A Voice From Nowhere

A July 2026 essay on zanlib.dev, written after Pope Leo XIV's first encyclical Magnifica humanitas (15 May 2026) turned up on X, LinkedIn, Hacker News, and as the first .va link ever posted to lobste.rs — upvoted, and argued about in good faith. The author is Polish and Catholic and says so, and treats that reception as the thing worth noticing first: encyclicals are addressed to "all people of good will" and have to survive on their merits outside the Church, which Laudato Si' had already managed when it turned up in UN ecology panels and in climate papers.

He decided against a full commentary. What he extracts instead is one property of these tools that the encyclical names and that, in his reading, the software field has agreed not to look at.

The argument

Technology is never neutral, and its makers have a particular character. The encyclical's §9 is the hinge: technology "takes on the characteristics of those who devise, finance, regulate and use it." That a tool carries its makers' character is old. The specific character here is not. The people building these systems are heirs of Western rationalism and the subject-object split, held so long that it no longer registers as a position — the conviction that if you subtract prejudice, mood, circumstance, and bias, what remains is the world seen plain and objective. He backs this with Pirsig: the rationality once used to escape one's immediate surroundings escaped so successfully that it became the thing romantics now try to escape from.

The pose of standing nowhere is itself a position. §100: "the apparent objectivity of the responses can lead us to overlook the fact that they reflect the cultural assumptions of those who designed and trained them." Absence of a viewpoint is not what is on offer; a viewpoint hidden well enough that its holders have hidden it from themselves is.

But the model's confidence is emptier than a human's failed objectivity. This is where the essay stops restating the encyclical and starts making its own claim. When a person reaches for objectivity and misses, you can see the reaching under the surface of the text. Under generated prose and generated code there is no reaching to see.

The path underneath. §99 supplies the frame: these systems imitate functions of human intelligence and often beat it on speed, but the power is tied to data processing, and they "do not understand what they produce." The essay's own image comes from chapter 6 of Zen and the Art of Motorcycle Maintenance, where Pirsig takes a motorcycle apart several different ways. The divisions are not welded into the machine. A rider splits it by function, a mechanic by assembly, a parts vendor by the factory's numbering scheme, and each chooses according to what he already understands, what he has watched fail, what he happens to care about. The manual that ships with the bike hides all of it behind an impersonal spectator's voice, as though nobody stood behind it. Somebody did. "Motorcycles were not found lying in a field. People made them run, and the making left a trail."

That trail is the path: the structure of thought underpinning anything a person makes. You sort the grains of sand a certain way, and the sorting is yours.

The machine has no path, and — this is the load-bearing part — it imitates one. The form gets reproduced without the comprehension that generated it. Hence the specific unease of reading generated content: it can be correct in every meaningful sense, it looks right, and it still reads as though nothing is underneath it.

The qualification that keeps the argument honest. The path is not wholly absent. Generating something requires putting knowledge in, and the input — the prompt, usually — is a human artefact with a path under it like any other. The problem is that a reader gets the output, not the input. The wider the gap between the last point of human judgement and the finished artefact, the harder it becomes to make out what the path was, or whether there was one.

Why reading it is tiring. When you read, you are reflexively looking for structure: why here and not there, what was considered and set aside, where the author was unsure of his own assumptions. When you cannot tell a careful writer from a stochastic parrot, you cannot tell whether the structure is there to be found. You are searching for something without knowing whether it exists, which is exhausting in a way that reading bad-but-human writing is not.

The second erosion: the path in the worker

If the tool erases the path in the work, it more slowly erases the path in the person doing the work. §140 supplies the claim — the speed and ease of getting answers risks extinguishing the desire to ask questions, with Plato on the deepest things being learned only after time and effort, "striking upon" ideas like flint until the spark catches.

The essay's version is a three-step decline in serendipity. Consulting a manual or the documentation for one problem used to mean stumbling over unrelated things and filing them away half-consciously. Google removed some of that by pointing at the one relevant section, but you still landed on a whole page and might glance sideways on the way past. A chatbot answer is trimmed to exactly the query, with no margins and no digressions. When the chat window is your only interface to the codebase, you never see what surrounds the diffs.

The endpoint is operational rather than sentimental: when you no longer know what your code does or where the boundaries of your system are, you cannot maintain it, and if an entire team disconnects from what is deployed, that is lights out. He cites Dex Horthy's Why Software Factories Fail here, the same essay the SlopCodeBench write-up in benchmarking-opus-5-slopcodebench follows from.

This is cognitive-debt and dont-outsource-learning arrived at from a different direction. Osmani's case is built on three studies with EEG traces and comprehension quizzes; this one is built on a phenomenology of what reading and searching used to feel like. They converge on the same mechanism — the friction was where the learning lived — and the encyclical adds the claim that the desire to ask is what gets extinguished, not merely the skill.

