Litotes: the last frontier of LLMs
Sixth article in the series on rhetorical figures in the age of AI. After correctio, negation in LLMs, the tricolon, anaphora and chiasmus, litotes -- the art of saying less to mean more, and the figure where the gap between human and machine output runs deepest.
In July 1940, in London, de Gaulle drafts a poster that goes up in a thousand copies on British streets. On it, the line that will stay in every schoolbook:
"France has lost a battle! But France has not lost the war!"
A first useful paradox to lay out at the start: this sentence is not in the Appeal of 18 June, contrary to near-universal belief. It belongs to a poster drafted in late June, printed in early July, and put up on 3 and 4 August 1940 in London. Even the most famous French political litotes lodged itself in collective memory at the wrong address.
A second, more interesting paradox: at the moment de Gaulle writes this sentence, France has, objectively, lost the war. The army is in rout. The government is collaborating. No Allied force is in a position to retake the territory. The sentence therefore states a false assertion. And yet it is exact -- because it does not state a state of affairs, it opens a semantic space that the positive assertion would have closed. "France will win the war" would have sounded like a lie. "France has not lost the war" creates a possibility.
This is exactly what litotes does. It inverts an implication rather than describing a reality, and produces an effect that the direct statement would shut down by its very frontality.
This mechanism is also the one that generative AI, for precise architectural reasons, masters least well of all the classical figures.
A figure that affirms by negating
Litotes affirms by negating its softened opposite. The canonical form: "he is not bad" to mean "he is excellent." The surface attenuation produces the underlying amplification.
Three neighboring figures not to be confused:
- Euphemism softens. It attenuates on the surface and underneath: "he passed away" for "he died." André Comte-Sponville sums it up: euphemism masks reality, litotes brings it to light.
- Antiphrasis says the opposite without negation, leaning on the irony of context: "nice one!" after a blunder.
- English understatement forms a broader family, of which litotes is a special case. British culture made it into a complete code.
The etymology says it all. Litotes comes from the Greek litos -- "simple, meager, devoid of ornament." The figure looks poorer than it is, in order to force the listener to restore what is missing.
Corneille, act III, scene 4
The founding litotes of French literature fits in five words. Act III, scene 4 of Le Cid (1637). Chimène to Rodrigue, who has just killed her father in a duel:
"Go, I do not hate you."
She loves him. She cannot say so. Her father died by his hand, and the aristocratic honor of the seventeenth century forbids any direct avowal. Here litotes serves as the only linguistic means available, more than as ornament: it allows the expression of what culture forbids stating head-on. It is a performative act of civilization, opening the unspeakable while respecting the code.
This trait defines litotes in every era. Racine, a few decades later, in Andromaque: "I no longer hate anything but extreme inconstancy." Voltaire in Candide, after a question on absolute evil: "This answer was not entirely satisfactory." Always the same movement: the softened negation hollows out the space for a meaning that direct assertion would have saturated.
Understatement as an entire culture
English made litotes a national trait. Two documented military examples are enough to gauge the power -- and the danger -- of the figure.
Admiral Beatty, Battle of Jutland, 31 May 1916. After losing three battlecruisers in an hour under German fire, Beatty turns to his officer and remarks:
There seems to be something wrong with our bloody ships today.
Three battlecruisers, thousands of dead, "something wrong." The litotes holds because all the interlocutors share the code: the gravity needs no naming, the understatement makes it more present than a cry.
Battle of the Imjin, April 1951. The 29th British Brigade, in Korea, encircled by four Chinese divisions, signals to the American command:
Things are pretty sticky down there.
The American command, foreign to the British code, reads "sticky" as moderate concern and delays the reinforcements. The brigade suffers a catastrophe. Here the litotes failed because it requires a shared cultural decoding. Out of its register, understatement is read literally -- and kills.
Kate Fox, in Watching the English, theorized this rule: the British code works between Britons, sometimes between English speakers, rarely beyond. At Debretts, the conversion table is explicit: not bad = excellent; a bit of a nuisance = disaster; I'd be rather disappointed if... = serious threat.
