De Suspensione

Translating a Latin sentence is not, in the first instance, a vocabulary problem. It is a suspensio problem. A noun appears before its role is fully determined. A verb arrives late, at the end of the clause, announcing the shape of everything that preceded it. Case endings and agreement oblige the reader to hold several possible parses alive, adjacent and undecided, until the morphology accumulates enough evidence to let resolution occur. The mental act this requires is neither parsing nor decoding. It is the voluntary deferral of closure: the willingness to keep thought open, suspended, in a condition of relational expectancy, until the form of the sentence arrives to discharge it. We call this practice suspensio: the inner act of composition in its receptive mode. It is one thing to read a sentence in which the work of relation has already been done by word order; it is another to perform that relational work oneself, inside, while the sentence unfolds.

The classical tradition already knew this practice, and trained it explicitly. Quintilian in the Institutio Oratoria I.1 lays out a curriculum in which the child's mind is drilled, morpheme by morpheme, on the recognition and deployment of grammatical relation; the discipline is not rote but attention. Cicero in De Oratore II distinguishes meditatio (the inward composition that precedes speech) from memoria (the cultivated storehouse that makes meditatio possible), a distinction that loses its meaning the moment composition is outsourced. Augustine, two centuries later, gives the phenomenology its deepest form: in Confessions XI he calls this holding-open the distentio animi, the stretching of the soul across the interval between what is remembered, what is present, and what is expected. All three are describing the same act under different aspects: the disciplined interiority that sentence-building, in a highly inflected language, makes unavoidable. Sister Miriam Joseph, in her twentieth-century recovery of the Trivium, names the three operations of the rational mind as simple apprehension, judgment, and discursive reasoning (Sister Miriam Joseph, 2002). Suspensio is not one of these operations but the connective tissue between them, the willingness to keep the first open until the second and third can be performed.

The modern empirical literature, approached from outside this tradition, arrives at the same object with weaker vocabulary. Vereeck et al., 2025 map the classical-language impact studies and conclude that the claim "Latin trains the mind" is real enough to be researchable but not yet well settled; Adema, 2019 finds the field scarce and often driven more by teacher intuition than by systematic evidence. The better-controlled studies find modest and specific effects, not generalized cognitive enhancement. Haag and Stern, 2000 report no meaningful difference in general intelligence between pupils who begin with Latin and those who begin with English; small advantages appear only on grammar-related tasks. Ortner et al., 2004 find no relevant effect on general reasoning after two years of Latin instruction. Bracke and Bradshaw, 2020, reviewing the wider literature, reach a similar middle position: positive findings exist but cluster around language-related outcomes rather than general cognition. Hauspie et al., 2024 report that prior Latin study predicts later achievement in higher education, particularly in linguistic subjects, though the effect is neither broad nor uniform. Masciantonio, 1977 surveys older educational projects claiming gains in English verbal skill among Latin students, but the methodology of those reports is by modern standards uneven.

The narrower picture is the one the argument needs. Classical languages do not, on the evidence, produce superior minds in general. What they do, according to the studies that survive replication, is cultivate specific forms of verbal analysis, delayed interpretation, and attention to structure: suspensio under its several names. Maharaj, 2020 treats translation as a practice of formation rather than as a technical skill, an interpretive overlay on the empirical finding but one that captures the phenomenology well. Suspensio is not a curiosity of Latin pedagogy. It is the common object of the three migrations traced in the sections that follow, the single inner practice that each successive externalization asks the speaker to perform less of.

What left the word had to reappear somewhere. That is the simplest summary of the first migration, which carries suspensio out of the inflected sentence and into the surrounding apparatus of fixed order, prepositions, and auxiliary verbs. The historical literature is firm on the facts: Latin had richer nominal case marking and freer word order than any of its Romance daughters, and over time much of that morphology was reduced or lost while fixed constituent order and functional structure took over more of the grammatical work (Ledgeway, 2010; Ledgeway, 2012). The literature is equally firm that this should not be framed as simple degeneration. What is less often said is that the typological trade-off has a cognitive correlate: the interior work that a Latin case ending demanded of its reader is work that an English positional constraint performs on the reader's behalf, before the interior has to hold it. Suspensio has migrated outward, into form.

Ledgeway, 2010 reads the shift from Latin to Romance as a move from a more synthetic system toward more analytic structures: fewer endings doing grammatical work, more reliance on separate words and on tighter order. Liu et al., 2025 quantify the same long pattern across Romance and find support for a trade-off between morphological richness and word-order freedom. Nijs et al., 2025, examining five Western European languages, find that declines in morphological complexity tend to precede increases in word-order rigidity: when endings stop carrying enough information, syntax has to become more disciplined. Figure 3 displays this schematic relationship, and Figure 2 makes the qualitative difference between the two regimes (morphological versus positional) concrete in a single minimal pair.

