confirmation-draft-document

confirmation-requirements

Title page and abstract at the front, as you’d expect.

Introduction (a few pages). Not the thesis introduction, but a shorter version that orients the reader to the project, the research questions, the inversion from the original proposal, and the “After Intelligence” framing. This sets up everything that follows so the panel has context before they hit the denser material.

Contextual and practice review (the substantial chapter-length piece). This is the bulk of the document and where the real work is. Opens with the literature and practice positioning, moves into the methodological justification, then into the practice review documenting the artefacts and what they’ve surfaced so far. Written as a draft chapter that could later split into the methodology and process chapters of the thesis. This satisfies elements (i) from the requirements in one continuous piece of writing.

Chapter outlines and plan to completion The outlines we’ve written, preceded by a short paragraph explaining the thesis structure and the distributed autoethnographic principle. Then a realistic timeline to completion: what gets written when, what still needs building, submission target. This satisfies element (ii).

After Intelligence - building possibility spaces for learning and making with and through generative systems

Abstract (what I submitted for the doctoral symposium, definitely needs a bit more work)

==what why how?

===where exactly I want to make a contribution to knowledge?

===keywords and glossary add after abstract

What can AI add to the processes of learning and knowledge-making, while preserving the boundary around the messy, generative thinking so central to creative work? And what assumptions about intelligence, agency, and knowledge are embedded in the commercial AI systems now entering educational contexts?
The research proceeds through building: developing speculative tools, live systems, and research infrastructure as primary modes of investigation, where each artefact operates as both a technical object and a theoretical proposition.

PhD-Live sits at the centre of this work: a live, public-facing digital research environment that treats the process of doctoral thinking as a core output rather than just the means to an eventual end state. Around it, a wider infrastructure is taking shape, including a ‘supervisor bot’ whose development has surfaced a central design tension: how do you build an AI tool that introduces genuine epistemic friction rather than simply reflecting the researcher back at themselves? This question points toward a principle that emerged through making rather than planning: that LLM reasoning and human associative thought operate with different logics. LLMs are built to resolve and optimise; creative research thinking often needs to stay open, contradictory, unfinished. Designing with both means holding that difference deliberately, rather than letting the efficiency logic of the tool quietly shape the thinking.

The research sits at the intersection of speculative design, live coding practice, and critical AI studies. It is shaped by its moment: the politics of knowledge, precarity in higher education, and the institutional pressures that surround the integration of AI. Within this, several threads have emerged as central: what a locally-based AI infrastructure might offer creative research practice; the distinction between liveness and performance; and how generative AI can participate in creative thinking without speaking for the researcher.

===could I hyperlink to the phd-live site

Introduction:

I will orient the reader to the project, introduce the research questions and talk about the inversion from the original proposal’s approach, additionally explaining the ‘After Intelligence’ framing ===ground the explanation in how it has changed

Research Questions:

• What assumptions about intelligence and knowledge are embedded in commercial AI systems, and how do those assumptions compound when these tools are adopted into learning and educational institutions?
• How does working with LLM-based tools change the practice of thinking itself?
• How can experimental and speculative approaches to working with AI move beyond the generic workflows of commercial systems, making space for new kinds of learning and teaching?
• What might a commitment to liveness (keeping knowledge public whilst still forming) offer as a model for learning and knowledge-making in an era of AI-generated outputs? =wikipedia as a connection?=

Contextual and practice review:

===criticality needs to be made clear, ie. I am aware of the critique

The problem space

• Commercial AI systems are entering education rapidly, carrying embedded assumptions about what intelligence is, how learning works, and what efficiency looks like
• These tools are designed to resolve, optimise, and produce polished outputs quickly. That collapses the space between question and answer, process and product
• For creative practice and creative education specifically, that collapse is a problem. A lot of what matters in creative work happens in the unresolved, unfinished middle, before things make sense
• This research sits inside that tension. I occupy three roles simultaneously: learner (PhD student), researcher (building and investigating AI tools), and teacher (senior lecturer in creative computing, teaching with and about AI). The triple position is the research site

