Every AI query passes through a system: a training corpus shapes a tokenizer’s vocabulary, that vocabulary is fixed inside a model, an API meters and prices access to it, a business layer wraps it in its own instructions, and only then does an interface let a user form their question. This thesis takes Donella Meadows’s framing of leverage points, which suggests places in a complex system where a small, well-chosen intervention produces disproportionate effects, as its lens, applied to one dysfunction in that stack: a tokenizer’s vocabulary, calibrated to whichever language dominates its training data, fragments non-English queries into more tokens than English ones for identical meaning. This token tax means two people asking the same health question pay a different price purely as a function of language. Every mitigation proposed so far, including expanding a tokenizer’s vocabulary, compressing an already-written prompt, and caching a repeated prefix, intervenes at the tokenizer or business layer; none touches the layer where a query is first formed. This thesis treats that interface layer as an unexamined leverage point and tests it directly, in the domain where the tax bites hardest: personal, multilingual consumer health questions.

The work proceeds in two movements. The first characterises the problem at its source: benchmarking real consumer health questions across English, Hindi, and Arabic, across three commercial models, confirms the tax is real, structural, and consistent in direction regardless of tokenizer, therefore a stable property of the model layer. The second tests whether the interface layer can counterweight it: MedQuery, a visual, non-verbal query interface built on a taxonomy of intents, open keywords, and ordinal parameters expressed through direct manipulation of a Kiviat diagram, replaces natural-language formulation with a compact, largely language-independent encoding, evaluated against free-text prompting across the same models and languages.

What this second study found is the clear demonstration of what a systems lens is for. The interface reduced token cost for some models, and reversed into a cost increase for another, therefore the largest effect observed anywhere in this analysis. Tracing that reversal upward through the stack shows it did not originate at the interface layer: it originated one layer above, in how differently each provider’s caching mechanism absorbed the structured system prompt. The interface behaved identically in every condition; what changed was a business-layer decision the designer cannot see or control without deliberately measuring for it. Response similarity between conditions stayed stable and moderate throughout, consistent with the interface reshaping emphasis rather than correctness; token savings, where they occurred, did not imply faster responses, suggesting that cost and speed are governed by separate mechanisms within the same stack.

The conclusion is therefore not simply that interface design is, or is not, an effective lever against the token tax. It is that leverage exercised at one layer of a multi-layer system cannot be evaluated, or trusted, from that layer alone—it must be verified against the layers it depends on, empirically and per provider, as standing practice rather than a one-time decision. That is both this thesis’s central finding and its methodological argument: the interface layer is real leverage, and knowing that required looking past the interface entirely.

Showcase:
https://youtu.be/0KlDJPVajj0 ... Read more

In the rapidly evolving healthcare landscape, integrating Artificial Intelligence (AI), particularly Large Language Models (LLMs), presents significant opportunities and complex challenges. This study examines the efficacy and implications of LLMs within healthcare systems, with a focus on enhancing user-centric evaluation methods for AI-generated healthcare advice. The research centres on three critical areas: Identifying the limitations of current evaluation methods, addressing the challenges related to the accuracy and reliability of LLM-generated advice, and proposing improvements to evaluation frameworks to enhance the practical application of these models in healthcare settings.

Our study utilizes a chatbot prototype trained specifically on healthcare datasets relevant to the Dutch context, exploring its application in real-world scenarios to validate and refine evaluation metrics. By involving healthcare professionals in interactions with the chatbot, we aim to ground our findings in practical, user-based experiences. The engagement with the prototype helps uncover vital insights into the AI’s performance, emphasizing the necessity for models that generate reliable and ethical responses and resonate with professional healthcare practices.

The proposed research contributes to the broader discourse on AI in healthcare by offering a novel framework for assessing AI-generated responses through a blend of empirical user studies and theoretical analysis. This framework aims to mitigate the subjective nature of current evaluations and provide a more robust, standardized approach to assessing the impact of AI technologies on healthcare outcomes. Through this research, we aim to forge a path toward more responsive, responsible, and user-centred AI tools in healthcare, ensuring that they align with both professional standards and patient needs.
... Read more