The chiasmus: the figure AI cannot yet produce — and why that matters
Fifth article in the series on rhetorical figures in the age of AI. After correctio, negation in LLMs, the tricolon and anaphora, the chiasmus. A figure that inverts its terms — and that, for precisely this reason, resists algorithmic imitation.
Washington, January 20, 1961. John Fitzgerald Kennedy takes the oath of office. Thirty-five minutes into his inaugural address, he utters the sentence that will set the tone of his presidency and ring on across the decades:
"Ask not what your country can do for you — ask what you can do for your country."
Fifteen words. One inversion. An effect that thousands of linguistic analyses have tried to reproduce, that every rhetoric manual cites as a model, and that modern language models, to this day, cannot produce with the same force.
This difficulty for LLMs is no accident. It stems from the very nature of the chiasmus — a figure whose power rests on a semantic intent that statistical imitation does not supply. To understand why is to understand what remains, in the age of generative machines, a properly human signature.
The figure, as the Greeks named it
The word comes from the Greek letter Χ (chi) — which, crossing, traces exactly the movement of the figure. Two terms set down in one order, then taken up in reverse. A-B becomes B-A. The two lines of reading cross visually at the center, like the arms of the chi.
Aristotle, in the Rhetoric (III, 9), treats this figure as a form of inverted parallelism, particularly effective for periodic antitheses. The Rhetorica ad Herennium — an anonymous work from the first century BC, long attributed to Cicero — gives it its Latin name in book IV, 28, 39: commutatio, the exchange, the permutation. Quintilian, in the Institutio Oratoria (IX, 3, 85-86), takes up the term while explicitly warning against its ornamental use: a hollow commutatio, he cautions, is nothing but vain symmetry. The ancient rhetorical criterion is already there: a formal inversion with no semantic gain is a flaw, not a virtue.
Heinrich Lausberg, in his Handbook of Literary Rhetoric (1998, §§ 800-803), formalizes the modern distinction. The French rhetorical tradition carries the analysis forward in Henri Morier (Dictionnaire de poétique et de rhétorique, PUF 1961), in Georges Molinié (Dictionnaire de rhétorique, LGF 1992), and in Olivier Reboul (Introduction à la rhétorique, PUF 1991).
Chiasmus, antimetabole, antithesis, parallelism
A technical distinction is needed at the outset, because it lies at the root of many misidentifications.
The chiasmus is an inversion of order between two clauses. The AB / BA relation is syntactic, not necessarily lexical. "One must eat to live, not live to eat" (Molière, L'Avare, III, 5) is a chiasmus: the verb-infinitive relation is inverted, even though the repeated words change places.
The antimetabole is the strict chiasmus — with exact repetition of the same words in reverse order. AB exactly becomes BA exactly. Kennedy's 1961 line is a canonical antimetabole: country / you → you / country. So is Shakespeare's line in Macbeth (I, 1, 11): "Fair is foul, and foul is fair." The antimetabole is the strictest form, the most powerful, and the hardest to produce.
The antithesis opposes two terms of contrary meaning without inverting their order. Malraux, in his Funeral Oration for Jean Moulin (December 19, 1964): "Culture is not inherited, it is conquered." The sentence sets inherit against conquer, but it builds no syntactic AB/BA inversion. It is an antithesis, not a chiasmus. The distinction is commonly missed — and the error instantly gives away the reader who has not looked closely at the structure.
The parallelism, by contrast, repeats a structure in identical order. "To win without boasting, to lose without renouncing" (Pierre de Coubertin, attributed). Two members built on the same pattern, not inverted. Parallelism is the mechanical opposite of the chiasmus.
Anthony Paul, in Chiasmus and Culture (Wiseman & Paul eds., Berghahn Books, 2014), further distinguishes four sub-types of chiasmus by level — phonetic, syntactic, semantic, narrative — and their respective effects.
The great chiasmi of history
The figure runs through cultures and centuries with striking regularity, precisely because its power depends on no specific cultural convention.
The New Testament. Nils Lund, in Chiasmus in the New Testament (1942, repr. Hendrickson 1992), shows that several Gospel passages — along with certain psalms of the Old Testament — are structured globally as large-scale chiasmi, the architecture of the entire text tracing an X. Matthew 19:30 / 20:16 offers a local case: "The first shall be last, and the last shall be first." The inversion falls on the same words, in exactly reversed order. A perfect antimetabole.
Lao Tzu, Tao Te Ching, chapter 81 (Liou Kia-hway translation, Gallimard): "True words are not beautiful, beautiful words are not true." The chiasmus fits Taoist paradoxical thought perfectly. Form follows substance.