Speculation as method

• The research adopts a speculative stance as a critical method.
• Speculative design (Dunne and Raby) provides the foundational tradition: building artefacts that ask “what if?” as a way of making alternative possibilities thinkable
• The tradition is read here as politically situated: Benjamin’s “Imagination: A Manifesto” frames imagination as a political act, not a luxury. Who gets to imagine, and what gets imagined, are shaped by power
• The artefacts in this research are not prototypes or products. They are thought experiments made buildable: propositions about what AI-supported learning and knowledge-making could look like if the assumptions were different
• This led to the concept of possibility spaces, which the thesis takes up further: both as a speculative method and as something that meets real institutional constraints

Liveness and process

• The speculative approach led to utilising live coding as a practice to inform how I explore liveness as a way of keeping process visible and unfinished
• Live coding provides the methodological vocabulary:
  • Blackwell et al’s “Live Coding: A User’s Manual” establishes liveness as durational, embodied, and non-repeatable
  • Cocker’s “Performing thinking in action” frames live coding as thinking-through-doing
  • McLean et al’s “The Meaning of Live” raises the question of liveness without audience, which is directly relevant to PhD-Live
• The central tension: is PhD-Live genuinely performative, or is it live only in a temporal sense? This question now drives a research question of its own: what might a commitment to liveness, keeping knowledge public whilst still forming, offer as a model for learning and knowledge-making in an era of AI-generated outputs?
• Tools for thought and digital gardens provide the second lineage:
  • Caulfield’s “Garden and the Stream” as the founding distinction between exploratory networked knowledge and linear feeds
  • Appleton on digital garden history and ethos
  • Matuschak’s “How Might We Learn?” as the bridge between tools-for-thought and AI-and-learning
• Recent work on creative process traces (Kreminski and Mateas, Hammad et al) speaks directly to what PhD-Live does: making the traces of creative research visible and treating them as meaningful objects, not just documentation
• The argument: liveness and process visibility are not features of the platform. They are a central way of working. Commercial AI collapses process into product. Making process visible and keeping it unfinished is a deliberate counter

The thesis will also engage with Personal Knowledge Managment systems and “second brain” traditions (Ahrens and adjacent) but distinguishes its critical research register from their productivity register.

Intelligence, cognition, and the politics of AI

• The research doesn’t start from a history of AI. It starts from the question of what intelligence means and who gets to define it
• Crawford’s “Atlas of AI” provides the political economy of AI: how intelligence gets operationalised, who profits, what gets extracted
• Agüera y Arcas’s “What is Intelligence?” provides a contemporary reframing of the question relevant to current LLMs
• The chatbot lineage matters as critical precedent: Weizenbaum built ELIZA as a critique, not a product. He was alarmed by how readily people projected understanding onto a system that had none. The bots in this research are built in a similar spirit: to interrogate assumptions about intelligence rather than reproduce them
• Sycophancy in LLMs (Cheng et al, Malmqvist) names a structural tendency that connects directly to a finding in this research. When the supervisor bot is fed context drawn from my own notes, it tends to agree with positions I’ve already taken and reinforce framings I’ve already used, rather than push back or open new lines of thought. The literature suggests this isn’t incidental: it reflects how these models are trained and rewarded, which means designing for genuine challenge in a system that holds the researcher’s own material is harder than it looks

The thesis will engage further with histories of cognition and computation (Hayles, the broader critical AI field), but the key references for confirmation are above.

