The "not X, but Y" formula is not an AI tic. It's a semantic optimization.
Yes, the title of this article is itself a "not X, but Y." The choice is deliberate. Read it through before you cry tic.
For two years now, a particular mockery has taken hold in LinkedIn comments, writing workshops, and brand newsrooms: AI-generated texts can be spotted by the pattern "it's not X, it's Y" or, in its denser form, "not X, not Y, but Z." It has supposedly become the tell of the algorithmic ghostwriter, the 2025 equivalent of the "as an expert" that haunted LinkedIn pages in 2018.
The verdict is convenient. It is also incomplete.
This formula existed before AI. It carries a name in classical rhetoric. Its growth in American management literature predates LLMs by several decades. And when you go beneath the surface, you find a structural reason for its algorithmic amplification: negation is cognitively more effective than assertion at narrowing the field of possibilities -- exactly what a reader is trying to do when presented with a thesis.
So it is not a tic. It is a semantic optimization. And like every optimization, it has a cost: saturation.
The too-easy diagnosis
Before defending the formula, let us concede what is true. Many AI-generated texts do indeed saturate the pattern. Three, four, five occurrences in a single post. The reader checks out. The structure turns mechanical. The tone loses its variation. On the level of the pure finished product, the criticism holds: AI overuses it.
But "overusing a form" does not mean "using a defective form." A chef who adds too much salt does not prove that salt is bad. He proves that he failed to measure.
The problem is the dosage. The diagnosis should be: AI applies a powerful figure without calibrating its intensity to the context of the human reader. The diagnosis in circulation is poorer: AI invents an artificial pattern that betrays its nature. The second is false. The first is interesting.
Part 1 -- A rhetorical figure that has carried a name since Aristotle
The "not X, but Y" pattern was not born with GPT-3. It is called correctio in Latin, or epanorthosis in Greek -- literally "rectification." Quintilian treats it in the Institutio Oratoria (Book IX, chapter 3), in the first century AD, as one of the rectification figures most useful to argumentative discourse. Heinrich Lausberg, in his Handbook of Literary Rhetoric -- the English translation (Brill, 1998) of his 1960 Handbuch -- devotes sections §784 to §786 to it among the major figures of thought in the Western tradition.
[VERIFIED] Correctio = a rhetorical figure consisting in rejecting one term to replace it with another judged more apt. Sources: Quintilian Institutio Oratoria IX.3; Lausberg §§784-786; the "correctio" entry in Silva Rhetoricae (rhetoric.byu.edu), the reference work online for classical rhetoric.
Correctio is distinct from three neighbouring figures with which it is often confused:
- Antithesis merely opposes two terms without rejecting the first (life is short, art is long).
- Antimetabole reverses the order of words (ask not what your country can do for you, ask what you can do for your country).
- Epanalepsis repeats a single word at the beginning and end of a sentence.
Correctio, for its part, makes a precise move: it posits an expected term, rejects it, then installs a more apt one. It is an argumentative pivot. A miniature 180-degree turn. The mechanism is ancient, stable, studied.
It is not a novelty. It is a resurgence.
Part 2 -- The growth before AI: the American business corpus
The pattern became massively present in American management literature well before LLMs went mainstream. Simon Sinek, in Start With Why (Portfolio, 2009), uses it as the central backbone of his book and of his TED Talk How Great Leaders Inspire Action (2009, roughly 65 million views). The book's thesis is built entirely around a cascade of correctio: people don't buy what you do, they buy why you do it.
[CLAIM -- to be verified by N-gram study] The growth of the pattern in the American business corpus between 1980 and 2020 would deserve a dedicated linguistic analysis. Two methodological avenues:
- Google Books Ngram Viewer on the expressions
"it's not about X, it's about"and variants, on the English 2019 corpus. - Corpus analysis on the best-selling management books of the period: Drucker The Effective Executive (1967), Collins Good to Great (2001), Goleman Emotional Intelligence (1995), Sinek Start With Why (2009), Duhigg The Power of Habit (2012).