The social explanation for this drift is well established. Trudgill, 2011 argues that morphological complexity is sustained by small, tight-knit speech communities (what he calls "societies of intimates") and simplifies when communities grow, disperse, and admit adult learners at scale. McWhorter, 2007 provides the most-cited statement of the specific mechanism: sustained adult-L2 contact leaves measurable "analytic drift" even in languages that never creolized. Bentz and Winter, 2013 show cross-linguistically that languages with more L2 learners tend to lose nominal case. Dale and Lupyan, 2012 frame the same pattern under a "linguistic niche" hypothesis: large populations, adult learners, and frequent cross-group contact favour simplification of inflectional paradigms and greater transparency.

On this view, simplification is partly a democratic success: language becomes easier to learn across wider networks. The cost is not loss of meaning. The cost, more narrowly, is that less meaning is packed into the word itself. Rich morphology places more structure inside the utterance; analytic language spreads that structure across position, helper words, and context. Neither freely, the mind still interacts with structure, but with a different structure, one whose demands track different attentional habits.

Inflection matters, on the evidence, not because it is old, difficult, or noble, but because it externalizes less and demands more active internal tracking of relation. That is the moderate linguistic-relativity position, and it is the one this paper defends. The strong Whorfian form of the claim, that one language family makes its speakers globally smarter than another, is not supported by any current evidence, and was never the argument worth defending. Havelock, 1963 reminded the twentieth century that literate rationality is itself a specific cognitive achievement rather than the default state of the mind, and the same caution now applies to grammatical form: what a given grammar makes easy to think, or hard to ignore, is a contingent training, not an imprisonment.

Morrow, 1986 showed that grammatical morphemes help organize discourse processing by mapping conceptual distinctions into the flow of comprehension. Wolff and Holmes, 2011 review the modern linguistic-relativity literature and explicitly reject strong determinism, but find evidence that language can make certain distinctions hard to ignore and can augment some forms of thinking. Thierry, 2016 goes further in a neurolinguistic review, reporting that lexical and grammatical patterns can shape perception, event framing, semantic association, and aspects of executive function. Boroditsky, 2001 provides a widely cited demonstration outside the European tradition, showing that Mandarin and English speakers, whose languages conventionally spatialize time differently, differ reliably in some temporal-reasoning tasks. That particular finding has had a contested replication record, and is cited here as illustrative of a pattern rather than as itself decisive.

The evidence is strongest where it is narrowest. Samuel et al., 2019, in a systematic review of forty-three empirical studies on grammatical gender, find that relativistic effects are real but task- and context-dependent, and that strong claims about gender-shaped cognition often have equally viable alternative explanations. The same caution applies here: grammar does not imprison the mind, but it does tilt perception and conceptualization in measurable ways.

Inflection matters, in this moderate form, because it externalizes less and demands more active internal tracking of relation. Case marking is a built-in guide to agency and dependence. Aspect trains attention to the contour of action. Mood and voice keep distinctions alive that analytic prose lets fade into paraphrase. Tal and Arnon, 2022 show that redundant grammatical marking can aid learning: more marking is not always clutter; sometimes it gives the learner more handles on structure. Grammar is not just a delivery vehicle for thought; it is one of the recurring ways thought is disciplined. What cannot be said, on present evidence, is that the move from inflected to analytic systems straightforwardly caused cognitive decline.

The sharpest contemporary evidence on externalization is not in the history of Latin; it is in the recent psychological literature on cognitive offloading and generative artificial intelligence. Here the two remaining migrations arrive: first the externalization of content, then the externalization of production, and with them the closest empirical proxies for suspensio. The inner act itself remains an interpretive construct, but some of its likely residues and conditions (recall of self-produced sentences, felt ownership, cognitive offloading, and the shift from composing to adjudicating) can be studied in laboratories, surveys, and mixed-methods designs.

A clarification is owed up front, because the AI literature attracts careless versions of the argument from every direction. The claim is not that using AI makes people dumber, nor that delegation to tools is itself a failure of mind; tools have always extended the mind, and nothing in what follows denies that. The claim is narrower. In its stronger uses, AI assistance now performs an act that used to be a constitutive part of being literate: the holding of a draft open against one's own judgment, while form catches up with thought. That act does not survive long when nobody is performing it. Faculties not exercised attenuate; what is at stake is not intelligence but a practice.