Sallust, Catiline 54: "esse quam videri" — to be rather than to seem. The formula, often misattributed to Cicero, inverts the poles of Roman aristocratic morality: what counts is being, not appearance. The chiasmus carries Stoic ethics in three words.
Shakespeare, Macbeth (I, 1, 11): "Fair is foul, and foul is fair." The three witches open the play on this antimetabole, which announces the moral inversion that will structure the entire tragedy. The chiasmus is not merely a verbal ornament — it is the dramatic argument condensed.
La Rochefoucauld, Maxims 149 (Truchet edition, Garnier 1967): "The refusal of praise is a desire to be praised twice." This is not a strict chiasmus, but a movement of thought that belongs to the same paradoxical family. The Maxims abound in these structures where inversion serves psychological revelation.
Kennedy, Inaugural Address, January 20, 1961, drafted with Ted Sorensen, who publicly acknowledged the antimetabole as the speech's signature figure. "Ask not what your country can do for you — ask what you can do for your country." The two pronouns you and country swap on either side of the verb. The inversion carries the moral pivot JFK means to install: from dependency toward civic commitment.
One observation runs through these examples. The chiasmus never works by its form alone. It works because the syntactic inversion carries a semantic inversion. First → last is not just permuting two words. It is overturning a hierarchy. Fair → foul is not inverting two adjectives. It is announcing that good and evil will blur together in the play to come.
Why the chiasmus leaves its mark
The figure produces, in the reader or listener, a specific cognitive effect that Gestalt psychologists would call a closure effect. The human mind, meeting a structure AB, instinctively expects either a continuation or a variant. It does not expect a strict inversion — because strict inversion is statistically rare in natural speech.
When it arrives, the surprise is immediate. Then the mind rebuilds the link between the two clauses, grasps that one is the mirror of the other, and retains the whole as a closed unit, shut upon itself. Cognitive neuroscience has not, to my knowledge, directly formalized this phenomenon for the chiasmus — so this hypothesis must be presented as such. But the kinship with the classical Gestalt principles described by Kurt Koffka in Principles of Gestalt Psychology (1935) is direct: a closed form is retained more firmly than an open one, a symmetry more firmly than an asymmetry, a finished pattern more firmly than an interrupted one.
The chiasmus concentrates all three. It is closed (two members bound together, no continuation expected), symmetrical (the inverted order traces a mirror axis), and finished (the inversion cannot be continued without redundancy). It is a rhetorical form that human memory takes up almost without effort.
The intellectual value of the chiasmus: correcting thought, overturning perspective
The chiasmus shares with correctio a functional kinship: both correct thought. Correctio does so by explicit negation — "it is not X, it is Y" — which posits an expected term only to reject it in favor of another. The chiasmus operates more discreetly, by syntactic reversal, without needing to declare that the first reading was wrong. It simply offers the inversion, and lets the reader's mind do the rest of the work. The correction is more elegant, more profound, and often more effective.
The mental operation it triggers goes beyond the mere correction of a phrasing. The chiasmus unsettles the angle of reading. It invites us to see the same phenomenon from the other side — from its cause rather than its consequence, from the effect rather than the decision, from the victim rather than the actor. When the inversion succeeds, the reader performs a conceptual reorganization of what they thought they already knew. The sentence brings them no additional information; it offers them a new way of holding the old — and that reorganization engages them more than an extra fact would have.
The expression "is it the dog that wags the tail, or the tail that wags the dog?" — proverbial in English as in French (the tail wagging the dog), popularized by Barry Levinson's 1997 film — is a textbook case. A canonical antimetabole: dog-tail / tail-dog. But the chiasmus does not merely invert two nouns. It signals an inversion of causality. It says, in eight words: you think A causes B; check that it is not B that causes A.
This capacity makes the chiasmus a critical tool of great power. In politics, it lets you turn an opponent's own framework back against them — "you protect the institution the better to betray it, you betray the institution the better to protect it." In philosophy or social theory, it serves to expose false causalities: you believe poverty produces crime; consider whether crime does not produce poverty. In consulting or business, it lets you signal to a client that they have the order of causes wrong, without contradicting them head-on, while demonstrating that you have seen their problem from an angle they had not explored.
It is this third application that deserves dwelling on. A consultant who opens a recommendation with a well-set chiasmus demonstrates, in a single sentence, two things at once. First, that they have understood the client's situation in depth — deep enough to have identified the useful inversion. Second, that they have enough independence of mind to question the obvious truths the client has probably not dared challenge themselves. The chiasmus signals that you have not come to confirm the ambient consensus, but that you are able to step outside it.
