point is that this is no longer an isolated initiative. Article 50 of the AI Act, the European Parliament resolution and the planned DFA are three parallel paths leading in the same direction For compliance departments in the gaming industry, this means that these developments are worth analysing them collectively from the outset, rather than in isolation
Article 50 of the AI Act now follows this same line of thinking, introducing a technical ‘obviousness’ test that requires designers to take into account the actual diversity of a particular game’s audience, rather than an abstract, statistically average user. A game featuring generative NPCs, return-to-play mechanics and an engagement-based monetisation system now prompts the regulator to ask a different question from ‘Is this legal?’ - namely: legal for whom?
More specifically, was this product designed with the fact in mind that some of its users are children who are only just learning to recognise the mechanisms of persuasion? This question no longer concerns AI law alone - today it intersects with consumer law and data protection law, and none of these areas permits the industry to treat them separately any longer. Article 50 of the AI Act is a good example of this: formally, it is a provision on technological transparency, but in practice it functions as a provision supporting the protection of minors.
Three things to check with a client today
If I were to sit down with a studio’s or publisher’s legal team this week, I would start the discussion with these three points.
Target audience and game mode. This determines whether a given AI-NPC meets the ‘obviousness’ threshold set out in Article 50, or whether it requires explicit disclosure. The same mechanic may be lawful in one game mode and unlawful in another - depending on who is playing and whether they might mistake the NPC for another player.
Shopping personalisation systems. It is worth checking which of these might border on the prohibited manipulation referred to in Article 5 - in practice, there are usually more of them than the product team originally anticipated.
Engagement mechanics. It is worth checking whether a young player - or their parent - has any real tools at their disposal to understand their own behaviour within the game, rather than just a ‘buy more’ button. It is worth doing this in advance, before the forthcoming Digital Fairness Act also sets out similar requirements.
These questions are a real starting point for an audit - and it is better to run one before the regulator’s first inspection than after it.