Too Much, Too Little, Never Just Right? The Labelling Dilemma of Article 50 of the EU AI Act
Imagine scrolling through TikTok and encountering a hyperreal “person” doing a goofy dance with uncanny precision—only to discover it’s not a Veo 3 creation but a flesh-and-blood human pretending to be one. A parody? A social commentary? Or just another viral challenge in the digital circus tent?
Welcome to the new frontier of AI-laced confusion, where even real humans now cosplay as AI outputs. And right in the middle of this shimmering hall of mirrors, we have Article 50 of the EU AI Act, waving its legal wand and requiring “prominent and clear” labeling of AI-generated content.
But what happens when the boundaries of reality and synthesis are so murky that not even the human participants know what label to slap on?
The Fine Line Between Overkill and Oversight
The intent behind Article 50 is clear: empower individuals to understand when they’re interacting with machine-made media. But the execution? Far less so. Article 50’s transparency requirements face substantial implementation challenges, particularly in relation to watermarking technology limitations and cross-media compliance difficulties. Transparency requirements could result in two problematic outcomes:
- Over-labelling: Every instance of generative software use (hello, spellcheck and photo filters) could technically qualify under some readings of the law. The danger? Warning fatigue. Research on content warning effectiveness demonstrates that AI content labels have significantly smaller effects than traditional warning labels, e.g. approximately 3x smaller than “False” content labels. Labels become the digital equivalent of elevator music: ubiquitous, ignored, and utterly pointless.
- Under-capture: Meanwhile, manipulative content like our TikTok AI-impersonators could slip through the cracks if regulation clings too closely to technical definitions. If no algorithm “generated” the person, does the performance still need a label, even if its intention is to mislead?
The risk? We end up regulating with a blindfold: catching everything and nothing all at once.
Sectoral Minefields and Creative Landmines
This isn’t a one-size-fits-all terrain. Consider the starkly different stakes across sectors. Analysis of 52 global news organisations found that 90% require AI usage disclosure, reflecting journalism’s commitment to transparency. But creative industries face different pressures entirely.
Research reveals that AI disclosure has no meaningful effect on creative or descriptive stories, but negatively affects evaluation of emotionally evocative first-person poems. The reactions are more negative when content is viewed as distinctly “human,” suggesting disclosure requirements may create unintended constraints on artistic expression.
Meanwhile, advertisers deploying hyper-personalised avatars, and everyday users adding a generative zing to their family vacation photo navigate different risk landscapes entirely. Label everything, and we delegitimise creative expression and overload consumers. Label nothing, and we legitimise deceit.
The real task is meaningful transparency—not legalistic tick-boxing. But achieving this requires understanding that transparency requirements may serve industry interests over human rights concerns, potentially reinforcing existing power relations rather than empowering users.
The Warning Fatigue Reality Check
Perhaps most concerning for policymakers is the mounting evidence about warning fatigue. A comprehensive meta-analysis of trigger warnings and content warnings found something striking: content warnings have no effect on emotional responses, do not improve educational outcomes, and reliably increase anticipatory anxiety.
If traditional content warnings don’t work as intended, what makes us think AI labels will fare better? Early research suggests they won’t. Small-scale studies found that while labels affected belief about AI generation, engagement behaviors (like, comment, share) were not significantly changed. People see the labels but continue behaving as before.
This points to a fundamental policy disconnect. We’re designing regulatory solutions based on the assumption that information equals empowerment, but the evidence suggests that’s not how human psychology works in practice.
Beyond Labels: Toward Operationalised Integrity
So where do we go from here? The research points toward several promising directions that move beyond the binary logic of “label or not label.”
Context-aware standards: Meaningful labeling should possibly consider purpose, risk, and audience. A deepfake political speech demands a different kind of flag than a generative filter in a selfie. Research on progressive disclosure shows users benefit from initially simplified feedback that adapts based on context and user needs, rather than universal one-size-fits-all approaches.
Platform-specific approaches: Meta’s transition from “Made with AI” to “AI Info” labels revealed that 82% of users in 13-country survey favor warning labels for AI content depicting false statements, but labels based on industry standard indicators didn’t align with user expectations for minimally edited content. The lesson? Perhaps generic technical standards do not map well onto user mental models.
Dynamic disclosure systems: Rather than relying on static text overlays, labeling could perhaps adapt to the platform and medium, be it watermarking, visual cues, or metadata traces. Research demonstrates that adaptive, personalised approaches show more promise than static warning overlays, particularly when they account for cross-cultural differences in AI terminology interpretation.
Media literacy investment: Let’s be honest about this: educational research suggests that AI literacy development may be more effective than labeling alone for helping users understand and evaluate AI-generated content. The four core competencies—understanding AI concepts, using AI tools, evaluating outputs, and addressing ethical issues—require investment in education rather than just regulatory mandate.
Guidelines over gavel: Hard law must be complemented with evolving best practices, co-developed with stakeholders (industry, users representatives, civil society, etc.) across sectors. The EC has begun this work, but this cannot be a theoretical rigid exercise: it needs flexibility and real-world test cases.
The Path Forward: Evidence-Based Pragmatism?
The evidence base is clear: current approaches to AI transparency are failing to achieve their stated goals. Warning labels don’t fundamentally change behavior. Universal labeling requirements ignore sectoral differences and cultural contexts. And technical watermarking solutions are still evolving.
This doesn’t mean abandoning transparency, quite the opposite. It means designing evidence-based systems that account for human psychology, technological limitations, and real-world implementation challenges. We need transparency approaches that are contextual rather than universal, adaptive rather than static, and educational rather than merely informational.
The stakes are too high for policy theater. In a world where humans cosplay as AI and AI mimics humans with increasing sophistication, regulatory solutions must be as nuanced as the problems they’re trying to solve. The alternative—a regulatory framework that satisfies neither transparency goals nor practical implementation needs—serves no one’s interests.
The question isn’t whether we need AI transparency. It’s whether we’re brave enough to design it based on evidence rather than assumptions. The research exists. The frameworks are emerging. What we need now is the policy courage to acknowledge them.
Written by Caroline De Cock, LL.M. , Head of Research.
