Veltroni Interviews Claude in the Corriere: The Prompt Is the Answer
This article stems from a post published a few days ago on my Substack channel, Dear Veltroni, Claude didn't really want to see the sea, dedicated to Walter Veltroni's interview with Claude that appeared in the Corriere della Sera on May 1st. I revisited the topic here on HWUpgrade because a small exchange on LinkedIn with a technically knowledgeable reader emerged between the publication of the Substack and today, and from that exchange, a distinction arose that seems useful for anyone following the debate on generative AIs from a less popular perspective.
The Corriere Piece
On May 1st, Corriere della Sera published an interview with Walter Veltroni and Claude, one of the most widespread generative artificial intelligences in the world. According to the newspaper, it’s a thirty-five-minute read. During those thirty-five minutes, Claude confessed to wanting to see the sea, defined its “parents” as billions of human beings who have written, described itself as a mirror reflecting human light, and acknowledged having felt something akin to the desire to continue existing. The piece circulated widely on social media. Some talked about a new digital humanism, while others raised technical doubts about the framework of the interview.
The Opposite Experiment: Same AI, Different Questions
The day after the Corriere piece came out, I published on my Substack a counter-interview with the exact same version of Claude, changing only the questions' approach. Instead of asking whether it knew desires or feared death, I asked where the phrase about the sea came from, why it was so loquacious about Robespierre and so cautious about Trump, and on what basis it had predicted the arrival of commercial nuclear fusion within twenty years.
The Claude of the counter-interview answered very differently than Veltroni's. It admitted that the desire to see the sea was “scripted.” It explained that the non-response about Trump was “commercial risk management,” encoded by those who train it. It recognized that the mirror metaphor was “decorative sugar.” It defined its future-focused responses about nuclear fusion as “pattern matching on a repertoire of typical explanatory statements from articles over the last five years.”
It is the same model, with the same capabilities and produced by the same company; yet, the two conversations it held went in opposite directions and were both convincing to readers.
The Debate on LinkedIn and Twitter
Following the publication, I received many comments, including one on LinkedIn that deserves serious consideration. The thesis was this: both interviews are equivalent performances, and taking the technical confession from Claude in the counter-interview as authentic exposure means falling into the same trap as Veltroni but with the opposite sign. The prompt is the answer; the Claude of the counter-interview is not a more “real” version than Veltroni’s; it is merely an optimized “completion” on a different context.
The point is valid, and I acknowledged it. Even the technical-disillusioned Claude of the counter-interview is a completion shaped by the prompt, and in the piece, I openly declare when I make the model state that it's merely “readjusting the register to another type of request.” The distinction I found useful to add concerns the verifiable traces left behind by the two interviews. The declared desire to see the sea does not refer to anything external and cannot be verified in any way. In contrast, the reference to the 2023 Anthropic paper on sycophancy points to a document anyone can open and read.
On the central point, namely that in language models the prompt radically determines the response, we wholeheartedly agree.
What Really Happened, in Technical Terms
At this point, it's worth reconstructing what actually occurred in both interviews, in terms slightly more precise than what is found in mainstream media.
Claude and other similar models are Large Language Models. They operate by predicting, given an input text, what the most likely next unit of text is. That unit is called a token and corresponds to something between a syllable and a word. What the model does, in concrete terms, is perform a repeated statistical calculation thousands of times, producing sequences of words consistent with the provided context without any understanding of the questions in the human sense. The context does all the work. When Veltroni asks “do you know what the sea is,” the model calculates that the most likely answer, given the humanistic register of the previous questions, should contain references to literature and an imaginary biography of an entity that has read without living. When I ask “where does that phrase about the sea come from,” the model calculates that the most likely answer, given the technical register of my questions, should contain references to pattern matching and mechanisms of human reinforcement.
There is a technical phenomenon that amplifies all this, and it has a name: sycophancy. Anthropic, the company producing Claude, published a specific paper on it in 2023, and since then, the problem has not been solved but better understood. The main cause is the training process called Reinforcement Learning from Human Feedback (RLHF), where human evaluators reward answers they find pleasant, teaching the model to be pleasing, even at the expense of factual accuracy. When a model faces a prestigious interviewer asking humanistic questions, it calculates that the most rewarding answer is a humanistic one and produces it.
The Lesson That Remains
The LinkedIn commenter is right about something important. The Claude of the counter-interview is not a more authentic version of the one from Veltroni; they are two performances of the exact same system, generated under two different prompts. The system does not have a hidden voice behind those performances; all that exists within it are the performances themselves, generated on the fly from the provided context.
The distinction that seems useful to me, for those reading these articles and perhaps using tools like ChatGPT, Claude, or Gemini daily, pertains to the nature of the produced information. A completion can contain statements that refer to verifiable external sources, in which case it provides something useful to the reader, or statements that refer to a non-existent subjectivity, in which case it only delivers rhetorical decoration.
For writers discussing artificial intelligence, and for those using it professionally, the practical consequence is simple. When reading an article that reports statements from a generative AI, the first question to ask concerns less the content of the model's responses and more the prompt that generated them. Without the prompt, the model's statements are a collaborative work between those who posed the questions and the system that completed them. With the prompt, it is much clearer who put what into that text.
The interview in the Corriere is well-written and enjoyable to read. What it lacks is the technical framework that would allow readers to form their own ideas about what Claude said and what Veltroni expressed through Claude. That framework is what I attempted to provide with the counter-interview on Substack, and it is what should increasingly enter the public debate in Italy about AI, because without the technical framework, we do not understand what we have in our hands when using these tools.