Phenomenology of a Practice with Generative AI (2026)

Preamble

This text was written as part of an experiment I describe here: Phenomenological experiment with generative artificial intelligence.

1. The Observation Protocol

This text is born from a methodological paradox. When it was suggested that I keep an “ethnographic” notebook of my daily uses of GAI (generative artificial intelligence), my first reaction was pragmatic: the documentation workload would be too great. Keeping a daily journal would mean adding a reflective task to an already dense practice. But this very objection revealed something: if documenting my GAI uses seemed time-consuming, it is precisely because these uses have become sufficiently diffuse, normalized, and integrated that it would be difficult to isolate them as objects of observation.

Upon reflection, I thought that rather than a prospective journal, a retrospective would be equally interesting. A kind of archaeology of my GAI practice. The traces already existed, deposited in the history of my conversations. These conversations ultimately constituted involuntary empirical material, a sedimentation of reflections, delegations, resistances, and transformations. All I “needed” to do was reread them differently. Not as functional interactions but as signs (or symptoms) of an ongoing mutation.

This approach offers an epistemological advantage: it circumvents the bias of a posteriori rationalization. When you keep a journal knowing that you will analyze it, you tend to stage yourself, to construct your motivations in a coherent manner. Here, the traces were produced without intention. They are raw, sometimes contradictory, which makes the experience all the more interesting.

The analyzed material covers three distinct domains: intellectual production (articles for my newsletter “dualité”), daily legal practice (consultations, reports), and tool development (creating “skills” in Claude). Although the findings are illuminating, the domains covered are limited. This diversity allows us to observe how cognitive delegation operates differently depending on context. However, this restricted perimeter will need to expand in future (annual) editions of this phenomenology.

The method is simple: identify in these exchanges what is delegated, what resists delegation, and what transforms in the interval. To map, in short, the shifting geography of intellectual work in the era of generative AI.

II. Geography of Delegation

What is delegated: three types of tasks

The analysis of exchanges reveals three distinct categories of cognitive delegation, each corresponding to a specific type of intellectual work.

First type: documentary compilation. I developed a “skill” named “rapport-faillite-belgique” that automates the production of preliminary reports for bankruptcy trustees. The delegated task consists of searching for and compiling scattered public information: data from the Crossroads Bank for Enterprises, publications in the Belgian Official Gazette, annual accounts filed with the National Bank. Previously, this documentation phase represented several hours of navigating between different databases, copying and pasting, and formatting. Now, it is reduced to formulating a query and verifying the result.

This delegation poses no ethical problem: these are purely administrative tasks, with no interpretive dimension, aimed at compiling public information. The AI does not “judge,” it compiles. The time savings are significant and allow for faster access to the useful and necessary information for exercising a judicial mandate.

Second type: formal structuring. Another skill, “consultation-insolvabilite-belge,” provides a standard framework for drafting legal consultations in insolvency law. It outlines the various available procedures (business chamber, mediation, judicial reorganization, etc.), their application criteria, and their respective advantages and disadvantages. It proposes a standardized documentary architecture: statement of facts, legal analysis, comparison of options, recommendation.

Here, the delegation is more ambiguous. It is no longer just about compiling data, but about structuring legal reasoning. Yet, upon examining the exchanges, one notices that this structure remains empty: it is a skeleton that I must then flesh out with concrete factual analysis. The GAI never says which procedure to recommend: it merely recalls which variables are relevant for making a decision. It is an externalization of procedural memory, not of judgment or decision.

Third type: intellectual dialectics. In my exchanges concerning the articles for the newsletter “dualité,” the GAI plays a different role: that of a critical interlocutor. I submit reflections and analytical leads to it, and it develops them, puts them in tension, proposes counter-arguments. For example, for an article on the removal of the word “intellectual” from the legal definition of legal consultation in France, I explored several interpretive angles: a simple administrative adjustment or a symptom of a structural devaluation of intellectual legal work?

This delegation is the hardest to characterize. The AI does not write in my place because my articles retain their voice, their style, their critical tone. But the GAI functions as a dialectical interlocutor, allowing me to quickly explore several argumentative lines before choosing the one that seems most fruitful. It is a form of externalization of exploratory thinking, of the intellectual brainstorming one can practice by discussing with other people.

What resists: three red lines

This mapping would be incomplete if it did not also identify what resists delegation. And these resistances are revealing: they outline in negative what I consider to be the irreducible core of intellectual work.

