Visible Thought: Conversational Discipline in Practice
I used to think I’d be a writer; writing was the first talent I exhibited from a young age. Then I built a career in technology. Go figure. Still, there’s something romantic about writing—about the creativity of turning thought into words. Writing is the intersection of thought, intuition, and the ability to convey an idea in a way that will resonate with an audience. It’s the ability to take a complex idea or story and make it. . .sing, if you’ll pardon the misaligned metaphor. Writers talk about their process, and they acknowledge that processes can be messy, frustrating, iterative, recursive, and ultimately uncomfortable. And yet, when they achieve flow—when the words just “pour” out of them — they experience focus and a feeling of euphoria. I’ve developed and honed my own process, and the pure joy that comes from translating an idea, making a point, or sharing a story eloquently is real. After reading this romanticized rant about writing, you might think I’d fall into the camp that all AI-written products are slop. You would be wrong. While frontier models, unchecked, produce prose that’s mediocre, the outcome can change entirely if we bring intention and discipline to the process.
Three Problems – One Solution
Problem 1: There’s been a lot of talk about how AI slop is “everywhere.” Here, I define “slop” as unoriginal, shallow or ill-conceived thought packaged in polished prose. In 1909, Ambrose Bierce in Write It Right said, “. . .good writing . . . is clear thinking made visible.” More than a century later, that remains the ideal. Frontier models, however, effectively create the ability to disconnect polished writing from human critical thought. That’s what yields the “slop.” 2025 research by the MIT Media Lab confirms the cost of passive reliance on LLMs: measurably lower cognitive engagement (a 55% reduction in functional neural connectivity) and a weaker sense of ownership of the product (83% of participants couldn’t quote from the text they just produced). No thinking, no visibility. Just polished prose and homogeneous patterns.
Problem 2: Critics argue that AI house style writing all sounds the same, and they're not wrong. And yet, academic papers all sound the same, too. So does legal writing, medical literature, and financial reporting. We engineer those genres to have a consistent rhythm and tone because sameness signals credibility and reduces cognitive load. The issue with AI house style is that it’s ubiquitous; it appears across all genres, even those requiring a distinctive human voice and style.
Problem 3: This pervasive uniformity has created a massive cultural backlash. As Te-Ping Chen of the Wall Street Journal reported in May, the stigma of generative AI writing is now so great that writers who leverage frontier models are twisting themselves into pretzels to “sound human.” These tactics, such as forcing casual slang or summarily rejecting established literary devices, completely undermine good writing.
All three problems are solvable, and the solution is entirely within human control: conversational discipline. Borrowed from theology, political science, philosophy, and organizational studies, the phrase traditionally means protecting the clarity and integrity of human exchange. In my previous paper, Beyond the Prompt: Learning to Lead the Machine, I modified the definition for the AI era: the rigorous application of linguistic precision, logic, and meta-analysis to an interaction with a frontier model to create an authentic, objective output.
Let’s make it concrete.
The Human Baseline for Proficient Writing
Skilled writers have always leveraged expert collaborators. Ghostwriters, editors, and trusted readers have shaped some of the most celebrated writing in history. The individual brings the ideas, the point of view, and the governance; the expert brings craft, structure, and an honest critical eye. The goal of using frontier models to assist with writing is the same: to employ them like an expert. Like any expert collaboration, it only works if the human brings “something” to it.
That “something” is critical thinking, discipline, and craft. Humans have always produced junk; Bierce was writing about it over a century ago. Technology has simply allowed that junk to be on display and to sound authoritative on a massive scale. The hard question is what percentage of writers bring those three elements to the page. Because most research focuses on literacy instead of writing proficiency, a clean, global metric does not exist. I hypothesize that frontier models amplify whatever the human brings. Bring all three, and the work gets better. Bring none of them, and it’s a faster way to produce slop.
The Writing Process
Using frontier models for writing is a thinking, brainstorming, and collaborating exercise. It’s not just a content-generation exercise. That’s where conversational discipline becomes critical. An April 2026 NYU study supports this approach. Researchers studied over 400 professional writers and found that the strongest professional outcomes come from combining a collaborative orientation and a rivalry orientation to LLMs simultaneously. While pure collaboration yields short-term efficiency at the cost of human skill decay, treating the model simultaneously as a rival introduces a critical, self-preserving friction. This dual orientation forces the writer to actively maintain their skills and challenge the machine.
