Michael Rowe

February 2026

Essays

  • AI tutor accuracy in health professions education: The accuracy-engagement paradox12 February 2026A mathematical framework demonstrating that AI tutoring systems with 10–15% error rates can achieve superior learning outcomes through dramatically increased engagement compared to more accurate but largely unused alternatives. Drawing on evidence from health professions education, this essay shows that the multiplicative relationship between accuracy and utilisation creates an accessibility paradox: imperfect but engaging systems outperform perfect but unused ones. The argument carries three critical qualifications—errors vary in consequence and safety-critical content demands high accuracy; generative AI poses distinctive epistemic challenges that may undermine conventional error correction mechanisms; and engagement is necessary but not sufficient for learning, with superficial use patterns capable of nullifying predicted benefits entirely. A framework for calibrating accuracy requirements to context and consequence is proposed.
  • Documentation becomes infrastructure when AI agents are the readers9 February 2026When AI agents consume documentation as operational input, it undergoes a category shift from reference material to operational architecture; inaccuracies no longer merely inconvenience readers, they cause system failures. This essay argues that the primary bottleneck for institutional AI integration is not AI capability but information architecture: how institutional knowledge is structured, maintained, and made available to AI systems. Documentation written for human readers cannot function as reliable AI input without deliberate restructuring around explicit relationships and rigorous maintenance workflows. Treating this transition as a governance imperative, rather than a technical afterthought, determines whether AI integration delivers on its institutional promise.

Presentations

  • Beyond document management: Graph infrastructure for professional curricula10 February 2026Professional education curricula face a fundamental infrastructure problem: while comprehensively documented, they lack systematic queryability. This presentation introduces a three-layer architecture using graph databases as the source of truth for curriculum structure, supported by vector databases for content retrieval and the Model Context Protocol for stakeholder interfaces.

