The Act of Creation: Learning by Writing

A personal method, built and tested again on a new subject

This is not a PhD study, and I am not trying to prove a theory. It is a personal method I built before several years, just after finishing my postgraduate studies, for learning by writing: put the question on paper first, say in advance what would count as evidence, and test each idea against it. I have used it for years. For this article I looked at what researchers say about learning and about writing, to see where my experience is consistent with the evidence and where the evidence qualifies it. I also tested the method again on a subject I did not know, devolution in the UK, and recorded each step.

The article follows that order: what learning is, what it means as mental activity, what writing adds, what the research on writing and learning shows, the chain I rely on through argument and critical thinking, and then my own test step by step, with a way for you to repeat it.

What is learning? Definitions in different contexts

There is no single definition. Most share a core idea: learning is a relatively lasting change in knowledge, understanding, skill, belief or behaviour that comes from experience, practice or reflection, and not from growing up or from a passing state. Beyond that core, each tradition puts learning in a different place.

• Behaviourist: a change in behaviour produced by association and reinforcement (Pavlov, Skinner). It can be observed, but it says little about understanding.
• Cognitive: a change in mental structures, in how information is processed, organised and stored. Piaget described people building and reorganising their own mental frameworks.
• Constructivist and sociocultural: meaning is built by the learner, and thinking develops through social interaction and language (Vygotsky).
• Social learning: people learn by watching and imitating others (Bandura).
• Experiential: knowledge is created by reflecting on experience (Kolb). In architecture, Schön described designers learning through a conversation with the situation.
• Situated learning: learning is taking part in a community of practice and becoming a practitioner (Lave and Wenger).
• Transformative learning: the learner’s frame of reference for interpreting experience changes (Mezirow).
• Organisations and policy: Argyris and Schön separate single-loop learning (fixing errors inside existing assumptions) from double-loop learning (questioning the assumptions). In political science, Hall describes learning as a change in policy tools, their settings, or the underlying paradigm.

For this article, the useful distinction is the last one. Adding facts is one kind of learning. Changing the framework, you use to think about the facts is another.

Learning as mental activity

In cognitive psychology, learning is something the learner does. The basic flow is:

• Attention: you select what to notice.
• Working memory: you hold and work on a small amount of material at once. Its limits are why overload blocks learning.
• Encoding and storage: you turn the material into a form that connects to what you already know.
• Retrieval: you bring it back and use it. Retrieving strengthens memory, which is why testing yourself teaches more than rereading.

Three further ideas matter here. Mayer describes learning as selecting relevant information, organising it into a structure, and integrating it with prior knowledge. Craik and Lockhart showed that thinking about meaning produces stronger memory than repeating words. Piaget separated assimilation (fitting new information into an existing structure) from accommodation (changing the structure because it does not fit). Above all this sits metacognition: planning, monitoring whether you understand, and correcting yourself.

Learning from writing: what a person does mentally

Writing is not one mental activity but several that alternate. Flower and Hayes described three that recur: planning (generating ideas, setting goals, organising), translating (turning ideas into sentences) and reviewing (reading back, evaluating, revising). A writer moves among them constantly. In terms of mental work, a writer:

1. retrieves what they know and selects what is relevant;
2. organises it, deciding what comes first, what depends on what, and what is a main point;
3. makes the idea explicit, because a feeling of understanding has to become a full sentence, and vague ideas fail at this point;
4. states the relations between ideas (because, therefore, however);
5. reads the text as a critical reader would, noticing gaps and unsupported claims;
6. restructures when the draft does not hold, which changes the framework and not just the wording;
7. offloads memory onto the page, so working memory is free for the next problem.

Bereiter and Scardamalia separated two modes. In knowledge-telling the writer puts down what they already know in the order it comes. In knowledge-transforming the writer works between two problems, what do I believe, and how do I say it so a reader will accept it, and each changes the other. Galbraith adds that some ideas form during writing itself. In one study of writing under different planning conditions, discovery was linked to spontaneous sentence production and to revising the text as a whole, and these two activities had opposite relationships with the quality of the finished text (Baaijen and Galbraith, 2018).