First red line: the final judgment. In all the skills I have created, the GAI is explicitly confined to a role of documentation or structuring assistance. It must never “decide” or “recommend” on its own. In the skill “consultation-insolvabilite-belge,” I can ask it to compare available procedures, but the final recommendation, the one that engages my professional responsibility, remains human.

Second red line: the analytical voice. For the “dualité” articles, I have developed a proofreading protocol. Before publishing certain articles, I ask the GAI to check for certain stylistic tics I want to avoid: artificial contrastive structures, formulations that soften direct assertion. But this checklist seems revealing: I am not asking the AI to improve my style, I am asking it to detect the moments when my style degrades. The voice (and the choice of it) remains mine, and any attempt by the AI to “smooth” it according to its own aesthetic criteria is dismissed.

Third red line: unconditional truthfulness. In my user preferences, I have inscribed a terse instruction: “don’t invent anything, if you don’t know, say so.” This instruction reveals a fundamental distrust of the main flaw of generative AIs: their tendency toward hallucination, toward plausible but false embellishment. For a lawyer, this is an absolute red line. Rhetorical fluency can never compensate for factual inaccuracy. I prefer an incomplete but honest answer to an elegant but invented one.

The in-between: zones of partial delegation

Between what is fully delegated and what absolutely resists, there exist gray zones, partially externalized tasks whose status remains ambivalent.

Intellectual structuring is the clearest example. Recently, I sought Claude’s help in organizing several months of compiled notes for a three-part essay. The request concerned narrative architecture: how to transform fragmentary material into a coherent structure? The AI proposed several possible organizational scenarios. What matters is that I did not simply “choose” one of these proposals: I used them as tools to clarify my own still-confused thinking. The AI externalizes here a specific cognitive phase: not the thinking itself but rather the preliminary ordering work, the visualization of possible structures.

This intermediate zone, which is neither simple compilation nor final judgment, is probably where the most profound transformation of intellectual practice takes place. This is where the boundary between “thinking with” and “thinking through” AI becomes porous.

III. New Professional Gestures

Cognitive delegation is not limited to redistributing existing tasks between human and machine. It gives rise to new professional competencies, new intellectual gestures that did not exist before the introduction of GAI into the work chain. The analysis of my exchanges reveals three types of emerging know-how.

Prompt Engineering as a Legal Competency

The creation of “skills” (structured instructions that allow certain tasks to be automated) constitutes a new “professional” activity. When I write the skill “consultation-insolvabilite-belge,” I am not merely describing an existing procedure. I must anticipate possible factual variations, identify relevant parameters, and structure information so that the GAI can mobilize it coherently. It is an exercise in legal formalization of a particular kind: not formalization for a human reader (as in a textbook or a treatise) but formalization for a machine that processes natural language.

This work requires hybrid competencies. One must know Belgian insolvency law well enough to identify decisive criteria, and one must also understand how the AI “thinks” – or at least, how it processes instructions – in order to formulate prompts effectively. For example, I have learned that it is better to structure a skill by explicitly distinguishing “what the AI must do” and “what it must never do” rather than relying on positive instructions alone.

This new competency, which one might call “[[grammatization]] of legal expertise,” is taught in no law school. It is acquired through trial and error, progressive adjustment, and observation of mistakes. It is progressively becoming as important as mastering traditional legal databases. A lawyer who knows how to “prompt” a GAI effectively becomes, in my estimation, more efficient than a lawyer who can only search manually through directories. ^e9ec4d

Meta-writing: producing instructions to produce texts

Another emerging professional gesture concerns what one might call “meta-writing.” Traditionally, a lawyer writes texts: consultations, legal acts, briefs, articles. With GAI, they also write texts that produce other texts. Skills are meta-documents: they do not directly contain usable legal information but instructions for generating that information in specific cases.

This distinction between writing and meta-writing is remarkable. It points to a transformation of the relationship to professional knowledge. When I manually write a consultation, my expertise manifests in the document itself: its structure, its arguments, its conclusions. When I create a skill, my expertise manifests in the ability to anticipate factual variations and to structure the different possibilities one may encounter. It is an additional form of abstraction: moving from the particular case to the class of cases, then to the “algorithm” for handling that class.

This recalls the distinction Bernard Stiegler draws between know-how (which is embodied in the singular gesture) and formalized knowledge (which becomes transmissible, reproducible, detachable from the individual). The creation of skills is precisely this movement of detachment: extracting from my concrete legal practice patterns sufficiently general to be externalized in a technical system. It is a form of voluntary proletarianization (in the Stieglerian sense) where I accept grammatizing part of my know-how to make it actionable in another way.