What can be asserted without a dedicated quantitative study is that correctio is documented as a cognitive figure of choice in the modern rhetorical literature. Max Atkinson, in Lend Me Your Ears (Oxford University Press, 2004), demonstrates on a television corpus that the contrast pair is one of the three most reliable applause triggers in British and American political speeches. Jay Heinrichs, in Thank You for Arguing (Crown, 2007, 2020 reissue), explicitly treats correctio as a tool of modern corporate communication.
The formula was not invented by AI. It was ingested by it.
Part 3 -- Why AI amplifies this figure
Two mechanisms overlap.
(a) Training on the corpus
Large language models are trained on billions of tokens drawn from the web, digitized books, press articles, and corporate content. The American management and thought-leadership corpus is massively represented within it. If correctio is a frequent figure there -- which the Sinek and Atkinson analyses suggest -- then an LLM will naturally reproduce it in output. It is a mirror effect of what it was fed.
(b) Alignment through RLHF reinforces the preference for discriminating forms
Ouyang et al. (2022), in the founding InstructGPT paper [Training language models to follow instructions with human feedback, arXiv:2203.02155], demonstrate that alignment through human preferences (RLHF) pushes models toward structured, explicit, discriminating output. Human annotators prefer, at equal content, formulations that take a side -- those that clearly state what a thesis is not before stating what it is. RLHF embeds this preference into the model.
Liang et al. (2024), in Monitoring AI-Modified Content at Scale (arXiv:2403.07183, Stanford), go further. Their study detects the GPT stylistic signature in a massive volume of academic reviews submitted to ICLR and NeurIPS. Among the stable markers: the over-representation of correctio-type turns and structured antithesis.
So AI does not merely reproduce. It amplifies, because human alignment pushes it toward the forms with the greatest measured power of persuasion -- and correctio is one of them.
Part 4 -- Negation narrows the field faster than assertion
There is a third, deeper reason why AI (and humans) favour correctio in an argumentative context: negation is a more powerful cognitive operator than assertion when it comes to narrowing a field of possibilities.
Take a concrete example. If I tell you:
The executive summary is a value proposition.
You understand the assertion, but your mental space stays wide. What is a value proposition? What does it stand against? The concept floats.
Now, if I tell you:
The executive summary is not a summary. It is a value proposition.
Something different happens. The negation immediately excludes the most probable interpretation ("summary" is the default reading of the word "summary"). The mental field narrows at once. Then the second proposition fills the cleared space with precision.
This is exactly what Laurence R. Horn describes in A Natural History of Negation (University of Chicago Press, 1989, CSLI reissue 2001), the reference work on the pragmatics of negation: marked negation carries asymmetric semantic information -- it does not merely state that X is false, it signals that X was the salient expectation and that it must be corrected.
Herbert Clark, in Linguistic Processes in Deductive Reasoning (Psychological Review, 1969), showed that processing a negation costs more cognitively than an equivalent assertion -- but that the discrimination it produces is stronger. The reader pays slightly more in cognition to gain far more in precision.
Grice, in Logic and Conversation (1975), supplies the theoretical foundation: the scalar implicature triggered by correctio signals to the reader that the author has thought about the contrast, has rejected the salient reading, and is committing to a precise formulation. It is a signal of argumentative rigour, not a tic.
Correctio does not spare the reader. It returns in precision what it costs in effort.
Part 5 -- So it is not a tic, it is a semantic optimization
Let us put the pieces back together.
We have a classical rhetorical figure, studied since Quintilian, documented in the treatises of Western rhetoric. We have a growth in American management literature from the 1980s to the 2020s -- predating ChatGPT. We have an RLHF alignment mechanism that pushes LLMs to favour its use because humans prefer it. We have a cognitive foundation explaining why negation narrows the field of possibilities faster than an assertion alone.
The conclusion is almost inevitable: the "not X, but Y" formula is not an accidental flaw in AI. It is a semantic optimization, drawn from a long rhetorical tradition, reinforced by human alignment, and structurally more effective than a plain assertion at discriminating a thesis.
