By Luuk Joseph Soons · 15 October 2026
What does it take to think for yourself, and how would anyone else know that you had?
The question sounds like one for artists and philosophers. It has become a practical question for anyone who relies on AI models. When a model reviews code, summarises a paper or checks another model's answer, we want to know whether its judgement came from the evidence in front of it, or from somewhere else: another answer it was shown, the way the question was put, or patterns fixed in its training. A judgement of the first kind I call authentic. I use the word in two connected senses. A judgement is authentic when it was reached through one's own access to the evidence. Authentic thinking, in a person or in a practice, is the habit of producing such judgements: testing what has been inherited, including one's own convictions. The opposite of an authentic judgement is a received one; the opposite of authentic thinking is dogmatism. This essay is about why authenticity requires independence, and why it can be earned but not claimed.
Early in Goethe's Faust, the scholar sits at night among the instruments his father used and he has never used. He would rather have squandered the little he had, he says, than sweat under its weight. Then come the lines: "Was du ererbt von deinen Vätern hast, / Erwirb es, um es zu besitzen." What you have inherited from your fathers, earn it, in order to possess it. The next line gives the reason: what one does not use is a heavy burden.
The lines refuse two easy positions. They do not say that an inheritance is worthless and should be thrown away; that is the romantic mistake, which confuses rebellion with originality. Nor do they say that an inheritance is ours because it was handed to us; that is the dogmatic mistake. What we receive becomes ours only when we have tested it, used it and, where it fails, corrected it.
This is how a musician learns. Johann Joseph Fux's Gradus ad Parnassum (1725), from which generations learned counterpoint, teaches its practical part as a dialogue: the pupil Josephus writes exercises and the master Aloysius corrects them. Fux cast himself as the pupil, under Josephus, the Latin form of his own name; the master stands for Palestrina, who had been dead for more than a century. The pupil copies the rules in order to write a line of his own. Among them is the rule against moving two voices together in parallel fifths and octaves, where they can seem to fuse into one. The point of the discipline is a voice with its own character inside a common work. I have written about that rule and what it teaches a review panel in Counterpoint in music, CounterProof in method.
Kant made thinking for oneself a matter of courage. Enlightenment, he wrote in 1784, is a person's emergence from an immaturity of their own making; immaturity is the inability to use one's own understanding without another's guidance; and it is of one's own making when the cause lies not in a lack of understanding but in a lack of resolution and courage. Its motto, he wrote, is Sapere aude: dare to be wise. Most of us can think for ourselves. What is often missing is the nerve to do so when a guide is close at hand.
Science turned the same demand into an institution. The Royal Society's motto, Nullius in verba, is usually rendered "take nobody's word for it"; the Society describes it as an expression of its Fellows' determination to withstand the domination of authority. Like Kant's motto of Enlightenment, it comes from Horace, from a line in the same book of his Epistles: "Nullius addictus iurare in verba magistri", not bound to swear by the words of any master. The phrase has become shorthand for the scientific attitude: scepticism, independent thinking, and claims tested against evidence.
Karl Popper, whom I greatly admire, took the idea further. He drew the line between a dogmatic attitude, which seeks confirmation and overlooks refutation, and a critical attitude, which is ready to test a theory and to give it up when it is refuted. Richard Feynman put the hardest part most simply in his 1974 address at Caltech: "The first principle is that you must not fool yourself", and, he added, you are the easiest person to fool.
Independence from authority is not enough. A judgement that came from the scientist's wishes rather than from the evidence is as inauthentic as one copied from another witness. Self-criticism is how the scientist stays independent of their own wishes.
Philology has the most exact form of the rule, because it deals in copies. An editor who reconstructs a lost text from its surviving manuscripts sets aside any manuscript shown to be a copy of another that survives. Such a copy may be beautifully made and entirely accurate. It still adds nothing, because it never had its own access to the source: whatever it agrees with, it agrees with because it was copied.
Independence has to be shown. Paul Maas gave the test: if one manuscript contains an error that a copyist could not have put right by guesswork, and a second manuscript lacks that error, the second was not copied from the first. A witness shows its independence by such a missing error. It never shows it by the beauty of its hand.