The best way to make the concept of conversational discipline concrete is to bring you into my writer’s room. A side-by-side comparison of how I write analytically (i.e., academic papers, thought pieces, research-grounded arguments), with and without a frontier model as collaborator, follows. While both processes have the same intent and require the same rigor, there are key differences. Those differences are where conversational discipline becomes visible.
Phase 1: Brainstorming
Starting a new project is one of the most exciting and uncomfortable experiences in writing — that stage when the ideas are swirling, and nothing has taken shape yet. For me, having a capable collaborator and sounding board available at any hour changes that experience. Engaging with a frontier model means that I don't have to navigate my chaotic mental space alone. Unlike a human collaborator, who knows what they know, a frontier model can connect ideas across fields in ways no single expert can. That changes what's possible in brainstorming before a single word of the actual piece is written.
| Human Process | Frontier-Model-Assisted Process |
|---|---|
| I start out with a ton of ideas and very little structure. So, I make some notes. I pose a lot of questions. I make some hypotheses. I evolve the topics and outline. I catalog what I know, what I don’t know, and where I need to go to find more information. | I start from the same place. I make some notes. I pose a lot of questions. I make some hypotheses. Then I start talking to the model(s). The quality of that collaboration depends entirely on the quality of my questions; a lazy question produces a lazy answer. I use targeted questions, exploratory questions, and critical debate to stress-test my early thinking before it hardens into assumptions. I evolve the topics and outline. I catalog what I know, what I don’t know, and where I need to go to find more information. |
Brainstorming is about asking unfiltered questions and exposing radical ideas. I love this phase. The ideas multiply, and the connections surprise me. In the human process, the transition to investigation happens naturally. I reach the edge of what I can generate alone and my questions start focusing outward. The model has no such natural edge. In the frontier-model process, the model could keep generating indefinitely, but eventually I notice that we’re covering the same ground. That's either a signal that the brainstorming phase is done, or a signal that the context has gotten muddled and needs a reset. Knowing which is which is part of the discipline.
Phase 2: Investigation
| Human Process | Frontier-Model-Assisted Process |
|---|---|
| I use a variety of tools to surface current research and conduct initial reading and investigation. I look for seminal research, trace citations, seek out dissenting views, and identify arguments that both confirm and challenge my hypotheses. Then I correct my hypotheses and re-catalog what I know and what additional information I need. The entire process is internal; it’s just me and the material. I test ideas. I make many more notes. As the process goes on, my thoughts become clearer. The work continues until I’m satisfied with the key ideas and the research support. |
I use a variety of tools to surface current research and conduct initial reading and investigation. The model absolutely increases the speed of the research. It surfaces seminal research and foundational thinkers more quickly than I'd find them on my own and gives me enough context about each to know if they’re truly applicable. The targeting isn't always right, so I do my own searches as well. I look for seminal research, trace citations, seek out dissenting views, and identify arguments that both confirm and challenge my hypotheses. The frontier model is both a thinking partner and a research assistant. The process is genuinely messy. I bring observations to the model (e.g., ‘I've seen this pattern; is it a thing?'), and we work through it. I push the model to name its sources; a confident answer without attribution gets verified. I go check the model's work, find things it missed, and bring them back. We banter about which ideas are strongest, how new findings change what we thought earlier, and where the reasoning holds or doesn't. Then I correct my hypotheses and re-catalog what I know and what I still need. The work continues until I’m satisfied with the key ideas and the research support. |
I usually start this phase with too many ideas and too many directions. I use the research and investigation to ultimately figure out what I actually think and know; humans can be confidently wrong, too. In the human process, that clarity arrives internally and gradually. Core ideas outlast the others. In the frontier-model process, core ideas are identified, evaluated, and accepted or rejected quickly. It feels different because the process is “witnessed.” The model and I arrive at a core idea together, and both recognize, "That's exactly the right core idea." That external confirmation changes the feeling of the moment, even when the moment itself is the same.
Let me make that “shared moment” concrete with a simple example. In early notes for this piece, I used the phrase "writing is inextricably linked to thought.” My frontier model collaborators both said, “True, but overused and boring.” One proposed “thought made visible." It sounded good, but that turn of phrase was just too “pretty.” I was immediately suspicious. So, I searched for where the phrase originated. The investigation revealed that Ambrose Bierce said it better in 1909. That discovery turned a more than century-old contrarian into the anchor for an essay about frontier models. That's the “aha moment” in this phase. It’s an unexpected discovery achieved through collaboration that honed what I thought I was building.
