Michael Borck
I lead LocoLabo, an applied AI research initiative studying what local, privacy-first AI can do for teaching and learning — on modest consumer hardware, honestly benchmarked.
"We map the floor because most people live there, and nobody is documenting it honestly."
Investigates whether structured conversational prompts help students move beyond passive AI use toward active engagement. Tested with advanced frontier models, measuring depth of inquiry, iteration rates, and learning outcomes.
Applies the Design Science Research methodology to the development and evaluation of AI-powered workplace simulations for professional skills education. Reports design cycles, artefact evaluation, and generalised design principles.
Analyses interaction logs from AI-assisted learning environments to identify recurring patterns in student questioning behaviour, AI response quality, and correlation with learning outcomes.
Explores whether intentionally introducing friction into AI-assisted workflows — slower responses, constrained outputs, forced reflection steps — improves long-term retention and metacognitive development.
Effective use of large language models has accumulated as a catalogue of prompting techniques, documented but little explained. This paper argues the catalogue has a single underlying logic: the techniques that most reliably improve LLM output, from chain-of-thought prompting to spec-driven development, are transfers of the strategies people have long used to manage cognitive and organisational complexity. Similarity and deficit explanations are reconciled through affordances theory and a form/substance distinction. Prompting returns the strategy's form; the strategy is actualised only when the user supplies, from outside, the substance the form lacks, the grounding and judgment the model does not provide itself. The framework specifies when transfer holds, predicts where experts outperform novices, and recasts AI literacy as the discipline of supplying substance, above all independent verification.
Tests whether students nudged toward deeper engagement with weaker local models can match the learning outcomes of un-nudged users working with frontier systems. Examines the role of model capability as a confound.
Examines how model context window sizes impact performance, coherence, and throughput on consumer-grade GPU hardware. Establishes practical guidance for deploying local models in resource-constrained educational environments.
Explores the relationship between a model's token generation speed and users' perception of its intelligence and usefulness. Investigates whether slower models are systematically underestimated in educational contexts.
A two-study programme examining whether structured conversational nudges help students engage more deeply with AI — and whether that engagement transfers across model capability tiers.
A research programme on why human cognitive and organisational strategies for managing complexity also improve LLM performance. The first paper is a conceptual framework accepted at ACIS 2026. The programme extends into empirical, design-science, and pedagogy work.
Design science research programme applying rigorous artefact-evaluation cycles to AI-powered professional simulation tools used in workplace readiness education.
Large-scale analysis of interaction logs from AI-assisted learning environments, examining questioning patterns, response quality, and correlations with outcomes.
Investigates whether introducing intentional friction into AI-assisted workflows — slower responses, forced reflection, constrained outputs — improves long-term retention.
Systematic performance testing of local language models across consumer GPU memory tiers — context length effects, token rate, and perceived intelligence.