The common mockery ("it's an AI tic") makes a poor diagnosis. The pattern does not signal a machine. It signals a powerful cognitive form. Its omnipresence in AI output reveals not a bug -- it reveals the importance it has always held in argumentative writing, an importance that a generation of readers discovered at the same moment they discovered AI.
Let us own the framing: what we call an "AI tic" is in reality a classical figure whose visibility AI has accelerated.
Part 6 -- The editorial stake: recognize, value, master
Recognizing the value of a figure is no licence to saturate it. A good writer measures. An uncalibrated AI writer saturates. That is the difference between a spice and seasoning, between music and noise.
For a human who writes -- with or without AI assistance -- the rule for using correctio comes down to three points.
One. Use correctio when the salient default reading must be explicitly rejected. If the reader is going to read "summary" by default upon seeing "executive summary," it must be corrected before building. If no default reading gets in the way, the figure is a cost without benefit.
Two. Limit it to one occurrence per argumentative unit (paragraph, section, post). Two or three occurrences in the same short text signal that the author -- or his AI -- is applying the figure by reflex, without structural reason. Syntactic variation is what distinguishes prose written by an attentive human from token output generated in series.
Three. Own the figure when it is right, even if it sounds like AI. A good argument does not turn bad because it resembles AI. An effective title does not become suspect because it uses a correctio. Refusing the figure out of fear of the algorithmic stigma is a form of over-correction that sacrifices the precision of the message to the cosmetics of style.
The real criterion is not does it sound like AI -- it is does this figure, here, discriminate better than a plain assertion. If yes, keep it. If no, cut it. The professional reader, contrary to common belief, is no more sensitive to form than to substance. He is sensitive to both, coupled: a heavy form betrays soft thinking, an elegant form betrays clear thinking.
For the discerning reader, form is substance in motion.
Conclusion
The "not X, but Y" formula is a revealer. It pits two rhetorical cultures against each other: that of practitioners who have seen it as a powerful cognitive figure for two millennia, and that of a segment of the contemporary public who perceive it as an AI signature because they rediscovered it through ChatGPT.
Both are half right. The figure is ancient. Its amplification is recent. Its effectiveness is structural. Its saturation is a failure of dosage. The stake, then, is neither to ban it nor to impose it. The stake is to master it.
For practitioners of pre-sales, of bid management, of consulting -- all those professions where one must convince an evaluator in few words -- correctio remains a first-rate tool. Provided it is used where the reader has an implicit expectation to rectify. And removed everywhere else.
It is not a tic. It is a semantic optimization. But an optimization without limit becomes a constraint. And an unmastered constraint becomes a tic.
That is where, and nowhere else, the difference plays out between an AI that assists and an AI that smothers: in the human's capacity to reclaim control over the forms it produces in bursts.
A future article will return in detail to the more technical question: why negation is computationally more efficient for an LLM than the equivalent assertion, and what that implies for the design of an explicit cognitive system.
Principal sources
- Quintilian, Institutio Oratoria, Book IX, chapter 3 (classical treatment of correctio).
- Heinrich Lausberg, Handbook of Literary Rhetoric, Brill, 1998 (Eng. trans.), §§784-786.
- Silva Rhetoricae (rhetoric.byu.edu), "correctio" entry.
- Simon Sinek, Start With Why, Portfolio, 2009 + TED Talk How Great Leaders Inspire Action, 2009.
- Max Atkinson, Lend Me Your Ears: All You Need to Know About Making Speeches and Presentations, Oxford University Press, 2004.
- Jay Heinrichs, Thank You for Arguing, Crown, 2007 (2020 reissue).
- Ouyang et al., Training language models to follow instructions with human feedback, arXiv:2203.02155, 2022 (InstructGPT / RLHF).
- Liang et al., Monitoring AI-Modified Content at Scale, arXiv:2403.07183, Stanford, 2024.
- Laurence R. Horn, A Natural History of Negation, University of Chicago Press, 1989, CSLI reissue 2001.
- Herbert H. Clark, Linguistic Processes in Deductive Reasoning, Psychological Review, 1969.
- H. P. Grice, Logic and Conversation, 1975 (scalar implicature).