Two papers on writing and learning

Scientific writing as an epistemic tool (Serrano-Moreno and colleagues, 2025)

This paper from Latin American universities starts from the idea that writing transforms thought and creates knowledge. It analysed 35 texts by education students, 18 articles and 17 project or degree reports, using a matrix of five categories and twenty criteria. Essays were excluded. It found weaknesses in handling other authors’ voices, in stating a research problem, in building the state of the art, and in discussing results against theory. The authors say writing’s epistemic potential arises when it builds argumentative reasoning, and they link the failures to little access to models, little supervision, and writing left until the end with no time to reflect and revise. The study did not measure learning, and it reports no counts for its findings, so it shows where students struggle, not that writing taught them.

Writing versus studying (Penrose, 1989)

Penrose tested the popular assumption that writing is a way to learn. Forty college freshmen each wrote a report on one 1,200-word text and studied another for a test, thinking aloud as they worked. Studying gave better scores on simple recall, on applying the material, and overall, even though the writers spent about 35% more time (54 against 40 minutes). There were no differences on tests eight to nine weeks later. Process mattered: one student copied the source sentence by sentence and remembered its structure, while another wrote key words and her own reflections and could connect facts across the text but not recall its structure. Penrose concluded that writing may suit critically examining information better than gathering it, and that teachers should say which kind of learning a writing task is meant to produce. She also warned that, with 117 analyses, many of the process findings could be chance, and that the tests missed the reflection some writers did.

What kind of learning for graduate level, and what the evidence says?

Graduate learning is usually described as a move from acquiring knowledge to transforming it and taking part in a discipline. The evidence for particular methods is uneven. Most rigorous studies come from schools and undergraduates. Graduate-level evidence is smaller and mostly qualitative, and I found no graduate-level study comparing teaching methods head-to-head. What exists:

• Retrieval practice and spaced study rated high in Dunlosky and colleagues’ review of study techniques (2013). Summarising, highlighting and rereading rated low. The studies are mostly school and undergraduate.
• Problem-based learning: a meta-analysis of 43 studies in tertiary education found a robust positive effect on skills, slightly less knowledge gained but more of it remembered, and variation linked to students’ expertise (Dochy and colleagues, 2003). A synthesis of reviews found long-term retention and skills favoured PBL, while conventional teaching did better on short-term standardised tests (Strobel and van Barneveld, 2009). Much of this comes from medical and professional programmes.
• Doctoral writing groups, feedback and supervision: studies describe feedback on writing as the main teaching tool, the supervisor as central, and groups and retreats as valued. Aitchison (2009) notes there is little empirical research on how writing group pedagogy works, and her own evidence is self-report. These studies are real but small and qualitative.

So does learning by writing create learning, or only a limited effect?

The largest body of experimental evidence is a meta-analysis by Graham, Kiuhara and MacKay (2020). It pooled 56 experiments with 6,235 students in Grades 1 to 11, comparing writing about content with instruction that used less or no writing. The average effect was g = 0.30 (95% interval 0.20 to 0.41), positive in 82% of experiments, and similar across science, social studies and mathematics and across elementary, middle and high school. The variation between studies was high.

The details matter more than the average:

• Writing against no writing gave a smaller effect (0.25). More writing against less writing gave 0.49.
• Writing with metacognitive prompting gave 0.40, against 0.15 without it, but the authors’ test of that difference was borderline (p = .06).
• Analysis and interpretation tasks gave 0.36, against 0.18 for recording information. Activities that promoted evaluation gave 0.44, against 0.16 for knowledge and comprehension. Graphical representation did not show a reliable effect.
• True experiments averaged 0.21, below quasi-experiments (0.42), so 0.30 may be generous. There were no follow-up data on whether learning lasted, no Grade 12 studies, and no university or graduate studies.