That independence sells. The client does not pay to hear what they already think. They pay to be shown what they do not see. And the rhetorical form that best condenses this service is precisely the chiasmus — because it inverts the conceptual order without having to explain that it is doing so.
It is for this reason that the finest aphorisms of the moralists (La Rochefoucauld, Chamfort, Cioran) are so often chiastic. These authors do not describe the world as one expects to see it. They turn it over. And their turning leaves a mark because it is syntactic before it is conceptual — the form inverts the reading before the meaning has finished being grasped.
Why AI struggles with the chiasmus
Here is where the story of this series of articles takes an unexpected turn. The chiasmus is one of the few classical figures that today's large language models reproduce poorly — or rather, that they produce in appearance but with a semantic quality far below their anaphoric or tricolonic output.
Three reasons combine.
First, the structure matches no privileged attention circuit. The induction heads identified by Elhage, Olsson and their colleagues at Anthropic (In-context Learning and Induction Heads, arXiv:2209.11895, 2022) favor the continuation of a detected pattern: when the model has seen [A][B] ... [A], it predicts [B]. The chiasmus demands exactly the opposite. Having seen [A][B], the model must produce [B][A] — a reversal, not a continuation. The attention circuit does not push in that direction; it pushes the opposite way.
Second, the chiasmus demands a strong semantic intent. Inverting two terms in a sentence does not automatically produce meaning. The inversion must carry a new meaning, often paradoxical, that did not exist in the direct phrasing. This operation presupposes conceptual understanding — the model must know that "the first become last" reveals a morality inverse to the natural hierarchy. The exercise belongs to conceptual production, not to pattern completion in token space. The Transformer architecture, trained by statistical prediction of the next token, does not naturally carry this form of intent.
Third, chiasmi are statistically rare in training corpora. Anaphora abounds — every political speech, every pedagogical text, every marketing post contains some. The chiasmus appears in classical literature, in aphorisms, in moments of high condensed thought. An LLM has seen tens of millions of anaphoras, probably only tens of thousands of chiasmi. The gap in data mass is reflected in the gap in produced skill.
The practical consequence is striking. Ask a 2026 LLM — Claude Opus, GPT-5, Gemini 3 — to produce an anaphora on a given theme: the output is immediate, fluent, often very elegant. Ask that same model to produce a chiasmus or an antimetabole: it generally offers an antithesis, a parallelism, or a flat inversion devoid of semantic tension. The formal figure is there. The conceptual charge, rarely.
Common false recognitions
Rigor about the chiasmus requires denouncing a few bad examples that circulate in school manuals and Wikipedia pages.
Pascal, Pensées (fragment 277 Brunschvicg / 423 Lafuma): "The heart has its reasons of which reason knows nothing." The sentence plays on polyptoton — the repetition of the same word reason in different grammatical forms, here with and without a conceptual capital. This is not a chiasmus. There is no syntactic AB/BA inversion.
Malraux, Funeral Oration for Jean Moulin (December 19, 1964): "Culture is not inherited, it is conquered." Pure binary antithesis. No inversion of order. This is not a chiasmus.
The frequent quotation "esse quam videri" attributed to Cicero in fact comes from Sallust (Catiline 54). The correct attribution restores the historical context and the original political reach.
These errors share a common root. They take for chiasmus any paradoxical sentence, any well-struck opposition, any superficial symmetry. The true chiasmus demands an inversion of the syntactic order of the terms. Without that inversion, the figure belongs to another family.
The chiasmus as a human signature
We reach the point that ties this series together. Correctio, the tricolon, anaphora — these three figures are massively produced by LLMs because they correspond to patterns that the Transformer architecture amplifies naturally. AI replicates them with ease, sometimes to the point of saturation.
The chiasmus escapes this mechanism. It demands an intent of meaning that statistical prediction does not supply, an attention architecture that would push against the expected pattern instead of prolonging it, a conceptual density that exceeds lexical imitation.
This is what makes the chiasmus, today, one of the last properly human signatures in professional writing. An author who produces a successful chiasmus — an antimetabole that carries a genuine reversal of meaning — demonstrates a cognitive mastery that current models do not reach. The discerning reader senses it instinctively, without always knowing why: they feel that the sentence comes from a thought that held the two poles together, that turned the idea over in its head, that searched for the exact form where the syntactic inversion carries the semantic inversion.