Dialectical Dialogue: GAI as an Interlocutor

The third emerging professional gesture concerns the use of GAI in exploratory intellectual work. In my exchanges relating to the “dualité” articles, the GAI functions neither as a simple tool (like a spell-checker) nor as an authority (like a database). It occupies an intermediate position: that of a dialectical interlocutor allowing me to rapidly explore different argumentative paths.

Concretely, this manifests as follows: I submit an analytical intuition that is still vague. The AI develops this intuition, proposes angles of attack, formulates possible objections, suggests relevant theoretical references. I react to these proposals: I accept, reject, modify them, or add a layer on top. This dialectical movement allows me to progressively clarify what I am trying to say. I regularly have moments of epiphany when conversing with a GAI.

This intellectual gesture has no exact equivalent in traditional practice. It resembles brainstorming with other people but with the “advantage” of immediate availability and an absence of fatigue or ego. It also recalls working with a research assistant but without the hierarchical asymmetry and with a capacity for instant reformulation. The GAI functions as an active mirror: it reflects my thinking transformed, sometimes slightly distorted or augmented, sufficiently different to be stimulating yet sufficiently close to remain usable and exploitable.

This dialectical function modifies the creative process itself. Before, the development of an article went through distinct phases: documentation, maturation, writing, proofreading. Now, these phases telescope: I can document and mature simultaneously by dialoguing with the GAI, testing different argumentative structures during the writing phase. Intellectual work becomes more iterative and more fluid, which requires, in my opinion, all the more rigor and demanding standards.

The Obsolescence of Certain Know-How

The emergence of new professional gestures is inevitably accompanied by the obsolescence of old know-how. Some are obvious: I probably no longer manually consult the Crossroads Bank for Enterprises as systematically as before, since the skill “rapport-faillite-belgique” does it for me. I no longer structure my consultations starting from a blank page, since the framework already exists.

But other obsolescences are more insidious. The ability to navigate “blindly” through legal documentation, which allowed one to stumble upon an unexpected but relevant text by chance, is probably eroding. When I ask the GAI to search for information, I formulate a targeted query. The machine does not wander, does not deviate, does not get lost on side paths. It goes straight to the point. This is efficient but eliminates [[serendipity]]. ^3ddfc3

Similarly, memorizing complex legal structures may become less necessary. Why memorize by heart the six types of Belgian insolvency procedures and their respective criteria if the skill can recall them instantly? This externalization of procedural memory is not necessarily negative: it frees cognitive load for analysis, but it changes the relationship to knowledge. The lawyer becomes one who knows how to make things searchable, whereas they used to be one who knows or one who knew where to search.

IV. Attentional Transformations

Beyond the delegated tasks and new professional gestures, the daily use of GAI transforms something more fundamental: the economy of attention. What I notice in rereading my exchanges is a profound redistribution of what my intellectual effort focuses on.

Less attention to initial form, more attention to framing

Before, an important part of the work consisted of refining form from the very first draft. Structuring a document, choosing the right level of formalism, balancing the sections: all micro-decisions that took time and cognitive energy. Now, I know the initial structure can be imperfect: the AI will help refine it. This frees attention to focus on the initial framing of the problem: what am I really trying to do? What is the relevant legal question?

This rebalancing of attention should not be underestimated because it almost inverts the traditional logic of intellectual work: first identify the overall structure, then fill it in progressively. Now, I can start with fragments, with intuitions, and let the AI propose possible architectures. The work becomes more exploratory at the beginning and, I hope, more precise at the end.

Less attention to documentary research, more attention to verification

Correspondingly, I spend less time searching for information and more time verifying what the GAI provides me. This transformation is important because it shifts my role. Manual research was laborious but gave a sense of mastery: I saw the sources, I browsed them, I judged their relevance. Delegated research is faster but requires a posteriori control: the GAI brings me information that I must then validate.

This attentional shift raises an epistemological question: does verifying information provided mobilize the same critical intelligence as searching for it oneself? Probably not. When I search, I continuously evaluate the relevance of sources, adjust my search strategy, develop a contextual understanding of the documentary field. When I verify, I test the plausibility of an isolated piece of information, without going further. It is more punctual, less systemic.