Risko and Gilbert, 2016 define cognitive offloading as the use of the environment to reduce internal cognitive demand. The phenomenon is real: when people expect information to be available externally, they are less likely to remember the information itself and more likely to remember where to find it (Sparrow et al., 2011). The effect has had a mixed replication record (Storm et al., 2017), and a 2024 meta-analysis reports that "Google effects" on memory are real but moderated by perceived reliability of the external store, device type, and cognitive load (Gong and Yang, 2024). The cumulative picture is not one of a single dramatic effect; it is one of a distributed, domain-dependent shift in what is retained and what is offloaded. Skulmowski, 2023 synthesizes this literature as a model of digital externalization: offloading does not leave the mind empty, but it leaves the mind holding biological pointers and gist traces in place of detail, and the paper warns of an illusio cognitionis, an illusion of knowing in which easy access is mistaken for mastery. This is the second migration. Suspensio is touched here only at the edges: the inner act of composition does not, in itself, require remembering every quoted sentence, only performing the act of composing around it. What thins, in this regime, is the ecology in which suspensio takes place: the half-remembered phrase that rises to meet the new sentence, the resonance, the allusion.

The third migration is qualitatively different. Generative AI does not merely store; in its stronger uses it proposes or performs substantial stretches of composition. Stadler et al., 2024 report that large language models reduce mental effort in student scientific inquiry while compromising depth. Lee et al., 2025, surveying knowledge workers, find that AI changes critical thinking rather than eliminating it: the worker is shifted from composing toward adjudicating. Confidence in the AI is the strongest moderator of whether engagement rises or falls. That is to say: the very suspensio-shaped act of holding one's own draft open against one's own judgment is liable to be replaced by the more frictionless acceptance or rejection of an externally generated one. The empirical pattern here recalls Pieper, 1998, whose distinction between ratio (the discursive labor of composition and judgment) and intellectus (the receptive contemplation that discursive labor serves) maps directly onto the concern: AI-assisted work threatens to collapse ratio into the mere adjudication of generated outputs, leaving neither the labor nor what the labor is for. The most quantitatively robust evidence currently available is Gerlich, 2025, a mixed-methods study of 666 participants that finds a significant negative correlation between frequency of AI-tool use and critical-thinking scores, mediated by increased cognitive offloading, with the effect strongest among younger users. Figure 4 formalizes that mediation as a path diagram.

Kosmyna et al., 2025 provide the most vivid, though still preliminary, evidence. In an essay-writing task with EEG monitoring and n=54 at baseline (n=18 at follow-up), participants using ChatGPT showed weaker brain connectivity than brain-only writers, lower ownership of their texts, and striking difficulty quoting sentences they had just produced; 83% of LLM-assisted writers could not reproduce a sentence they had written minutes earlier. The paper is a preprint and its own authors resist generalization, but its image is consistent with what the larger Gerlich data and the Lee survey already indicate: when the tool helps too much, authorship can thin, and the text may exist with a weaker interior trace of having been composed.

The pattern across these four studies is worth noting in itself, because no one of them could carry the claim alone. Gerlich, 2025 is a large-N peer-reviewed mixed-methods survey of 666 participants: correlational, generalizable, but unable on its own to ground a causal inference. Kosmyna et al., 2025 is a small EEG-monitored experimental preprint of 54: invasive, vivid, and at the wrong end of every generalizability dimension. Lee et al., 2025 is a CHI-published self-report survey of knowledge workers: peer-reviewed and ecologically valid, but vulnerable to the usual self-report distortions. Stadler et al., 2024 is a controlled experimental paradigm in student scientific inquiry: peer-reviewed and causally clean, but narrow in domain. Each study's weakness is another study's strength. What gives the claim its evidential weight is not any single effect size but the fact that four independent literatures, with four different methodologies, converge on the same directional finding: under AI assistance, the internal trace of composition is attenuated.

Figure 6 is a heuristic rather than an estimate. It arranges the regimes on which the argument turns as stacked columns and asks what portion of each communicative act remains internal to the speaker. The values are illustrative, not measurements, and the figure should be read as a synopsis of claims argued elsewhere in the text rather than as an independent result.

Across the Latin-to-Romance simplification, the digital externalization of memory, and the offloading of sentence production to generative AI, the same inner practice (suspensio, the inner act of composition) is asked to bear less of the communicative act. Together the three movements are the Tres Migrationes: the migrations of form, of content, and of production.

They are not points on a single continuum, nor are they equally documented. The first is a redistribution of grammatical work; the second alters the mnemonic ecology in which composition takes place; the third can move part of composition itself outside the speaker. Their social meanings differ and their evidence bases differ. What they share is a family resemblance: more infrastructure outside the speaker, less obligatory work inside. That family resemblance is the central claim of this review. Figure 7 sets out the argumentative shape by which the claim is reached.

The honest form of the cultural worry, then, is not that Latin was lost, nor that AI is arriving, nor that any single migration is a catastrophe. It is that all three migrations are now co-occurring, and that the inner practice which remains after all three have moved outward is no longer obvious, no longer performed by default, and no longer visible to the person from whom it has migrated. That is a cultural claim, not a linguistic or a neurological one, and it is defended here on cultural grounds.