Learning, institutions, and knowledge

• The pedagogical framing is grounded in:
  • Manning on learning otherwise: learning exceeds and refuses the categories institutions impose on it, as such the application of AI in this context is not straightforward
  • Naidoo and Whitty on students as consumers: the commodification of learning under neoliberalism shapes what knowledge-making looks like and what gets valued
  • Matuschak’s “What’s worth learning if we have AGI?” poses a foundational question this research is exploring: when AI can do much of what learning used to develop, what is learning actually for?
• On creative knowledge specifically: a lot of what matters in creative practice can’t be fully put into words. Schon calls it knowing-in-action, Polanyi calls it tacit knowledge, Dreyfus argues expertise is embodied and intuitive rather than rule-based. LLMs only work with what’s been written down. So there is a disconnect here: the kinds of knowing creative work depends on most are exactly the kinds these systems are not able to perform well with
• The argument: learning and knowledge-making happen within institutional, political, and technical conditions. The autoethnographic position makes those conditions visible

Methods: building with and through

• The research is practice-based and autoethnographic, with building tools as the primary mode of investigation
• Methodological traditions drawn on:
  • Frayling’s “Research in Art and Design” as the foundational distinction between research into, through, and for art and design. This work sits in the “through” tradition: research conducted through making
  • Candy on practice-based research specifically, distinguishing it from practice-led: the artefacts themselves are the contribution, not just the means of generating findings
  • Skains on designing and conducting practice-based research in the creative arts
  • “The Auto-Ethnographic Turn in Design” on using the researcher’s own experience as research material
  • Gaver on research through design and the annotated portfolio method
• “Building with and through” LLM systems is the specific formulation: the LLMs are both the material (what the tools are made of) and the infrastructure (the substrate the research practice runs on). Each artefact operates as both a technical object and a theoretical proposition
• The stated “playful approach” is grounded in beliefs about how learning works, drawn from my teaching practice and experience. It connects to the Montessori tradition and to creative and art school approaches to teaching (studio pedagogy, crit culture, learning through making as a norm). The characteristics of play, experimental, imaginative, self-directed, learning through doing, describe both the method and central values about teaching, which is what connects the methodology to the pedagogical thread
• I will address Jowsey et al’s rejection of generative AI for reflexive qualitative research directly: this work uses AI within reflexive research, but AI is the object of inquiry, not a substitute for reflexive analysis.
• Pedagogy and the autoethnographic position are distributed across every chapter rather than isolated in one. This is a deliberate structural choice that the methodology chapter will name and justify

Practice review: the artefacts

Research Infrastructure
This is the overarching system that contains all computational systems and components in my research process. More information about these components below:

PhD-Live
A public website that publishes my working research notes as I make them, so anyone can see the thinking happening rather than just the finished writing.

• A public-facing digital research environment built with Obsidian, Eleventy, and Vercel
• Publishes the working vault (all of my research notes) as a digital garden, making the ongoing process of doctoral thinking visible as it happens rather than retrospectively
• PhD-Live is not documentation of the research. It is the research: a proposition that process visibility, maintained in public and in real time, constitutes a meaningful form of knowledge output
• The activity log (making AI interactions part of the public research record) is the most novel element of the architecture

Supervisor bot
A chatbot I built to act as a kind of additional supervisor: it listens to recordings of my real supervision meetings, transcribes them, and offers reflections and questions back to me afterwards.

• Started as a terminal tool using whisper.cpp for transcription and Ollama for local model inference
• Rebuilt as v1.5: a modular FastAPI application with server-sent event streaming, a browser-based interface, and an LLM-driven interjection gate
• Draws on the shared research context layer to ground responses in my own material
• Key finding: the echo chamber problem. Feeding the bot context drawn entirely from the research vault collapses its response space around my existing framing. This is a structural consequence of how LLM systems process correlated context, not a failure of implementation
• This points to a central design question: how to build an AI tool that introduces epistemic friction rather than fluency

Study companion
A chatbot I talk to when I’m thinking through ideas, working out half-formed thoughts, or processing the anxieties that come with doing PhD work, essentially a thinking partner.