My reading: writing produces a modest, real gain in learning content, and the gain depends on what the writing asks the writer to do. It is stronger when the task is analytical or reflective than when it only records information. None of this measures producing new knowledge, and none of it is at graduate level.

The link I rely on: writing, argument, critical thinking, knowledge

Since graduation my working explanation has been a chain: writing builds the ability to argue; the ability to argue is part of critical thinking; and critical thinking builds knowledge. For this article I asked an AI literature search tool, in a series of short statements, whether published studies point the same way.

It returned a table of about ten sources, later fifteen. On the first link they were mainly intervention studies of argumentative writing: writing-intensive instruction (Pappas and colleagues, 2019), argument-driven inquiry (Sampson and Walker, 2012), guided experimentation with a database (Radhakrishnan and colleagues, 2010, a quasi-experimental study), and students evaluating examples of writing (Braaksma, 2005), with Longo and Cisotto (2014) describing writing as a thinking tool. Later sources compared critical-thinking teaching with conventional teaching and looked at later argumentative writing (Li, 2012; Prayogi and colleagues, 2022; Birhan, 2023; Yamin, 2025). I have not read these studies myself. I am reporting how the tool summarised them.

Here is what I think that supports and what it does not:

• Writing builds argument capabilities: reasonably supported. In most of these studies the outcome is better argumentative writing, which is the skill that was practised. That shows practice helps. By itself it does not show that thinking has improved in general.
• Argument capabilities are part of critical thinking: argument is one part of critical thinking, so “part of” is accurate and “means” overstates it.
• Critical thinking builds knowledge: The studies show better arguing and better course performance. Research suggests that critical thinking possesses a generative capacity, enabling the creation of new knowledge or the testing and refinement of previously established theories (Facione, 1996), (Kirk et al., 2023).
• Future work: in these studies, it means later assignments or test scores, not performance in professional practice.

A study source is the meta-analysis by Abrami and colleagues (2015). It pooled 341 effect sizes from experimental and quasi-experimental studies that used standardised measures of critical thinking. The average effect was g = 0.30, with large variation between studies. Both generic and subject-specific teaching worked, at all educational levels. Three features helped: dialogue, authentic or situated problems, and mentoring. Combining them gave a medium effect (about 0.57), and teaching critical thinking inside a subject did better on average than teaching it as a generic skill. Graduate and adult students formed a small group of studies (about 17 effects, average near 0.21, from a summary table I have not checked against the original). So critical thinking can be taught, modestly, and the conditions that help are known. My method (though different in context because its graduate level) has all three: an authentic problem (the research plan), dialogue (critique by readers and research), and a mentor-like reader.

My test: a research plan on devolution, step by step

I wrote a research plan on the effect of devolution on urban development in the UK. When I wrote it, I had no knowledge about devolution. I wanted to see how a question and a plan would turn that into usable knowledge, and I searched online for the procedure to describe my thinking in the language of cognitive psychology. What follows is what I did and how the cognitive theory read it. The mental activity described is the theory’s interpretation.

Step 1: I asked “Why devolution?”

I asked what posing that question means as mental activity. The cognitive theory reading   described that as executive attention isolating the topic, working memory holding what is needed, and four processes: problem-solving (a gap between knowing that devolution happened and knowing why), reasoning (causal and conditional links), judgment (weighing causes) and decision-making (choosing the best explanation), with fast intuitive responses (System 1) and slow deliberate analysis (System 2).

Step 2: I specified I had no knowledge of the content

this reframed the question (step 1) as an ill-defined problem. The cognitive theory reading   shows Without facts, the mind builds a search strategy and generates hypotheses by analogy. It also described the question as preparing an empty, high-priority slot in working memory that new information could anchor to. That matches Ausubel’s advance organizer: a high-level structure that gives incoming detail somewhere to attach. The slot is a metaphor for that anchoring, not a measured mechanism.

Step 3: I wrote my own first theory

Before reading anything, I wrote that the countries had perhaps been united by force, that a central power blocked them from sharing decisions about their people and wealth, and that devolution came from that conflict. The cognitive theory reading classed this as a naive theory or intuitive schema, built top-down from a general power-and-conflict pattern and mapped by analogy onto the unknown event.

