AI systems can now tailor and refine persuasive messages based on what influences us effectively. In my talk, I go beyond the familiar concern that they may change individual beliefs, and ask whether persistent, optimized influence could affect how we form and assess beliefs over time. In particular, I ask what happens when AI increasingly selects the ways it communicates for their behavioural effectiveness, rather than for their capacity to support users' engagement with reasons. By drawing on a particular account of manipulation, I outline how this could gradually weaken opportunities to practise independent judgement and identify the empirical questions needed to assess that risk. My aim is constructive as well as critical: I suggest that AI should not merely personalize persuasion, but help people reflect on, question, and respond to reasons for themselves.

Automated detection of dark patterns has become a prominent research agenda, and multimodal large language models (MM-LLMs) now promise to support enforcement at scale. This talk starts from our own contribution to that agenda. DeceptiLens (Kocyigit et al., 2025) combines retrieval-augmented generation and chain-of-thought prompting to assess deceptive design patterns in interface screenshots and to explain its decisions, and experts evaluated those explanations for clarity, correctness, completeness, and verifiability. Although promising, the approach assessed isolated elements on single interfaces, and its usefulness for enforcement actors remained assumed.

A subsequent interview study with regulatory practitioners (Rossi & Parkin, 2026) shows that, despite a pressing need for automation, most academic tools can at best flag suspicious interface elements. Enforcement, by contrast, requires evidence of specific legal infringements, gathered through processes that are traceable, auditable, reliable, subject to human oversight, and preferably open source. Many research tools, ours included, do not meet these requirements.

I argue that this mismatch is not merely a matter of engineering. As of last year, several tools we examined recognized interface elements and concluded that they were dark patterns, thereby misrepresenting the concept, as if any highlighted option was a digital nudge and every nudge a dark pattern. When a simplified version of dark pattern doctrine reaches ML engineers, and regulators who lack design and behavioural science expertise, form-based definitions such as "interface interference" or "non-neutral presentation" risk mislabeling designs that enhance privacy or security, some of which EU law requires, as with data protection by default. The prohibition of manipulative AI techniques in Article 5(1)(a) of the AI Act illustrates the point, since material distortion and significant harm cannot be read off an interface. Interface-only detection is therefore of limited reliability. The talk concludes by calling for a more nuanced, interdisciplinary approach, both in HCI research, from which ML detection typically derives its ground truth, and in legal scholarship.

Being manipulated by machine learning is bad, the law should do something! But wait: maybe the law is Teenagersdoing something about it? There is currently a lot of attention on addiction and manipulation online, with news reporting haunting stories of humans that divorce to marry chatbot companions, teenagers losing their lives, and lawsuits against social media addiction. The architecture of a digital environment and its interface design are powerful tools to guide and influence user behavior. Both experts and the European legislator have highlighted the possible effects of combining persuasive design (so-called dark patterns, but not only) with profiling based on personal data from recommender systems and other algorithms. This talk identifies the components of online persuasion that are relevant for the law, and the different regulatory approaches proposed by the European legislator, emphasizing the tricky legislative puzzle created by the interaction of the main European rules on data, platforms, and AI.