From personalized service to personalized practice

On 24 June 2026, I took part in the European Legal Sovereignty Days organized by the Legal Data Space and the Conseil National des Barreaux. This event was held in Brussels, in collaboration with the Brussels Bar. It dealt with the stakes of legal sovereignty, and notably with the importance of providing legal professionals with tools that comply with ethical principles.

In this context, I had to step in at short notice. I was contacted 2 days before the event to ask whether I could speak on the third panel, devoted to human oversight. This panel also dealt with the legal and ethical issues and with the limitations one ought to consider in the use of artificial intelligence in the professional practice of law. I responded positively to this request, because I have been thinking about these subjects for several months. To prepare my intervention, I retraced my wanderings and dug through my notes on this notion of human intervention that I had addressed in an article of 5 June 2026. In preparing this intervention, I discovered a thread of reflection that I had not really seen when I finished writing my article on the human in the loop.

With hindsight, I note that I had sown concepts: proletarianization, dispossession, grammatization, and, with hindsight, I see that my proposal aims at a shift from personalized service to a personalized practice. In this note I will lay out the path I suggest for arriving at this personalized practice.

First, let us come back to the regulatory amendment proposed by the CNB, which was voted in December 2025. It concerns the definition of legal consultation which, as a reminder, is a monopoly reserved to the lawyer in France. Within the framework of this amendment, the CNB decided to modify the definition of this service by replacing intellectual with personalized. Legal consultation is henceforth a personalized service of the lawyer.

Following this amendment, some commentators saw in it a form of proletarianization. This term must be understood in the sense in which Karl Marx used it, but also in the updated vision of Bernard Stiegler. Proletarianization in the sense of Karl Marx refers to the dispossession of workers of their productive capacity by machines. Bernard Stiegler updated this proletarianization by considering it not as proper to a particular class (workers), but as applicable to all classes. These can be dispossessed of their capacity to do by technical devices. In the age of generative artificial intelligence, or rather of generative devices, this vision finds a strong resonance. Bernard Stiegler died in 2020 and this reflection is becoming, for some, visionary.

Indeed, as I was able to write in a note on this dispossession of our operative capacity, generative devices will allow the user to accomplish a series of things, even though they do not have the competence to do so. Conversely, even though they do have the competence to do it, they will delegate the expression or the execution of that competence to the machine. It is through the formula “knowing without doing, doing without knowing,” which I developed, that a phenomenon of dispossession of our operative capacity can be generated. Also called cognitive atrophy or de-skilling, these notions target the same phenomenon, namely one where we no longer use a competence and which, because of this lack of use, will diminish or make this competence disappear.

So, at this stage, you may ask yourself about the link between all these elements. The link I propose to make is to consider that if the regulatory amendment of the Conseil National des Barreaux can be seen as a “simple” descriptive modification of a phenomenon that is under way, it can nonetheless become performative. Moreover, this performative mutation is not exclusive to the legal sector. It can be observed in every professional sector that uses generative devices, since the use of these devices amounts to delegating drafting, expressive or reflective competences, whereas they were generally executed by the individual, creating a form of dispossession.

For lawyers, I go a step further by considering that one can see in it a risk, indeed a danger, if one prefers to avoid euphemisms, to our democracy. Why? Because the Court of Justice of the European Union recalled, in September 2024, that the lawyer plays a fundamental role in our democratic societies. Consequently, if the profession mutates toward the execution of personalized services and no longer intellectual ones, this could mean that the lawyer would become a verifying automaton, abandoning, over the course of their use of generative devices, all critical and reflective sense. They will fill in the entries between the brackets of texts generated by machines to which they will have delegated, in the end, the entire process of reasoning, drafting and reflection. This vision is catastrophist, I am aware of that. It is no less plausible for it, and all those who observe the transformation these devices bring about will share, perhaps to varying degrees, this potentiality of dispossession. The progressive deterioration of humans’ cognitive, technical or social competences due to an excessive dependence on artificial intelligence tools is a subject under study and one that must be taken seriously because of the growing adoption of these devices (800 million weekly ChatGPT users in October 2025 according to OpenAI).