Either way, when it happens, I know. It won't be the last time core ideas shift or are refined. But it's the moment all my thoughts and ideas stop competing.
Phase 3: Writing
| Human Process | Frontier-Model-Assisted Process |
|---|---|
| The human writing process is sequential — at least at first. I start with the draft. For short-form pieces, I write the full first draft. For long-form, I start with a detailed outline and build from there. I write, and I keep writing, resisting the urge to edit until the thoughts are on the page and I have something substantial enough to dismantle. That draft is rough. It's supposed to be. | I still start the same way: a rough first draft for short-form pieces, a detailed outline for long-form. I hand the draft or outline over to the model with a sense of relief. My thoughts are on the page, and now I'll get real feedback, as long as I'm careful to ask for it. I read more carefully with a collaborator present. I don't have to ask and answer myself. The first pass answers two broad questions: What’s working? What’s not? Those results kick the collaboration off. If it’s a long-form piece, that initial read begins the hard work of writing point by point. The setup matters. I share the full piece along with my style guide and writing samples because the model needs to know whose voice it's working with. Then I assign the role deliberately. At this stage, my writing partner is warm but honest, focused on flow, voice, and whether the argument is coherent. The editor comes later, and I make that persona considerably tougher. |
Phase 4: Editing
| Human Process | Frontier-Model-Assisted Process |
|---|---|
| The editing phase is where the real work happens. I dissect every section, paragraph, and sentence. I evaluate structures, transitions, and word choice. I look for inconsistency in ideas. I evaluate everything rigorously against the central thesis, including things I loved when I wrote them. The effort is specific: it's the sustained mental focus of holding the central idea in my head while interrogating every sentence against it, for hours or days, across multiple drafts. Eventually, best practice calls for a first reader, a trusted critic who can tell me what’s working and what isn’t. I don’t always use one, but when my instincts tell me something isn’t working, that human perspective is irreplaceable. Then I iterate again. And again. New ideas jolt me awake at 2 a.m. throughout. The process is slow, effortful, and recursive. | The editing phase doesn't arrive as a discrete moment — it's woven into every exchange. I might debate a single paragraph for an hour. We test ideas, structures, transitions, logic, word choice, and metaphors relentlessly. We start at the section level, then iterate to the paragraph level, then to the sentence level. I eliminate content I love that just doesn't fit. The effort here is compounded: I'm maintaining the same sustained critical focus the human process demands, while simultaneously governing the collaboration. I’m evaluating every suggestion the model makes, pushing back, protecting my voice against a very capable entity that has its own patterns and tendencies. The 2 a.m. jolts still happen — that never changes — but now there's a collaborator ready when I wake up. The process is just as effortful, just as recursive, and in some ways more demanding. The setup matters here, too. As the piece evolves, so does the model's role. Throughout the editing process, my writing partner and I scrutinize everything together. That persona knows every decision, every iteration, every argument we've tested. At the close of each iteration (and there are many), I create a group of critics. I start new conversations with multiple models and assign tough editorial personas. This creates a blank slate for critique. They have no context about the iterations, the debates, and the things we discarded. Just the piece as it stands. And sometimes, when my instincts tell me something still isn't working, I bring in a human reader. Some things only a human can catch. |
Here’s how we make editing with a frontier model concrete. As I was going through one of the many editing iterations, I identified four sentences that were core but felt subtly “off.” My intuition was pinging. I spent over an hour with a frontier model debating over a verb and a new sentence in the introduction, a short paragraph about citing the definition for conversational discipline, and adding a single sentence to the conclusion. We analyzed what was “off” and why. Then we debated tone, word choice, sentence structure, and effects on the larger essay. Each sentence was an in-depth discussion. To many, I suspect spending that amount of time on four sentences looks like perfectionism or, maybe, over-engineering. It’s really about not letting a model default to a diluted, mediocre version of my intent.