Step 4: I turned the theory into driver questions

I converted my ideas into a list of questions about drivers of change, in two groups: why the reforms were made (political power, economic failure, public pressure, international change, party conflict) and why implementation failed (misapplied legislation, administrative failures, lack of experience, decision problems, too many bodies). I planned to check each against evidence. The cognitive theory reading described this as an explicit hypothesis space with verification and falsification and set out how to test it: define what proof looks like before looking, gather chronological, intention and structural evidence, judge each driver for necessity, sufficiency and proximity, and counter confirmation bias by looking for evidence against.

Step 5: I tested two drivers

Party conflict. The criteria were set first. The cognitive theory reading then supplied the historical data: eighteen years of Conservative central government while Scotland and Wales voted mainly Labour and for regional parties, Labour’s 1997 manifesto, the referendums and the Acts. Verdict: necessary, not sufficient, both background and trigger.

Administrative failure. The criteria were a missing overall strategy, a civil service that did not adapt, and role confusion between London and the new bodies. The data pointed to a piecemeal approach, strain on the civil service and Whitehall slow to adapt. Verdict: necessary, not sufficient, and mostly a longer-term problem. The cognitive theory reading called both hypotheses “proven correct.”

Step 6: What I take from it

The question was the active ingredient. Without a question that said what I needed, reading would not have reached it, because the question told me what to select and what to ignore. Writing the question down made it explicit, and an explicit question can be tested.

I also learned something concrete. My first belief about why the Regional Development Agencies were abolished and replaced by Local Enterprise Partnerships was that the change was fiscal, a matter of austerity. That is the only explanation I had. I have not yet checked the stated reasons in the official documents, and there may be others. This is knowledge about a subject, new to me, but already in the public record. As key characteristics of new knowledge is Rigorous Methodology: It is not merely a personal opinion or a summary of existing reading; it must be generated through systematic, peer-reviewed, or scientifically valid research methods as the fiscal matter is pure peer- reviewed knowledge.

Conclusion

Writing does not automatically create learning, and it creates knowledge only when the claims survive checking. What it does is make thinking explicit enough to question. A question written before you search, criteria set before you look, and a critical reader afterwards turned a blank starting point into a research design and into knowledge about a subject I did not know. The evidence I found is consistent with this, and it also says the effects are modest and depend on the task, on dialogue, on real problems and on someone who reads critically. That is a modest claim, but it is one you can test yourself.

Sources I read in full or in part

Graham, S., Kiuhara, S. A., & MacKay, M. (2020). The effects of writing on learning in science, social studies, and mathematics: A meta-analysis. Review of Educational Research. doi:10.3102/0034654320914744

Penrose, A. M. (1989). Strategic differences in composing: Consequences for learning through writing (Technical Report No. 31). Center for the Study of Writing, University of California, Berkeley, and Carnegie Mellon University. ERIC ED310402

Serrano-Moreno, M. S., Oliveros-Rodríguez, E., & Moreno-Balza, E. A. (2025). Escritura científica en educación superior: proceso epistémico y argumentativo. magis, Revista Internacional de Investigación en Educación, 18, 1–27. doi:10.11144/Javeriana.m18.eces

Abrami, P. C., Bernard, R. M., Borokhovski, E., Waddington, D. I., Wade, C. A., & Persson, T. (2015). Strategies for teaching students to think critically: A meta-analysis. Review of Educational Research, 85(2), 275–314.

P. FACIONE, “Externalizing the Critical Thinking in Knowledge Development and Clinical Ju dgment”, [Online]. Available: https://www.academia.edu/download/109659771/Exernalizing_CT__Nsg_Otlk_1996.pdf

M. Kirk, R. Tytler, and P. White, “Critical thinking in primary science through a guided inquiry pedagogy: A semiotic perspective,” Teachers and Teaching, pp. 1–23, Mar. 2023, doi: 10.1080/13540602.2023.2191181.