Churchill, in the House of Commons on 4 June 1940, sums up the mastered use of the figure better than anyone:
We must be very careful not to assign to this deliverance the attributes of a victory. Wars are not won by evacuations. What has happened in France and Belgium is a colossal military disaster.
Everything is here, in a rare movement: Churchill refuses the complacent understatement ("not a victory") and simultaneously practices a hard understatement ("colossal military disaster" in place of "apocalypse"). A meta-litotes. The figure recognizes its own temptation and corrects it.
The mechanism: Grice and implicature
Why does this figure work cognitively? The answer comes from Gricean pragmatics.
H. P. Grice, in 1975, laid out the conversational maxims that govern cooperation between interlocutors: quantity (be as informative as required), quality (say only what you hold to be true), relation (be relevant), manner (be clear, brief, orderly). When an utterance apparently violates one of these maxims, the cooperative listener infers an implicature to restore coherence.
Litotes violates two maxims at once. To say "not bad" of a masterpiece is under-informative (violation of quantity) and obscure (violation of manner: the double construction negation+antonym is more complex than direct assertion). The listener thinks: if he chose this roundabout phrasing, it is not for economy -- it is that he means something else. He infers the inversion.
This inference is not neutral. It demands what Catherine Kerbrat-Orecchioni, in L'Implicite (1986), calls an encyclopedic competence: decoding litotes rests on a knowledge of the extralinguistic world, without which it is impossible to calibrate the gap between what is said and what is meant. A French reader and a British reader, faced with the sentence "this candidate is not bad," extract opposite meanings: the first reads "meh, average," the second reads "excellent, hire him."
Laurence Horn pushed the analysis further on double negations. His 1991 work on negated negative antonyms ("not unaware," "not unhappy," "not without merit") shows that the double negation is never perfectly equivalent to the corresponding positive assertion. To say "I am not unaware" opens a calibrated zone of uncertainty -- a reserve about the exact degree -- that "I know" shuts down. Litotes is a matter of precision rather than vagueness.
Why LLMs produce it poorly
It is precisely this pragmatic finesse that the architecture of large language models fails to reproduce.
In 2024, Scientific Reports published a study led by Cong et al. -- Manner implicatures in large language models -- testing the ability of LLMs to decode "not unaware"-type double negations. The panel covers RoBERTa, GPT-Neo, Llama-2, Falcon-7B, MPT-7B, Mistral-7B, Qwen1.5-MoE, Gemini-flash-1.5 and GPT-4o-mini. The result is clear-cut: most of the models tested achieve null or near-zero accuracy on this task. RoBERTa tops out at 21.1%. More recent models do not significantly improve the performance. Humans, for their part, decode these implicatures almost automatically.
Four technical reasons converge.
The token distribution favors direct assertion. During training, the sequence "excellent" preceded by the context "this proposal is" has a far higher probability than "not bad." The model, which maximizes likelihood, spontaneously prefers the most frequent phrasing.
RLHF accentuates this preference. Reinforcement Learning from Human Feedback optimizes for informational completeness: human annotators overwhelmingly prefer clarity to obliqueness. The model rewarded by this loop learns to over-explain. Litotes, by construction, leaves the reader to do the work. It runs against the training gradient.
Induction heads propagate surface patterns. In a previous article we saw how these circuits, identified by Anthropic in 2022, continue repeated patterns: when the model sees "not un-", the most probable head continues on the classic softening negation ("I am not unhappy") rather than on the inversive litotes ("I am extremely aware"). The pragmatic reversal requires a second-order operation that induction circuits do not encode.
The absence of grounded encyclopedic calibration is the deepest factor. A senior bid manager who writes "this proposal is not without difficulties" means "it is unwinnable." The LLM may know, in its weights, both the sentence and the situation -- but it has not internalized the codes of professional courtesy that link one to the other. Without this calibration, it reads the sentence literally and reproduces it likewise.
A synthetic formula: where anaphora directly exploits induction heads, where chiasmus demands a rare symmetry, litotes demands a complete pragmatic inversion that neither the architecture nor the training encourages. It is the figure where the gap between human and machine output is probably deepest, because it asks exactly what RLHF was designed to prevent: holding back information.