Posts

  • AI agents for academic workflow27 February 2026For the last few months, my screen has been split between Obsidian and a terminal, with two or three AI agents running in parallel tabs. This post describes what that shift in academic workflow looks like and what made it possible. The change is not simply additive: the work has shifted from execution to direction. What that distinction means in practice — and why it matters for those of us working in knowledge-intensive academic roles — is what I try to work out here.
  • Seven principle for collaborating with AI25 February 2026Seven principles for extended AI collaboration, distilled from a week-long project to restructure a large note collection using Claude Code. The principles cover goal-setting, understanding what AI can and cannot contribute, investing in planning conversations, adaptive planning, safety infrastructure, treating AI output as drafts, and expecting to learn something about your own thinking. Offered not as rules to follow but as patterns to recognise.
  • Building an AI-ready knowledge base24 February 2026A detailed account of a week-long project to restructure 5,819 Obsidian notes using AI as a working partner. The project involved building a 23-category taxonomy, migrating thousands of legacy notes to a consistent metadata structure, and generating AI-written descriptions for every note in the collection. The piece describes not just what was done, but how extended planning conversations, external project documentation, and careful human review at each phase made the work tractable. The most unexpected outcome was that building infrastructure for a note collection required articulating, for the first time, precisely how I think about my academic field.
  • Is a smaller AI model good enough?18 February 2026A field note on switching my default in Claude Code from Opus 4.6 to the smaller Sonnet 4.6, testing it on a task I'd normally give the larger model, and finding it good enough so far.
  • The hidden inefficiency in how we work with AI16 February 2026Most academics use one AI model for everything, but different models suit different kinds of cognitive work. Choosing the right AI model for each step of a task can improve both efficiency and output quality. Michael Rowe describes how he noticed this in his own workflow, what splitting one workflow across three models looks like, and why the same logic applies when a university plans its AI strategy.
  • Querying the Zotero database with AI15 February 2026What happens when you query Zotero with AI and treat your whole reference library as context, instead of searching it one document at a time? This field note describes a proof of concept in which Claude Code reads the Zotero database file directly. The approach works, and what breaks shows how much your metadata practices matter. It also sets out the privacy trade-off of sending query results to a cloud-based model.
  • Context engineering for educators: infrastructure, not just policy14 February 2026Most universities have responded to AI by rewriting assessment policies and running prompt-writing workshops. Context engineering asks for something different: decisions about system prompts, memory, retrieval, and connectivity that commit an institution to a direction for years. This post sets out what context engineering involves, why integrative, longitudinal, and values-based clinical education needs it, and why the gap between changing words and changing structures is where most institutions are stuck. It ends with three questions a programme team can ask about the AI tool it already uses.
  • The quality of the challenge: AI as a thinking partner13 February 2026Most discussions of AI in writing focus on output. This post describes using AI as a thinking partner while revising a theoretical framework paper for journal submission, asking it for critique, arguing back where it was wrong, and testing the draft against my own voice. The exchange changed the paper's structure and how it describes its methodology, and showed that the value lay in the quality of the challenge.
  • Organising your notes for AI: why retrieval isn't enough12 February 2026Most advice on organising notes for AI treats it as a retrieval problem, as if the AI only needs to find the right documents. The harder problem is translation, making the relationships in your thinking machine-readable without losing what makes them yours. This post explains why search and retrieval fall short, shows what a typed link between two notes looks like, and suggests one small way to start.
  • Building an AI workflow for academics with structured documentation11 February 2026Most advice on building an AI workflow for academics concentrates on writing better prompts. After several weeks of restructuring years of scholarly output with Claude Code, I've found the leverage lies in the structured context around the model — documentation that works as external memory, a cycle of testing rules and locking in the ones that hold, and the discipline of making implicit connections explicit. The same applies to institutions, where AI tools are only as reliable as the information architecture they're given.
  • Similarities between AI and human thinking: What if we're the language models?6 February 2026Discussion of AI and human cognition usually dwells on the differences. Turn the question around and the terminology of language models describes human thinking surprisingly well: context windows, biased training data, tokenisation, temperature, hallucination, system prompts, and pattern matching all have human counterparts, some of which health professions education already has names for. The similarities say something about our own cognitive architecture, and the strength of our resistance to them says something about what our professional identities depend on.
  • Hidden prompts in assignments: the security theatre of AI detection5 February 2026Some lecturers now hide prompts in assignment materials so that any AI system reading them adds tell-tale keywords to its output. This post looks at what those tripwires assume about assessment, and why what they detect is carelessness with AI, not the absence of learning. The technique also invites students to turn it on AI marking, and starts an escalation neither side can win. The alternative is to ask what assessment is for when producing the artefact takes seconds.
  • A bitter lesson for higher education3 February 2026Rich Sutton's "bitter lesson" from AI research is that general methods using computation beat decades of carefully encoded human expertise. Higher education faces a parallel lesson about AI and assessment validity. Essays, reflective accounts, and other artefacts worked as evidence of learning mainly because they were hard to produce, and detailed briefs and rubrics now read as ready-made prompts for a language model. Handwritten exams, proctoring, and AI use scales all try to make artefacts hard to produce again, when the question worth asking is how we recognise learning at all.