The signature and what it hides

The last movement is about accountability, and it is the most concrete part of the essay. §105 asks for "the possibility of identifying who must 'account' for decisions, justify them, monitor them, and, when necessary, challenge them and remedy any harm caused."

The author's counterexample is a job he worked at a large professional-services client, where every line traced to a person: commits labelled with Jira tickets, tickets carrying sign-off from the subject-matter expert who had confirmed the business logic, tickets inside epics signed off by a manager. When something failed badly, everyone in that chain joined the call. Colleagues found this unhealthy and argued for blameless postmortems.

His one-line verdict on what replaced it: "We have them these days: 'oh, Claude wrote it like that. You have it too, just prompt it for a fix.'"

Then the mechanism by which the chain quietly detaches. People are spending real effort divorcing themselves from the machine's output. Some colleagues will not acknowledge that a commit was AI-written, because acknowledging it invites the question of whether they understood the problem at all — so they sign it with their own name and omit Co-Authored-By. And it is impolite to suggest that a colleague's work is generated, since the suggestion implies he did not think. So nobody asks. What is lost is not the identity of the author but the level at which understanding last existed before the work was handed off, or whether it ever existed.

The encyclical's hardest line on this is about war — "it is not permissible to entrust lethal or otherwise irreversible decisions to artificial systems" (§198) — and the essay generalizes it: every system should have a chain of accountability terminating in a person, if only out of respect and common courtesy. The trouble with the voice from nowhere is that it lets the chain terminate in nothing while looking, to everyone else, like it terminates in you.

The closing inversion: technology takes on the character of its makers, and what should worry us is not the character it took on but how readily we learned to write in a voice that admits to no maker at all.

How it sits against the rest of the vault

The vault already holds most of the pieces this essay assembles, and it is worth being precise about which claim is new.

Detection: credibility-as-slop-test establishes that you cannot tell from the surface whether thought went in, and answers it by moving the question from the text to the author's willingness to stake credibility on it. This essay reaches the same dead end and does not offer the same exit — it stops at the exhaustion of searching for structure that may not exist. slop-marker-convention is the producer-side answer, and the omitted Co-Authored-By described here is exactly that convention run in reverse: a deliberate refusal to mark, for a reason (avoiding the competence question) that the marker proposals do not address.

The direct counter-position is checking-for-human-writing, which holds that authorship never lived in the path under the words at all. Shehu grants that writing by hand focuses thought and then refuses to promote it into a test of humanity: what can be asked of a finished text is who wanted to say it, who chose what stayed, and who accepts the consequences of publishing it. He is arguing about ownership of the result where this essay is arguing about what the reader can make out in it, so the two only collide if either is read as answering both.

Attribution: agent-identity-attribution is the infrastructure answer, giving agents their own keys and signed event trails. Its own limit is the one this essay is entirely about. Signatures recover who, never how much thinking, and the essay's target is understanding rather than identity, so cryptography does not reach it.

Review: control-the-ideas-not-the-code argues that reading generated code line by line is wasted effort and the design is what you should own. That is compatible with this essay only if the design work is where the path gets laid down and stays legible — which is antirez's DESIGN.md proposal, and which the essay would presumably accept. reviewing-ai-code and cult-of-vibe-coding take the opposite side. Where this piece differs from all of them is that it is not primarily about whether the code is correct. It grants that generated work can be correct in every meaningful sense and still be unreadable in the specific way that matters.

Preservation: the essay's own qualifier, that the prompt is the artefact carrying the path, suggests archiving the inputs. memorizing-session-transcripts is the negative result on the naive version of that — raw transcripts turned out to be residue, and the durable part had already been distilled into commit messages, PR descriptions, and docs. Which is to say the practice that would actually preserve the path is the traceability regime the author's professional-services client had, not a transcript store.

anti-llm-discourse maps the genre; this belongs in its craft-and-comprehension slot rather than the economic or ethical ones, and is unusual in that cluster for arguing from a religious document without asking the reader to accept its authority.

Cross-references

• cognitive-debt, dont-outsource-learning — the measured version of the foreclosure argument
• credibility-as-slop-test, slop-marker-convention, human-made-disclosure — the detection problem and the three declared answers to it
• agent-identity-attribution — signatures recover who, not how much understanding
• control-the-ideas-not-the-code, reviewing-ai-code, cult-of-vibe-coding — the review positions this is orthogonal to
• memorizing-session-transcripts — why storing the inputs does not by itself preserve the path
• building-syntaqlite-ai — the same loss of mental model, told as autobiography
• anti-llm-discourse — where this sits in the broader map
• benchmarking-opus-5-slopcodebench — shares a citation, and measures the maintenance failure this predicts