This is why, within the framework of these potentialities, I propose to mobilize Bernard Stiegler once again, and the notion of pharmakon. I consider, with the French philosopher, that one must consider the whole set of technical devices as pharmaka and, for this reason, that one must propose therapeutic practices (that is, treatments and means suited to cure, to relieve the sick) and pharmacological ones. In this pharmacological perspective, it is therefore a question of dosage. The idea is thus not to sweep aside all technologies and techniques with the back of one’s hand, but rather to use them in a reasoned manner so that these techniques remain remedies and not poisons.

Grammatization is one of the paths I propose as a therapeutic practice. This grammatization is a therapeutic way of using the machine by imposing on it the work sequence to be implemented. Before the arrival of generative devices, we spoke of automation, of workflow, of process, of work algorithms, which made it possible, on the basis of atomic sequences, to delegate the execution of these sequences so as to obtain a result as compliant as possible with our expectations. But this grammatization is a demanding configuration process, and one that can appear complex to many.

During my reflections on human oversight in the use of generative devices and the famous human in the loop, theorized by Ethan Mollick, I suggested integrating grammatization into this practice in order to remedy the absence of clear content in this notion of human oversight.

Indeed, human oversight was integrated into a regulatory provision for high-risk AI systems in the EU Regulation on AI. Then this notion transformed into an ethical standard and became a must-have of the governance of companies claiming to use these generative devices. Unfortunately, beneath this ethically loaded intention, the content of human oversight is very often hollow. Whereas the human is placed in the loop, at the proofreading of the result, I propose to place the human upstream by integrating grammatization, as a work of configuring the generative device.

Grammatization aims to model production protocols and it is placed upstream of the process to be supervised. It then allows the machine to execute the user’s request on the basis of more precise and clear markers. This grammatization, which used to be complex, is today far easier to implement: it is “enough” to explain to the machine how we do the work in natural language. The machine will then transform the content into reproducible protocols on the basis of this interaction.

On the other hand, this human control, this attentive proofreading invoked like a magic formula that one need only pronounce to increase the quality of the deliverable, itself also undergoes the rhythm and the cadences imposed by these generative devices. As Marion Carré points out in Le paradoxe du tapis roulant, we are influenced by the cadence of these machines, which impact us in our own cadences too. We always do more and the time for verification shrinks.

So, almost to force oneself to slow down, I imagined a verification protocol that is placed downstream of the machine’s production. This protocol has for its objective to distinguish a reading path through the generated content. This reading path would be a way of drawing the (re)reader’s attention to the hypotheses that were taken sometimes unilaterally, to postulates that would have been used, indeed invented or assumed, depending on the case. This aid to proofreading therefore makes it possible to have matter to reflect on and to supervise (a little more) in depth the content of the generative device used.

On this subject, I also addressed the recommendations that this proofreading protocol could offer on the production protocol used. But in (re)posing the subject, I think one must dissociate the verification protocol from the recommendation protocol. Since the production protocol and the verification protocol are mobilized with the initial instruction, why not add a recommendation protocol that would come to suggest or optimize the production protocol? For example, if the production protocol is incomplete on certain aspects, the recommendation protocol would make it possible to suggest a modification to the user. This would make it possible to have a continuous improvement of the production protocol over the course of its use, like a self-optimization of one’s own production, assisted by the machine.

So, of course — and this was one of the questions put to me when I presented this hypothesis at the conference — one must not, blindly or without control, authorize the machine to update the production protocol on the basis of the recommendation protocol. There is the mastery and the competence of the user, which must remain, and they must validate the recommendation proposals that are made by the machine. I even consider that the updating process must be the object of an independent solicitation of the machine, by reformulating the recommendation in order to update the production protocol.

To synthesize: I started from the new definition of legal consultation, reduced, according to some, to a “simple” personalized service. This new definition is criticized because it can create a proletarianization of the lawyer in the face of the machine and could generate a (progressive) dispossession of competences. To fight against this phenomenon, I propose the setting up of three protocols that make it possible to obtain not a personalized service but a personalization of one’s practice. It is not the result that is personalized. It is not a standardized legal consultation that one will personalize by taking up the elements proper to the recipients. It is far more than that. It is no longer a matter of personalizing a consultation in the face of a client’s case, but rather of personalizing a practice of using a device.

I think that personalization must not take place at the level of the content proposed by the generative device but at the level of the manner of producing with the device. It is there, I think, that it is fundamental to insist and to work toward the training of the users of generative devices, because this personalization will then allow a real therapeutic practice.

You may perhaps have noticed that I use the notion of generative device? I will explain why in a future note.

Have a lovely summer!