The point is not to pile up chiasmi to signal the author's humanity. That would be to fall back into saturation — the same error as with anaphora or correctio, but out of the opposite fear. The point is to produce a chiasmus when the thought itself is chiastic — when the idea reverses naturally, when the syntactic inversion serves to bring to light a real conceptual inversion. In those cases, and only in those cases, the figure reaches its full power.
Practical implications
For professional writing in the age of LLMs, the chiasmus opens a precise avenue.
To set a human text apart from AI output. Inserting a worked chiasmus into an important text — an executive summary, a bid's conclusion, an op-ed — signals to the expert reader that the author has held a conceptual thought beyond mere phrasing. A single good chiasmus is enough. Better none than a flat one.
To conclude an argument. The chiasmus works particularly well in final position, as the closure of a thought. Kennedy closed on "Ask not...", Shakespeare opened Macbeth on "Fair is foul", La Rochefoucauld ended his maxims on inversions of this kind. The closure the chiasmus produces cognitively accords with the structural closing of an argument.
To test your own text. If you produce a chiasmus on the first attempt, check three things: is the inversion strict (AB/BA or approximate), is the semantic inversion real (or merely ornamental), and would the sentence withstand the test of an attentive proofreader. If all three conditions hold, keep it.
To proofread AI. When proofreading a text produced with LLM assistance, spot the places where the AI has offered a paradoxical formula that resembles a chiasmus but is not one. Correcting them is often what makes the difference between raw output and a signed text.
What the chiasmus reminds us
The series of articles on rhetorical figures in the age of AI began with an observation: several classical figures are now overused by LLMs to the point of becoming their mocking signature. Correctio, anaphora, the tricolon have tipped into the domain of patterns recognizable as automatic.
The chiasmus holds the opposite position. It is rare in AI output, it succeeds with difficulty, it carries a human trace when it is well made. The machines will eventually learn, as they have learned the rest. A future model, trained specifically on a corpus of annotated antimetaboles, will be able to produce convincing chiasmi. That will not render the exercise pointless — it will displace it, as every advance in AI has displaced the zones of properly human mastery.
For now, the chiasmus remains a territory where the writer keeps an edge. On the condition of using it when the thought calls for it, and never because the figure is pretty.
Cicero wrote in De Oratore that the accomplished orator must master form without being enslaved to it. Two thousand years later, faced with statistical models that reproduce certain structures better than we do, the lesson sharpens: one must know how to produce what the machine reproduces poorly, and beware of what it produces too well.
The next article in the series will explore litotes — the figure that says less to express more ("it is not bad" to mean "it is excellent"). Another figure AI struggles to master, because it demands an understanding of the implicit and of cultural innuendo.
Principal sources
- Aristotle, Rhetoric, III, 9 (+ II, 23 secondarily).
- Rhetorica ad Herennium (anonymous, 1st c. BC), IV, 28, 39 — commutatio. Ed. Achard, Belles Lettres 1989.
- Quintilian, Institutio Oratoria, IX, 3, 85-86. Ed. Cousin, Belles Lettres.
- Lausberg, H. (1998). Handbook of Literary Rhetoric. Brill. §§ 800-803.
- Fontanier, P. (1821-1830, repr. 1977). Les Figures du discours. Flammarion, preface by Genette.
- Morier, H. (1961). Dictionnaire de poétique et de rhétorique. PUF.
- Molinié, G. (1992). Dictionnaire de rhétorique. LGF.
- Reboul, O. (1991). Introduction à la rhétorique. PUF.
- Paul, A. (2014). From Stasis to Ékstasis: Four Types of Chiasmus, in Wiseman & Paul (eds.), Chiasmus and Culture. Berghahn Books.
- Lund, N. (1942, repr. Hendrickson 1992). Chiasmus in the New Testament.
- Forsyth, M. (2013). The Elements of Eloquence. Icon Books.
- Koffka, K. (1935). Principles of Gestalt Psychology. Harcourt, Brace.
- Elhage, N. et al. (2021). A Mathematical Framework for Transformer Circuits. Anthropic.
- Olsson, C. et al. (2022). In-context Learning and Induction Heads. arXiv:2209.11895.
- La Rochefoucauld, F. (1665, ed. Truchet, Garnier 1967). Maximes.
- Sallust, De Catilinae coniuratione, chapter 54.
- Kennedy, J. F. (1961). Inaugural Address, January 20, 1961. JFK Library.
- Shakespeare, W., Macbeth, act I, scene 1, line 11.
- Bible, Gospel of Matthew, 19:30 and 20:16.
- Lao Tzu, Tao Te Ching, chapter 81 (trans. Liou Kia-hway, Gallimard).