Litotes as a signature of authority
In consulting, the register of litotes is an almost infallible marker of seniority.
A junior consultant, faced with a difficult case, writes: "This case presents major risks requiring particular attention." A senior partner, faced with the same case, writes: "This case is not without difficulties." The second conveys more information to the informed reader, because understatement is a marker of mastery: only the one who has the case under control can afford to minimize it. Over-dramatizing betrays the fear of not being taken seriously -- a structural mark of the junior.
In bid management, the figure steers several key moments. In a chapter on understanding the need: "These specifications are not without echoes of the issues observed on [reference]" -- opening a comparison without claiming a complete parallel. On a weak point identified by the client: "This constraint is not without an answer" -- a signal of mastery without triumphalism. Faced with a major risk: "This point is not the easiest to address" reassures more than "we have perfect command of this," because the experienced reader knows that the one who claims to have everything under control has usually seen nothing.
In commercial negotiation, litotes inverts the pressure. "Your offer is not far from our expectations" forces the other side to raise it, where "your offer is good" freezes them. The figure is a tool of calibrated concession, leaving the interlocutor responsible for closing the gap.
In diplomacy, the code is shared among all professional actors, which lets litotes convey a precise signal without breaching courteous form. "The positions are not yet aligned" allows the negotiation to continue; "we are in disagreement" shuts it down. Same informational content, asymmetric value.
The pitfalls
Litotes has two structural flaws worth knowing.
The cliché litotes: certain formulas have so saturated the register that they no longer produce any effect. "He is not unaware," "this is not without interest," "we do not ignore" -- registered by the listener as set phrases, they short-circuit pragmatic decoding. LLMs produce a great many of them, precisely because their training frequency is high. A telling paradox: a language model generates dead litotes more readily than living ones.
AI over-correction: ask an LLM to "make a litotes" on any subject, and it almost always produces a heavy anti-litotes -- "it is not uninteresting to note that...," "it would not be incongruous to suggest that..." The model reproduces a syntactic pattern ("not un-") without the pragmatic pressure that makes the figure effective. A clear signal in a text suspected of automated origin: the mechanical presence of double negations disconnected from any gain in implicature is a stylistic AI signature as plain as the saturation of "Not X. It's Y." or the stacking of tricolons.
What remains for the human author
In the taxonomy of classical figures faced with large language models, litotes occupies a particular place. The tricolon, anaphora, and correctio are produced en masse, sometimes to the point of saturation that gives the machine away. Chiasmus already resists well -- it demands a symmetric inversion that induction heads do not favor. But it is litotes that marks the deepest frontier.
The figure demands three competencies that the modern Transformer architecture does not reproduce together: pragmatic competence (inferring an implicature from an apparent maxim violation), encyclopedic competence (calibrating the gap between what is said and reality by drawing on a knowledge of the world), and active restraint (choosing not to state what could be stated, against the incitement to proliferation imposed by RLHF).
In AI-assisted professional writing, the presence of authentic litotes -- living, calibrated, integrated into a discursive balance of power -- remains one of the most reliable markers of a human hand. For a bid manager, a consultant, a negotiator, it is a skill to cultivate rather than delegate. The machine can draft the explanatory paragraphs. The inversion that signs authority still has to be set down yourself.
Next article
In our next analysis, we will turn to aposiopesis -- the figure of interruption, the one that stops at the threshold of speech. "I could answer him, but..." Another figure where the unsaid works harder than the said, and one that tests the fundamental vocation of LLMs: producing the next token.
Main sources: Cong et al., "Manner implicatures in large language models," Scientific Reports 14:28907, 2024. Ruis et al., "The Goldilocks of Pragmatic Understanding," NeurIPS 2023. Kerbrat-Orecchioni, L'Implicite, A. Colin, 1986. Horn, "Duplex negatio affirmat," Chicago Linguistics Society, 1991. Fox, Watching the English, Hodder, 2004. Grice, Logic and Conversation, 1975.