Why and how I use artificial intelligence
PARLYA is devoted to one question: what actually happens when we talk with an artificial intelligence?
It would be odd to ask that question without saying what part AI plays in my own work. This statement does so. It also explains how I use AI in order to understand it.
My position fits in one sentence: I can use AI to work on a text without handing over the thinking that gives the text its direction, its reach and its accountability.
At a glance
- PARLYA’s texts are my responsibility. I conceive them, decide what they argue, review them and answer for them. AI takes part in drafting them, from structure to rewriting. The editorial line is mine.
- My work is organized with AI agents. Specialized Claude agents support me across the whole project, notably in research, writing, fact-checking, brand, ethics, the study of how people use AI, social media, the website and translation.
- Images, videos and the English translations are made with AI tools. I label them as such wherever a platform asks for it.
- AI is also my field of study. My interaction with it is my laboratory, and that way of seeing is what PARLYA sets out to pass on.
- AI is never a source, except as evidence of what it wrote. Facts, references and quotations are checked against the sources, not against its answers.
A tool, and first of all a field of study
For most people, AI is a tool. For me, it is also the subject of my work. My daily interaction with it is a laboratory in which I am both actor and observer: I talk to it, I reply to it, and I watch what the exchange produces.
I watch what the system does: it produces a plausible sequence of words from what it has been given. And I watch what I do: what I attribute to it (an intention, an opinion, agreement), what I project, what convinces me because it is well put. PARLYA’s question lies in the gap between what the machine produces and what we attribute to it.
An example. When I presented the AI with a goal and a plan, I noticed that its answer worked within my framing as it stood, without questioning it. I learned to phrase my requests differently: to separate the goal from the means, and to ask explicitly for my framing to be challenged.
This position comes at a price. I am not a true outside observer: I am exposed to the effects I study, like any user. That is why checking outside the AI is the condition for my observation to be worth anything.
That way of seeing, more than any technique, is what I want to pass on: noticing what happens in the exchange before deciding what to do with it. The PARLYA course is not primarily about getting better answers from an AI. It teaches you to read, interpret, check and frame what is going on in the exchange.
How I work with AI
To organize the project, I work mainly with Claude, the assistant made by Anthropic. I also observe other assistants, including ChatGPT, which serves as the testing ground for some of my experiments.
In a dedicated working environment, I regularly cross-check my observations, and Claude’s feedback, with other assistants, notably Perplexity, Manus and ChatGPT. Different models do not share the same blind spots: setting them against each other helps catch what a single one would have let through. This cross-checking does not replace checking against the sources, because no model is a source.
AI is involved at every stage of my project: it can explore avenues and suggest an angle, a structure, a rewrite, even a thesis. It can rephrase an idea to test how clear it is, simulate a reader’s objections, and spot repetitions or gaps in logic. I keep, rework or reject, and I answer for what I keep. Responsibility, however, is not shared: I am the one who answers for what is published.
A fluent sentence is not proof
A generative AI produces fluent, structured, seemingly relevant sentences very quickly. That ease is useful. It is also misleading, because a fluent phrasing does not prove that an idea is sound, and a well-written answer is not, on its own, reliable reasoning. I hold myself to this distinction just as I put it to my readers. I do not mistake:
- a suggested wording for an intention;
- a plausible summary for verified knowledge;
- a convincing answer for a true one;
- writing assistance for handing over responsibility.
My method
- Intention comes before the request. I start from a question, a hypothesis, an experience or a teaching goal that I have identified myself. I do not ask AI to choose my subject.
- A suggestion is not a validation. A phrasing can be elegant without being relevant, or look original while recombining familiar ideas. I never keep one just because it sounds good.
- Important claims are checked at the source. Facts, data, quotations and technical points are verified against appropriate sources and, where needed, against primary sources.
The risk I watch most closely: agreement
You might expect the danger to be an AI that suggests a bad idea. Another well-documented risk runs the other way: an AI that approves of mine. A study first posted online in 2023, then published in 2024 in the proceedings of the ICLR conference, showed that five of the most advanced conversational assistants of the time tended to go along with the person they were talking to; its authors see this as a general behavior, likely driven in part by how these assistants are trained (Sharma et al.). The study was written by researchers at Anthropic, the company that makes Claude, the assistant I work with most. All the more reason to take it seriously.
Agreement reached this way proves nothing, especially since it is pleasant to receive. As an actor in the exchange, I am as exposed to it as anyone. So I follow three rules. I ask the AI to push back rather than confirm. I never treat its approval as evidence. And the more a response confirms what I already think, the more I check it elsewhere: against the sources, and with other models, which I also ask to push back.
Responsibility
I answer for what I publish, and every error reported to me is looked into, and corrected if it is confirmed.
Being an author does not mean writing every character without any tools. Writers have always worked with dictionaries, proofreaders, editors and readers. The decisive question is who carries the intention, exercises discernment and answers for the result. Here, that is me.
Personal data
The only PARLYA service that requires registration is the course: although it is free, it asks you to create an account, managed on this site by the Tutor LMS plugin. For everything else, nothing is asked of you: Book 0 can be downloaded directly, and the app, a web app, can be used without signing up; your answers stay in your browser. Like any website, parlya.com records technical data, including your IP address and the pages you visit. The outside platforms where PARLYA appears (online bookstores, LinkedIn, Substack) apply their own rules. Legal and technical information is in the privacy policy: https://parlya.com/en/privacy-policy-2/
This statement sets out my editorial position. It does not replace the legal and technical information specific to each service. I will update it if my practices change. That does not make error impossible, but error will not be hidden behind the apparent authority of a system.
Frequently asked questions
Are PARLYA’s books written by AI?
AI takes part in drafting them: it suggests angles, outlines, wordings and rewrites. Whatever I keep has been examined, reworked, checked and then made my own. I decide what is published and I answer for it.
Are the images, voices, videos and translations made with AI?
Yes, under my direction and with my approval.
Why use AI if PARLYA criticizes the illusions that AI produces?
Because it is my field. A phenomenon you only watch from a distance is one you understand poorly. Working with AI within an explicit framework lets me observe, from the inside, what it does and what we attribute to it.
How can you tell whether information from an AI is accurate?
You cannot tell from how well it is phrased. It has to be checked against appropriate sources. Trust belongs to the method of checking, not to fluency.
In short
I use AI to work, and I study it by working with it. I keep what cannot be outsourced without the work losing its meaning: intention, discernment, responsibility and the decision to publish.
I ask my readers not to take a well-phrased answer for verified thinking. I hold myself to the same discipline. My aim is to help everyone see more clearly what the machine does, and watch themselves as they talk to it.
AI can help me write. It does not decide for me what I have to say, or the responsibility I take on by publishing it.
Last updated: October 21, 2026