Writing with probabilistic machines

Introduction

I come from an engineering and research background where structure, precision, and clarity matter above all. However, I have always been attracted to great storytellers (e.g., novelists, history popularizers, comedians, science communicators). There is something almost magical about their ability to weave ideas into compelling narratives. My own writing has always been more pragmatic: detailed specifications, internal notes synthesizing multiple perspectives, documents outlining a vision, research papers sharing results of experiments, … Over the last twenty years, I have learned that the person who writes things down holds a particular kind of power: the power to shape how ideas spread and take root both in my own mind and in others when I share them.

For me, writing has never been comfortable because I write to think. Writing forces my mind to meet reality: to capture multiple points of view, to form an opinion, to articulate a perspective. Without writing, my ideas remain vague, badly formed and poorly sustained.

I often share short texts with a colleague for a quick review, and sometimes I link them to publish an essay, structure a class, or deliver a public presentation. However, most of the time, I accumulate stacks of half-baked drafts that serve as notes for ongoing thinking. Until recently, this was mostly a solitary effort. It still is at its core. The thinking, the choosing, and the crafting of ideas remain my own. But the act of writing, the articulation itself, has changed.

Writing becomes conversational

With the advent of Generative AI tools, some parts of my writing process now occur through a messaging interface (Rockmore, 2025). Some refer to it as “vibe writing.” Suddenly, I could engage in a dialogue about my ideas, test different ways to articulate them, and get help when I felt stuck. For someone who finds storytelling challenging, this felt revolutionary. I had a tool to help convert a chaotic train of thought in my mind into smooth, readable texts, almost like a storyteller.

I rely on popular general-purpose LLMs (e.g. ChatGPT and Claude) to refine the flow of my thoughts and on AI wrappers (e.g. Perplexity) to find references and new research material. I use these tools to organize my notes, define the boundaries of my text, and better position my ideas. They are especially helpful in preventing me from “Magnum Opusing”: that common trap where the scope expands endlessly, and notes build up until the project becomes too complex to ever complete.

If I were to illustrate my writing process for this text, it would resemble a wandering line gradually gaining definition. The blue line shows how the text evolved as a living document through successive revisions, transforming from vague fragments into clearer, more developed thoughts. The light grey lines represent notes from observations, readings, and conversations that intersect with AI-assisted articulations in pink. Together, they create an exploratory path that splits into multiple directions, loops back, and ultimately converges. What starts with a narrow focus broadens into not just a finished piece but a wider landscape of thinking that only emerges through the act of writing.

This diagram mirrors my other creative work in software engineering and futures design, in which Generative AI tools have amplified my capabilities. I have been vibe-coding ideas and prototyping concepts without deep knowledge of the latest software libraries or fluency in certain programming languages.

Having spent years creating software and envisioning futures, I have learned to watch carefully for what we might lose in what we gain. In 1964, Marshall McLuhan argued in Understanding Media, that technology and society coevolve: every augmentation is also an amputation (McLuhan, 1964).

The hidden costs of writing with AI

LLMs have undoubtedly reduced the barriers to writing. Many people can now express their thoughts more confidently, even when writing in a foreign language (just like I am doing now). Something that would have been unimaginable just a few years ago. We worry less about grammar and form as these seem “easily” fixable by an AI writing tool. Or is it an illusion of confidence?

Loss of authenticity

My experience of outsourcing writing often results in safe, predictable prose instead of pushing the limits of my original thought. If I am not careful, my ideas fade and become average. They are stripped of personal nuances, they lose their wabi-sabi and soon are no longer authentic. When approached superficially, writing with AI can easily reduce thought to a statistical middle ground.

Flawed reasoning

Generative AI tools rush towards conclusions and good-enough solutions, sometimes introducing illogical shortcuts. They produce plausible, compelling, confident prose that can camouflage flawed reasoning. These biases are features of the stochastic probabilistic models used in Generative AI. They are not a flaw that will go away soon. Consequently, I find myself engaged in the constant effort to reappropriate the ideas, to make them my own again.

Shallow thinking

Finally, thinking is not linear as a conversation via a messaging interface. It is often erratic and messy. To develop stronger ideas, exposure to different perspectives and the discipline of working through them is necessary. Ideas need time to mature. The quick-fix nature of current Generative AI tools is undeniably useful, yet it also mirrors today’s shortcut culture, tempting us to avoid the difficult intellectual work required for genuinely original and deep thoughts.

If I were to illustrate my experience of AI-only writing, the diagram would tell this story: the trajectory quickly moves towards an apparent clarity and an “illusion of maturity,” as AI tools help me smooth sentences, tighten structure and define boundaries. However, the scope of the writing corpus remains relatively limited. The final text appears complete and mature, but it reflects what I lose: the deeper, more erratic thinking that arises from grappling with ideas outside the context of interacting with a probabilistic machine.

