20 items with this tag.
PhD students are using AI across their doctoral work, but current policies focus on permission rather than specification. Drawing on how software engineers solved a structurally identical problem, this post introduces the research harness: a seven-part operating context that makes AI contributions visible, traceable, and supervisable. The harness addresses drift, offloaded thinking, and attribution failures — not by restricting AI use, but by specifying what the agent can and cannot do within the work.
Current responses to AI use in doctoral research locate the problem either with institutions through policy, or with students through judgement and AI literacy. This essay argues that neither response addresses the structural source of the difficulty. The characteristic problems of AI use in doctoral research are not problems of policy or capability; they are problems of working with an agent in the absence of a defined operating context. Drawing on software engineering practice, it develops the concept of the research harness: a structured specification, negotiated between researcher and supervisor, of what an AI agent is doing in a doctoral project. The harness has seven components — knowledge base, interpretive permissions, tools, authority, scope register, process record, and amendment protocol — and can be entered at a minimal level and developed iteratively alongside the work.
One in seven people in the UK are using AI chatbots for health advice instead of seeing a GP. The institutional response has been to warn them off, but that response applies a standard it doesn't consistently apply to anything else in the system. This post argues that the risk comparison driving those warnings is systematically skewed, and that a more honest accounting points toward an entirely different kind of response.
AI has disrupted doctoral education in two ways: the immediate question of AI-assisted writing, and the deeper question of what the PhD means when AI can conduct research from scratch. This post argues that the thesis was always a proxy for the person; evidence of an identity shift, not the thing being assessed in its own right.
Language models can transform documents into interactive tools in minutes. This post walks through a concrete example, turning a 21-page Word questionnaire into a working web app, and reflects on what that capability makes possible.
AI-generated text is fluent regardless of whether its content is accurate or well-reasoned. Fluency was once a reasonable trace of genuine thinking — a student who wrote clearly had usually thought clearly. That relationship no longer holds. Worse, the AI literacy response of teaching output evaluation is a temporary fix: as models improve, output quality converges on expert-level across every artefact we care to measure. The question isn't how to spot current failure modes. It's what you'll do when those failure modes are gone.
Claude produced the word "contribuves" in a piece of writing, which is obviously not a real word. This is a different kind of error than hallucination, and the distinction matters.
Every week I annotate articles in Zotero, highlights in Reader, and podcasts in Snipd, all of which is synced to Obsidian. By Friday I have a week's worth of material, tagged and structured, but unreviewed. This post describes the weekly review command I built to surface what matters and create a reason to engage with it.
The structural features of an information source that enable its knowledge claims to be challenged, traced back to evidence, and evaluated against the source's track record. Traditional sources carry it; generative AI largely does not.
A headless AI model runs non-interactively — no chat interface, no conversation. You pass it text, it returns output, and it exits. This makes AI tools composable with the same scripts and schedulers that have coordinated Unix processes for decades.
An 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.
A 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.
A field note on switching from Claude Opus 4.6 to Sonnet 4.6 as my default in Claude Code, and what I'm noticing after the first hour.
Source details Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021).
Prompt injection is a technique in which instructions embedded in text cause an AI system to follow them as commands. In educational contexts it has appeared as an AI detection mechanism in assessment — which raises sharper questions about authorisation and trust than it might initially seem.
The process by which a trained language model generates outputs; the computational work that happens each time you send a prompt and receive a response.
The 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.
Most academics treat AI models as interchangeable general-purpose tools. They aren't. Different models have different characteristics that make them better suited to particular kinds of cognitive work, and matching tasks to those characteristics may improve both efficiency and output quality. This post explores what that looks like in personal workflows and how the same logic scales to institutional AI strategy.
Learn about LLM context drift (or context rot) and how the reduction in output quality as tokens increase affects complex AI workflows.
A 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.