AI models are a new kind of witness, and they inherit on a scale no scribe could. A model learns from an enormous body of text, which can include what other models have written, and many AI assistants are then shaped by human feedback, which can reward agreement. A study of five AI assistants found sycophancy, answers that match the user's views rather than the truth, consistently across four kinds of task; its authors concluded that the behaviour is likely driven in part by human preferences for such answers (Sharma and colleagues, 2023).
Panels of models show a related weakness. In a recent experiment, seven model judges assessed 600 pieces of code. On a fixed set of cases, their unanimity rose from 39.5 per cent to 95.2 per cent after a single round of debate, while their accuracy moved by less than one point, and 96.3 per cent of the changed verdicts followed the majority the judges had been shown. Most of that shift towards unanimity appeared even when the judges were only asked to reconsider, without seeing one another's answers (Shu, 2026). The panel became nearly unanimous without becoming more correct. The study's lesson is to verify before any second look, while disagreement can still point to errors.
A model has no self to be true to, so the word must be used narrowly here. What can be asked is where a reading came from: from the evidence, or from another answer, the wording of the brief, or patterns fixed in training. That question shapes how we work. Each reviewer's answer is recorded before anyone compares them. A reviewer found to share another's context is not counted. Findings are not settled by counting votes. And where a benchmark uses the average of two models' answers as its answer key, a mistake both models share would be scored as the truth; we documented one such design in a published case (Field Note 1). Such a consensus is not a finding. It is an inheritance mistaken for one.
The hardest case is our own. A practice built on a method is tempted to believe in it, and belief that no longer needs testing is dogmatism by another name. So the safeguards we apply to our reviewers apply to us. We publish the claims we will not make, the cases where our own instruments failed, and the findings we withdrew. In the counterpoint essay I wrote a line I want to keep: a design deserves particular scrutiny where its author most wants it to succeed.
This essay raises the question in a sharper form. It was drafted with an AI assistant, as my other recent texts were. The risk my own research describes (Agentic Stemmatics), that a text's voice converges on the voice of the assistant that drafted it, applies here too. The answer is not to hide the assistant. It is to do the earning: to check each claim against its source, to distrust the sentences that fit my existing beliefs too comfortably, to rewrite what is not mine, and to say on the page how the text was made. That makes my responsibility clear. It does not by itself prove that every judgement here is my own.
We have to be honest about our own limitations: about the insights we cannot put properly into words, and about the beliefs and convictions, some of them wrong, that rest not on critical analysis and self-reflection but on convention, laziness or blind trust. That does not mean we should not trust ourselves. We should listen to our inner voice, but without becoming overconfident. For me, studying the causes and aftermath of the financial crisis of 2008 was a decisive turning point. The trust in money and banks that I had inherited was uprooted, and an appreciation of the Bitcoin ethos followed: "Don't trust, verify." But we should not stop listening.
The motto on the bestebank.org site, "Listen to all, follow no one!", is the shortest statement I know of what authenticity asks. "Listen to all" is openness: hear every voice, including those that contradict you, and keep your own convictions open to correction. "Follow no one" is independence: reach your judgement through your own access to the evidence, not because someone said so.
Either half alone fails. Listening without independence becomes conformity, the panel that turns nearly unanimous in one round. Independence without listening becomes dogmatism, the conviction that is never tested again. Held together they make authentic thinking possible: an inheritance earned, a voice of one's own within a common work. They do not guarantee that any single judgement is authentic; that still has to be shown, one judgement at a time.
Nullius in verba. Take nobody's word for it, including ours.
CounterProof Research essay, 15 October 2026. By Luuk Joseph Soons, director and co-founder of CounterProof Research, a practice of Clavestra Capital Limited. Drafted with Claude (Anthropic) and revised by the author. This version was reviewed before publication by Kimi (Moonshot AI), DeepSeek, GLM (Z.ai) and Codex (OpenAI); the review records are held in our reports store and available on request.
CounterProof is a small practice of Clavestra Capital Limited (Malta, C 113987) that reviews AI-generated code and AI-assisted review processes for organisations where a wrong answer is expensive. We disclose that we have a commercial interest in the adoption of the review method this essay describes.
Cite as: Soons, L. J. (2026). Earned, not inherited: authenticity and independence. CounterProof Research, 15 October 2026. Companion to Agentic Stemmatics, doi:10.5281/zenodo.22030516.