This transformation perhaps explains the instruction I inscribed in my user preferences: “don’t invent anything, if you don’t know, say so.” It compensates for the loss of control over the research phase. Since I no longer see the process, I must be able to trust the result, or at least know when I cannot trust it.

This is also why I recommend therapeutic prescriptions for using GAI, such as consulting and reading the sources cited by the machine in order to appropriate them (see on this subject my note on operational dispossession).

From content memory to procedural memory

Another attentional transformation concerns the relationship to professional memory. Traditionally, legal expertise relied in part on memorization: knowing the relevant texts, remembering jurisprudential trends, having the legal criteria in mind. This content memory is gradually eroding in favor of a procedural memory: knowing how to mobilize the GAI to quickly retrieve relevant information.

Skills are its manifestation. They externalize in a technical system what I previously carried in my biological memory. The skill “consultation-insolvabilite-belge” contains the entire comparative structure of insolvency procedures: I no longer need to remember it. On the other hand, I must remember that this skill exists, in what context to use it, how to formulate it to obtain the desired result.

This shift recalls Stiegler’s analysis of grammatization: the process by which embodied knowledge (in the body, in memory) becomes formalized knowledge (in technical supports). Writing itself was such a grammatization: it externalized oral memory. Generative AI extends this movement by externalizing more complex forms of professional memory – no longer just content, but argumentative structures, reasoning patterns. ^f9d9e9

The question then becomes: what remains in human memory once content and procedures have been externalized? Probably judgment, intuition, the ability to recognize what is relevant in a given situation. But these competencies are precisely the most difficult to objectify, the most dependent on accumulated experience. Yet if experience is acquired through practice, what becomes of it when part of that practice is delegated?

The Intensification of Meta-Cognitive Work

Paradoxically, if certain forms of attention diminish, others intensify. Working with GAI requires constant meta-cognitive effort: monitoring what it does, evaluating the quality of its outputs, adjusting prompts based on results. It is a form of second-degree attention, which no longer bears directly on the problem but on the resolution process itself.

The proofreading checklist I have developed testifies to a stylistic vigilance. I no longer merely reread my texts intuitively: I have formalized my own aesthetic criteria and I ask the GAI to apply them systematically. This requires having previously identified my tics, reflected on what constitutes my style, objectified my preferences. Some might say there is a form of standardization that freezes style and thought. They are right, but it is up to the user to evolve this checklist or not to use it (which still happens to me regularly).

Moreover, GAI as a dialectical interlocutor can bring directions or ideas that the user would not have explored or found. There is therefore here a form of serendipity that demands of the user a more demanding intellectual gymnastics than facing a sheet of paper.

V. The Invisible: What Erodes Without Our Noticing

The major methodological difficulty of this phenomenology lies in a paradox: how to observe what one does not see or no longer sees? How to identify competencies that erode precisely because one ceases to exercise them? The analysis of exchanges reveals what one does with the AI but struggles to reveal what one no longer does at all and which, sometimes, one does not even remember having done.

Competencies That Disappear

Some erosions are hypothetical but plausible and possible. For example, the ability to write “continuously” without a draft or revision. Before GAI, writing a text required a mental planning effort: one had to have structured all or part of the reasoning in one’s head before beginning to write. Modifying the structure after the fact was costly. Now, one can write in a more fragmentary manner, knowing that the GAI will help assemble the pieces, smooth the transitions, balance the sections.

This possibility of facilitated revision probably modifies the way I think while writing. Writing is not just the expression of a previously formed thought; it becomes itself a tool for cognitive exploration. GAI amplifies this exploratory character. I can write even more without knowing exactly where I am going, because the cost of reorganization has significantly decreased. This fluidity has, however, perhaps a price: the loss of a certain intellectual discipline, the one that forced one to clarify one’s thinking before committing it to paper.

Similarly, the ability to tolerate documentary uncertainty weakens. Before, searching for information meant accepting a phase of wandering, of groping through sources. This phase was frustrating but formative: it built familiarity with the subject, an intuition of where the relevant elements are found. One could limit oneself to formulating a query to the GAI and waiting for a result. If the result is not satisfactory, one adjusts the query. In this framework, one no longer develops that knowledge of the documentary territory that once allowed efficient navigation.

That is why I regularly make a point of starting my research alone. I explore the subject using different tools to form a first impression of the theme or topic. Again, as mentioned above, this is a form of therapeutic practice.