• Built on the same infrastructure but designed as an everyday thinking partner: the bot I talk to when I’m working through ideas, iterating on half-formed thoughts, or processing the anxieties that come with PhD work. The intended role is what Claude has been doing for me, ported to local infrastructure I own and control
• Key finding: the capability gap between local models and frontier systems. The study companion is trying to do something specific and important, support the kind of open-ended dialogic thinking that creative research depends on, and that’s exactly where local models struggle most. The bot does not replicate what working with Claude makes possible
• The gap is treated as a research finding and as something the work actively pushes against. Closing it through pure model capability is constrained by what’s available in open-source models and what local compute supports, but there’s real work to be done at adjacent layers: prompting, context shaping, scaffolding, fine-tuning where appropriate, and architectural decisions about how the system uses the model.

Shared research context layer
A searchable database of all my research notes that the bots draw on, so they can respond to me with awareness of what I’ve already written and thought about.

• A ChromaDB vector store with nomic-embed-text embeddings, watching the Obsidian vault and serving a /context endpoint consumed by all bots
• A deliberate choice to build a shared knowledge substrate rather than treating each tool as a standalone application
• The architectural decision to use an event renderer paradigm rather than a conventional chat interface is itself a research position: AI interaction framed as a sequence of discrete events with visible structure rather than as simulated conversation

What remains unbuilt
A planned bot that would help me share what I’m working on more publicly: monitoring the research as it develops and drafting things like social media posts for me to review and approve.

• The confidence bot: an agentic tool that would monitor the research context and propose public-facing outputs for human (my) approval
• This may be the most research-generative component not yet built
• Likely to surface questions about agency, authorship, and voice
• Queued as priority for the next phase

emerging findings ===analysis of findings

• The echo chamber problem: correlated context collapses response space. A finding about LLM context handling and about the difficulty of designing for epistemic friction in tools that hold deep researcher context
• The capability gap: local models do not match frontier systems for open-ended dialogic thinking. A tension between sovereignty (owning infrastructure) and capability (needing what your infrastructure can’t provide)
• Liveness vs performance: an open conceptual question, now driving a dedicated research question
• LLM reasoning vs human associative thought: LLMs are built to resolve and optimise; creative research thinking often needs to stay open, contradictory, unfinished. Designing with both means holding that difference deliberately

Chapter outlines:

Every chapter (other than the introduction and conclusion) will have two movements. The first is the conceptual argument with pedagogy and learning as the orienting context. The second is an autoethnographic account where the pedagogical and learning questions deepen through reflection on my own practice.

1: introduction

• Frames the four research questions
• Unpacks “After Intelligence”: the critical interrogation of intelligence as a concept and engagement with AI beyond the doom and hype rhetoric
• Introduces PhD-Live and the wider infrastructure briefly
• Sets out the thesis structure and explains why pedagogy and autoethnographic reflection are distributed across every chapter

2: methodology, a playful methodology

Conceptual argument: practice-based and autoethnographic research, with speculative design and live coding as the methodological lineages. Possibility spaces as a foundational speculative approach. “Building with and through” LLM systems as the primary mode of investigation. Liveness named here as a methodological commitment, investigated fully in chapter 5.

Pedagogical grounding: “Playful” grounded in my own teaching practice and experience. Connects to the Montessori tradition and to creative and art school approaches (studio pedagogy, crit culture etc).

Autoethnographic account: The triple role of learner, researcher, teacher and introducing my positionality.

3: intelligence, power and ethics

Conceptual argument: How intelligence has been defined and measured, and how that history shaped AI development. The Weizenbaum/ELIZA parallel: building bots as critique rather than product. Political economy of commercial AI and the power dynamics of institutional adoption. The bots introduced here as critical responses.

Pedagogical grounding: When intelligence is defined as measurable and extractable, AI tools built on that logic squeeze out the kinds of learning that resist measurement.