Both processes ask everything of me: the thinking, the scrutiny, the 2 a.m. jolts, the willingness to discard something I loved because it doesn't support the argument. For me, a key element of the collaborative experience is having a sounding board. Frontier models give me the ability to think out loud and get an immediate, intelligent response. The differences in the processes are real because I invited frontier models into my writer’s room. They get a seat at the table, but never at the head of it. “We” is the right pronoun for the collaboration that happens in the writer’s room, and I’m not anthropomorphizing. As Ethan Mollick explains in Co-Intelligence: Thinking and Working with AI, treating the model as a partner produces better results. When that partnership “clicks,” it unlocks an intense, unbroken state of flow—the kind of immersion where I have spent eight consecutive hours locked in debate with models, entirely lost to hunger, thirst, or interruptions while holding a complex cognitive architecture together. At the end of both processes, I'm drained. I'm joyful. And, every single time, the product of conversational discipline is stronger than the product without it.
A Writer’s Reckoning
I’ve struggled with the stigma of AI-assisted writing. The experience reminds me of first grade, when a teacher took an inordinate amount of convincing to believe I had written a creative story myself, even though my mother had watched me write it line by line. That said, I refuse to yield to the stigma. I won’t deliberately insert typos. I won’t abandon literary devices. I won’t treat the em-dash like punctuation’s pariah just to prove “a human wrote this.” Early on, I worried that embracing the use of frontier models would mean giving up my identity as a writer. As my thinking has evolved, I've come to understand that disciplined collaboration with frontier models has made me better. Instead of taking my voice and style for granted, they’ve forced me to define both with precision—treating my prose like a data analysis exercise to map the patterns that are uniquely mine. When I engage in collaboration and critical governance, I sharpen my thinking. I organically stumble across new paths to evaluate. I fight with the model over and over and over. Ultimately, I create a better, more refined product.
I would be remiss if I didn’t share that the refined product can come at a cost. Conversational discipline is like a triathlon; it requires physical and mental stamina over a long period of time. I’ve treated it like a triathlon executed at the speed of a sprint. The model is always available, and default interaction patterns are frictionless. When I operate at maximum intensity, the model just reflects it back. I consume my biological reserves until I hit that proverbial wall. When I hit that wall, I’m tempted to compromise on discipline, which increases the risk of AI-generated slop. The lesson I’ve learned is that working with frontier models requires me to be more aggressive in self-governance to manage my own intensity in service to myself and my products.
The stigma associated with frontier-model-assisted writing misses the point. Producing good written products is a human responsibility, and it always has been. While researchers may be examining how to make models write in a “more human way,” there's no guarantee they'll protect us from ourselves. Bierce had it right in 1909 when he wrote, “. . .good writing . . . is clear thinking made visible.” A frontier model doesn't change that standard, but it does change how we uphold it. That's conversational discipline.
If you found this tactical breakdown valuable, you can read the earlier entries in the series here:
Read Part 1– Beyond the Prompt: The foundational journey to conversational discipline.
Read Part 2 – The Conductor’s Burden: Human fatigue and orchestration underlying AI interaction.
Feel free to Connect or Follow me on LinkedIn where I periodically share new thought pieces on artificial intelligence, business, and leadership.
Foundations
While this essay reflects my own practice and observation, its roots are in established research across several fields:
Business and AI: Chen, Te-Ping. Writers Are Going to Extremes to Prove They Didn’t Use AI. Wall Street Journal. May 6, 2026. This article is the initial catalyst for this essay. Link
Literary Tradition: Ambrose Bierce's (1909) Write It Right provides the piece's central standard: “…good writing . . . is clear thinking made visible.” That Bierce said it better than anyone since is either reassuring or humbling, depending on the day.
Conversational Discipline: Ethan Mollick's (2024) Co-Intelligence: Living and Working with AI informs my understanding of frontier models as collaborative partners. The emphasis on precision, clarity, and honest communication is grounded in Paul Grice's (1975) Cooperative Principle. The distinction between passive and disciplined use of frontier models is a practical application of Daniel Kahneman's (2011) Thinking, Fast and Slow, specifically his System 1 and System 2 framework. Conversational discipline applied to AI is defined in my 2026 essay Beyond The Prompt: Learning to Lead the Machine. Link
Experiential Foundation: Mihaly Csikszentmihalyi's (1990) Flow: The Psychology of Optimal Experience provides the experiential foundation for why disciplined collaboration feels rewarding.
Evidence:
Kosmyna et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab. The study included 55 college students, aged 18 to 39, from elite universities in the Boston area. Link
Varanasi, R. A., Nov, O., & Wiesenfeld, B. M. (2026). Investigating Writing Professionals' Relationships with Generative AI: How Combined Perceptions of Rivalry and Collaboration Shape Work Practices and Outcomes. Presented at CHI '26. Link