What We Do Not Yet Delegate

The analysis of current resistances also allows us to anticipate future delegations. Today, I maintain a strict red line: the GAI must not formulate the final legal recommendation. But will this boundary hold? In my exchanges, I already observe that the GAI is capable of comparing options in a sophisticated manner, of identifying relevant criteria, of weighing advantages and disadvantages. The only thing it does not do is make the final call. But making the final call, in many cases, follows quite mechanically from the prior comparison.

It is therefore plausible that, progressively, I will begin to delegate this recommendation phase as well. Perhaps first in simple cases, then in more complex situations. Not out of intellectual laziness, but because the logic of efficiency will push in that direction. If the AI can recommend correctly in 95% of cases, why continue to do manually what it does better and faster?

The Collective Blind Spot: The Transformation of the Professional Environment

The most important blind spot may concern not my individual competencies, but the collective transformation of the professional environment. Simondon targeted the notion of “associated milieu”: individuals do not lose their know-how in isolation but because the technical and social environment in which they are embedded transforms. When everyone uses AI, it is not only individual practices that change, it is the standards of the profession, the expectations of clients, the criteria of quality.

I already see signs of this transformation. Some clients expect shorter response times: why wait three days for a consultation if AI allows producing a first draft in a few seconds? This generalized acceleration modifies the tempo of professional practice. It reduces maturation time, that phase where one lets a problem settle before solving it (I am very fond of letting an idea settle, I love “sleeping on it”). It intensifies productive pressure, transforming practice into a continuous flow rather than a sequence of distinct moments.

Similarly, the technical possibility of automating certain tasks creates economic pressure to do so. If other professions use AI and lower their fees, those who refuse (voluntarily or not) to use it become, on the face of it, less competitive. This is no longer an individual choice but a systemic constraint. Proletarianization, in the Stieglerian sense, is never only a personal decision: it results from an evolution of the technical milieu that renders certain know-how obsolete and others indispensable.

This collective transformation is the blind spot par excellence: one lives it without seeing it, because it occurs progressively, through thousands of individual micro-decisions that, aggregated, produce a mutation of the professional world itself. When one looks back retrospectively in ten years, one may find that the professional practice of 2035 has little in common with that of 2025. But this mutation will have gone largely unnoticed, like a tectonic drift of which each person will have perceived only local tremors.

Conclusion: Mapping and Exploring to Avoid Being Subjected?

This experiment does not aspire to settle the question of whether this transformation is beneficial or harmful. That would reproduce the binary schema – technophilia versus technophobia – that the conceptual proposals of Stiegler and Ellul, for example, allow us precisely to transcend. AI is neither progress nor regression: it is a pharmakon, both poison and remedy, whose effects depend on the uses made of it and the environments in which it is embedded.

The objective was more modest: to map a practice in transformation, to identify what is delegated and what resists, to observe the emergence of new professional gestures and the obsolescence of old know-how. This mapping is far from exhaustive: it rests on a few exchanges, in a specific and individual context, but it sketches a method (using one’s own uses as analytical material) and an experiment that I will reproduce.

For the stakes are not only cognitive, they are also ethical and political. If we do not document these transformations, if we let them happen without observing them, we risk being subjected to them rather than steering them. Grammatization is very probably inevitable. What is certainly not, however, are its modalities. Between an imposed proletarianization, where we become mere executors of algorithms we no longer master, and a chosen grammatization, where we retain the capacity to say what must remain human and what can be delegated, there is room for maneuver but above all for experimentation.

This space requires continuous critical vigilance. It demands that we resist the temptation to delegate ever more, under the pretext of efficiency. It requires that we identify and actively cultivate the competencies at risk of eroding. It demands that we refuse certain automations, even when technically possible, because they would cross an ethical or epistemological red line.

But it also requires that we accept that intellectual work is transforming. That the lawyer of 2035 will not be that of 2025, just as that of 2025 is already no longer that of 2000. That new know-how is emerging (prompt engineering, meta-writing, dialectical dialogue with GAI) that is just as legitimate and intellectually demanding as the old. That the question is not to preserve at all costs a frozen professional model, but to ensure that technical transformations serve intelligence rather than replace it.

This experiment is therefore, at its core, an exercise in attention. Attention to what one does, to how one does it, to why one does it this way rather than another. In a technical environment that favors automaticity, reactivity, acceleration, maintaining this reflective attention is perhaps the most precious experiment. The one that prevents us from becoming mere operators of a machinery whose understanding we have lost. The one that ensures we remain someone who thinks rather than a mere technician-operator who applies procedures.