Autoethnographic account: Building tools that resist resolve-and-optimise logic. The echo chamber problem as a consequence of how LLM “intelligence” works. Navigating institutional AI adoption as both staff and student.

4: process, digital gardens, tools for thought, and knowledge infrastructure

Conceptual argument: Digital gardens, Personal Knowledge Management, and tools for thought as a lineage the local research infrastructure sits within and departs from: constructed instruments for thinking, not productivity tools. The infrastructure documented: supervisor bot, study companion, shared context layer. Architectural decisions as research positions (local/open-source-first, shared substrate, event renderer rather than chat).

Pedagogical grounding: Tools for thought are claims about how learning happens. What changes when the learner is also the builder?

Autoethnographic account: What building and using the infrastructure has surfaced. The echo chamber finding. The capability gap as a sovereignty/quality tradeoff.

5: liveness and performance

Conceptual argument: Driven by the research question about keeping knowledge public whilst still forming. PhD-Live as central object. Live coding as the methodological tradition, and the liveness/performance distinction as the central tension: is PhD-Live performative, or live only in a temporal sense? Creative process traces as a bridge between liveness and documentation. Liveness as a critical response to AI collapsing process into product.

Pedagogical grounding: Learning in public as a pedagogical proposition. Connects back to art school pedagogy: showing work in progress, crit culture, process as visible and discussable.

Autoethnographic account: What it’s actually like to maintain a live public research environment. The unresolved question of whether PhD-Live is performance or practice.

6: knowledge, power and ethics

Conceptual argument: From how intelligence is defined to what counts as knowledge and who gets to produce it. Tacit knowledge traditions (Schon, Polanyi, Dreyfus) and the gap between creative knowing and the explicit, extractable model LLMs embody. The politics of knowledge in higher education: what gets credentialled, what’s deemed legitimate.

Pedagogical grounding: Knowledge-making as a political act. The critical pedagogy tradition (Freire, hooks) acknowledged here as part of what this argument extends, but the foundation is lived experience inside the institution.

Autoethnographic account: Making knowledge in public via PhD-Live while navigating institutional structures that assess and credential it. The tension between unfinished thinking and the institutional demand for legibility.

7: possibility or (im)possibility spaces

Conceptual argument: A reflective, evaluative chapter where the speculative framing meets what actually happened. The tension between what can be imagined and what can actually be built given resources, time, and institutional pressure. The chapter situates these constraints in their wider context: the growing financial pressures on Higher Education institutions in the UK, and the political economy of commercial AI that determines what tools are available, sustainable, and accountable. The decision to build locally is both political and practical, a response to these conditions rather than a neutral technical choice.

Pedagogical grounding: What kinds of learning become possible or impossible depending on infrastructure and who controls the tools?

Autoethnographic account: Reflect on the real constraints that shaped the work: limits of time and capacity as both a lecturer and a PhD student, what local hardware can and can’t do, the institutional pressures of occupying both roles at once. Some of the most revealing moments in the research have been the impossibilities. The distance between what was imagined and what was actually possible is itself a finding.

8: conclusion

• Returns to the research questions and synthesises contributions across chapters
• Reflects on what “After Intelligence” means at the end of the project
• Limitations and future directions.

Plan to completion ===make clear with the build components and focus groups

timeline
Year 3: October 2026 – September 2027

Confirmation submitted in October 2026. First workshops and focus groups with students. First substantial chapter drafts begin in summer 2027, starting with chapters closest to existing material (methodology and process).

Year 4: October 2027 – September 2028

Main writing period. Multiple chapter drafts across the year, with summer as the highest-output period. Continue infrastructure work and workshops/working with students alongside writing.

Year 5: October 2028 – September 2029

Complete remaining chapters. Reflective chapters benefit from being written last when the work is further along. Draft introduction and conclusion. Full draft review with supervisors. Revisions and final preparation for submission. Target submission late 2029.

Buffer: October 2029 – 2030

One year of contingency between funded submission target and realistic worst case.