Beyond my work, I have observed the “illusion of maturity” in some students’ final projects and in the AI-generated content shared on LinkedIn, Medium, and blogs. Recent studies seem to support these observations (Kosmyna et al., 2025), although the research is still in its early stages and the findings should be approached with caution. The underlying cause seems to be our natural human preference for fast thinking, which is easier and requires fewer resources (Kahneman, 2011).

This raises a crucial design challenge: how do we maintain the slow, arduous work of thinking alive when Generative AI tools make writing so swift and effortless?

Writing is social

To write this text, I wanted to counterbalance my use of Generative AI tools with more human interactions. I conducted a small experiment in the form of a tertulia. A concept originally from Spain, a tertulia is a regular social gathering where people share their recent creations and discuss current affairs. These gatherings, also known as cénacle in France or salon in the English-speaking world, have long served as spaces where ideas flourish through human connection.

Practically, once a week, I gathered 4-6 colleagues (known as tertulianos) for a 1-hour online discussion. Each of us brought something in progress, such as a draft, a project, an outline for a presentation, readings, etc. We all regularly use AI tools, but the tertulia became a space to reflect and allow our ideas mature outside the rush of work.

The sessions felt like having a “writing circle”, which only popular storytellers or comedians have the luxury to enjoy. We bounced ideas around, challenged each other, and offered perspectives none of us would have reached on our own or with a tool. My role was to encourage these frictions, provoke collisions, and keep the conversation challenging enough to spark new thinking. It was erratic. It was fun.

After each session, I found myself immersed in notes from our conversations, along with related observations and readings. Each session challenged and deepened my thinking about “writing to think.” I was struggling to write this very text, and that struggle felt right. It pushed my ideas beyond the “illusion of maturity” in ways that happened entirely outside of AI tools.

The core of the writing process, the unexpected insights, the sensation of getting lost, obscure cultural references, inspiring analogies, and genuine leaps of imagination arose from social interaction. Laurent introduced the ongoing challenge of reappropriating ideas from AI tools. Andrés introduced the concept of “shortcut culture” developed by Carolina Sanín (Sanín, 2024). Lisa made connections to the book Thinking, Fast and Slow. My notes are filled with these breakthroughs, each sparking new connections in my mind.

A space to think

This text is a direct result of the practice it describes: engaging in conversations with machines to gain focus and clarity, and with humans to achieve maturity and depth. The goal is not to replace the struggle of writing, but to make it more fertile. It raises a larger question about how we collectively manage our relationship with AI tools.

There is a quote that captures this current coevolution between machines and humans: “The rapid spread of AI adoption is made possible through human collaboration”. It comes from my friend and accomplice Lisa Gansky, a member of the tertulia. Serial entrepreneur and author of The Mesh (Gansky, 2010), Lisa knows what she is talking about. She is an expert on technology, collaboration and networks. Together, we share a concern: when speed replaces depth, something is lost in what makes us human, both as professionals and as citizens.

Deep and authentic thinking requires time, curiosity, vulnerability, and a willingness to sit with questions. As Lisa often says, it is not a “spectator sport.” Thinking involves active making (e.g., writing, sketching, prototyping, …) to develop and practice its core skills.

The tertulia experiment shows that anyone could benefit from it. But today’s frenetic AI world needs more than occasional experiments. It needs a community of practice for this kind of slow and deliberate thinking. This is exactly what Lisa and I are pursuing: a space where a community of perpetual learners from diverse backgrounds immerses and engages with each other through tertulias and other collaborative explorations (e.g., hands-on studios, guest talks, etc.). We see this type of trusted, diverse learning lab as a foundational element for our lives.

References

GANSKY, Lisa (2010). The Mesh: Why the Future of Business is Sharing. Portfolio/Penguin.

KAHNEMAN, Daniel (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.

KOSMYNA, Nataliya; HAUPTMANN, Eugene; YUAN, Ye Tong; SITU, Jessica; LIAO, Xian-Hao; BERESNITZKY, Ashly Vivian; BRAUNSTEIN, Iris; MAES, Pattie (2025). “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task”. arXiv preprint arXiv:2506.08872 [online]. Available at: https://arxiv.org/abs/2506.08872

McLUHAN, Marshal (1964). Understanding Media: The Extensions of Man. McGraw-Hill.

ROCKMORE, Dan (2025, August). “What It’s Like to Brainstorm with a Bot”. The New Yorker [online]. Available at: https://www.newyorker.com/culture/the-weekend-essay/what-its-like-to-brainstorm-with-a-bot

Sanín, C. (2024). “Inteligencia artificial | CAMBIO ”. YouTube [online]. Available at: https://www.youtube.com/watch?v=WmhsX3qtiOs

Recommended citation: GIRARDIN, Fabien. Writing with probabilistic machines. Mosaic [online], January 2026, no. 206. ISSN: 1696-3296. DOI: https://doi.org/10.7238/m.n206.2517