Notes

  • AI agents27 February 2026An AI agent is a system that autonomously executes multi-step tasks using language model reasoning — distinct from an AI assistant, which responds to individual prompts. Agents plan, act, observe results, and adapt, using tools such as file access, code execution, and web search. They perform best when given clear goals, explicit constraints, and well-prepared context.
  • Agentic workflows26 February 2026The human role in agentic workflows: agents handle execution while humans decide what should be made, define what good looks like, and judge the results.
  • Prompt injection17 February 2026Prompt injection is a technique in which instructions embedded in text cause an AI system to follow them as commands. In education it has appeared as a way of detecting AI use in assessment, which raises questions about authorisation and trust that the technique itself can't answer.
  • Inference16 February 2026The process by which a trained language model generates outputs; the computational work that happens each time you send a prompt and receive a response.
  • Token budget16 February 2026The idea that different kinds of cognitive work have different computational costs in large language models, and that matching task complexity to model capability matters for both efficiency and output quality.
  • Context rot14 February 2026Context rot, also called context drift, is the way a language model's output gets worse as its context fills with drafts, dead ends, and its own earlier mistakes. What causes it, why a larger context window does not fix it, and how to manage it in long pieces of work.
  • Claude Code12 February 2026Anthropic's AI agent that works in a folder of files on your own machine, reading, editing, and checking them under your direction instead of answering in a chat window.
  • Documentation debt12 February 2026The accumulated cost of outdated, ambiguous, or poorly structured institutional knowledge — long paid invisibly by readers who work around the gaps, and exposed when AI agents read the same documents literally.
  • Higher Education Reference Model (HERM)12 February 2026A shared vocabulary for universities, set out in business, data, application, and technology reference models, and a practical starting point for making institutional knowledge readable by software and AI agents.
  • LaTeX12 February 2026A typesetting system in which you write the structure of a document in plain text and the system lays it out — the standard for theses, mathematics and anything with hundreds of cross-references.
  • Markdown12 February 2026A way of writing formatted documents in plain text — a hash for a heading, asterisks for emphasis — so the source stays readable and converts to anything.
  • Open access licensing12 February 2026Open licences, mostly Creative Commons, state in advance what others may do with a work — read it, share it, adapt it, build on it — so nobody has to ask. The Creative Commons licence on teaching materials decides whether a colleague can adapt a resource for their own programme or has to rebuild it from scratch.
  • Open source software12 February 2026Software whose source code is published under a licence that lets anyone read, change, and redistribute it — and what that means for the tools an institution depends on.
  • Pandoc12 February 2026A command-line converter that reads a document's structure in one format and writes it out in another — markdown to Word, Word to HTML, anything to PDF — with citations intact.
  • Plain text12 February 2026Files made only of readable characters, with no formatting locked inside them. The format every tool can open, and the one that will still open in thirty years.
  • Programmatic assessment12 February 2026An assessment programme that gathers many low-stakes data points over time and makes high-stakes decisions on the aggregate — assessment designed as a system, with no single exam deciding a student's progress.
  • Embeddings10 February 2026An embedding is a list of numbers, learned during training, that represents a piece of text so that similar meanings get similar numbers. Embeddings let AI find material by meaning rather than by matching words, but they record that ideas are associated, not how they are related.
  • Graph database10 February 2026A graph database stores entities and the typed relationships between them, and is the storage layer behind a knowledge graph. For a programme team it can hold the curriculum map itself, so questions about how standards, outcomes, and assessments connect are answered by following the links instead of reading the handbooks.
  • Retrieval augmented generation10 February 2026A technique that improves a language model's answers by retrieving relevant passages from your own documents and adding them to the prompt, so the model answers from sources instead of from memory alone.
  • Single-hop reasoning10 February 2026Answering a question in one step, by retrieving the passages most similar to it and generating a response from them. It's the default pattern in RAG systems, reliable when the answer sits in one place and weak when it has to be assembled from several.
  • Vector database10 February 2026A database that stores embeddings, so that it can find content by meaning instead of by matching words. It's the storage layer behind most RAG systems.
  • Arms race dynamics in higher education5 February 2026How arms race dynamics in higher education create adversarial relationships between institutions and students, and what drives these cycles
  • Contextual interoperability5 February 2026Contextual interoperability is what lets you use your own notes, frameworks, and commitments with any AI model, local or hosted. It has two halves: the context has to be legible, with relationships a model can follow, and portable, held by you instead of inside one provider's system. With a curriculum mapping example from a pharmacy programme.
  • Intelligence as a service5 February 2026A way of using AI in which you draw on a model for a particular task while your notes, documents, and working knowledge stay private and under your control.
  • MCP server5 February 2026An MCP server is a small program that connects an AI assistant to one data source or service, such as a reference library, a folder of notes, or a calendar, through the Model Context Protocol. What the server offers sets the limit of what the assistant can see and do there.
  • Model Context Protocol5 February 2026The Model Context Protocol (MCP) is an open standard for connecting AI applications to external data and tools through one common interface, in place of a custom integration for every pairing. It makes connections technically cleaner; whether they are wise ones is still a decision for the people who set them up.
  • System prompt5 February 2026Instructions placed in front of every conversation with an AI model, setting how it behaves, what it knows, and what it won't do, usually without the user seeing them.
  • Large language models4 February 2026A plain explanation of large language models: what they are, how they are trained, where their limits come from, and what their ability to produce fluent text means for assessment in health professions education.

Frameworks

  • AI literacy development framework1 February 2026A framework for embedding AI literacy development into existing modules and courses, enabling students to develop AI capability while learning disciplinary content.