Figure 8 places the intellectual history of the worry on a single timeline, a genealogy of recurring anxieties about externalization rather than evidence for the present claim. What changes between one voicing and the next is which dimension of the cognitive act is being moved outside.

Four objections deserve explicit reply.

Simpler languages are a democratic gain, not a loss.

Partly true, and the paper grants it. Simplification of inflectional paradigms is closely tied to large speech communities, adult learners, and cross-group contact (Bentz and Winter, 2013; Dale and Lupyan, 2012; McWhorter, 2007; Trudgill, 2011); analytic structure makes a language easier to acquire past childhood. The paper's claim is not that the shift was bad (it often was not), but that it was a redistribution. What left the word reappeared elsewhere. The democratic gain is real; so is the migration of suspensio outward from morphology into position and context. Figure 3 is precisely that redistribution, not a decline.

Tools have always extended the mind.

True, and the argument is not against extension as such. What is distinctive about the present moment is specific: the extended system now drafts the sentences, not only stores or retrieves them. Three independent methodologies converge on the empirical point that this particular extension changes the internal trace of the act: Kosmyna's EEG, Stadler's experimental paradigm, and Gerlich's mixed-methods survey (Gerlich, 2025; Kosmyna et al., 2025; Stadler et al., 2024). The argument turns on which part of the mind this particular tool extends past, and on what remains to be performed when it has.

The offloading literature is less settled than the paper implies.

A fair objection. The original "Google effect" (Sparrow et al., 2011) has a mixed replication record (Storm et al., 2017); the 2024 meta-analysis reports the phenomenon as real but strongly moderated (Gong and Yang, 2024). The AI-effect literature is young: Kosmyna et al., 2025 is a preprint; Lee et al., 2025 is self-report; Stadler et al., 2024 is a single experimental paradigm. The strongest single data point is now Gerlich, 2025, and although it is peer-reviewed and large-n, a correlational survey cannot by itself ground causal claims. The paper's argument does not require any single study; it rests on the directional agreement of four independent literatures. The empirical picture is young; the long arc is not.

The thesis risks nostalgia.

The objection deserves an answer but not the defensive one. The paper does not claim that Latin was good and English is bad; it claims that suspensio is a practice, and that practices, in the sense MacIntyre, 1981 gives the word, are traditions of making that require communities to sustain. When no one performs a practice, the practice does not merely become harder to find. It ceases to exist as a practice, because the practice is the performing. What the paper defends is not classical pedagogy as such but the conditions under which the inner act of composition can still be trained. That those conditions have historically been classical ones is a contingent and, under the pressures of the three migrations, an increasingly important fact.

These four literatures land on a single result. Classical-language study defends itself, on the evidence, as a training of specific habits of verbal analysis and attention to structure, though not of generalized cognitive capacity (Bracke and Bradshaw, 2020; Haag and Stern, 2000; Hauspie et al., 2024; Vereeck et al., 2025). The Latin-to-Romance simplification is a redistribution of grammatical load from the inside of the word to the outside of the sentence, not a decay (Ledgeway, 2012; Liu et al., 2025; McWhorter, 2007; Nijs et al., 2025; Trudgill, 2011). The moderate linguistic-relativity position is that grammatical systems train what a speaker must notice, hold, and discriminate, without imprisoning thought (Morrow, 1986; Samuel et al., 2019; Thierry, 2016; Wolff and Holmes, 2011). And the contemporary evidence on AI-assisted composition shows, consistently across four independent methodologies, that under conditions of tool use we now have proxy evidence for an attenuation of the interior trace of composition (Gerlich, 2025; Kosmyna et al., 2025; Lee et al., 2025; Skulmowski, 2023). Read together, these four literatures converge on the same object. Each of the three migrations this paper has traced (of form, of content, of production) asks less of the same inner practice, and that practice, the one we have called suspensio, is the common object.

Suspensio leaves no artifact. It is legible only in its exercise; it produces no object one could point to outside the act of composing, holding, or deferring. This is why the Kosmyna finding is so unsettling. What was measured there was not suspensio itself, but one possible residue of it: whether a sentence produced minutes earlier had left enough of an interior trace to be recovered as one's own. For eighty-three percent of the LLM-assisted group, it had not (Kosmyna et al., 2025). The practice in question does not fail visibly. It simply leaves a weaker residue, and its attenuation is inferred from that weakness rather than directly observed.

Convenience is not neutral. Every gain in ease raises the question of what inner faculty is no longer being exercised, and that question does, at least partially, answer itself. Faculties not exercised attenuate. The faculty in question here is the one on which considered speech, considered judgment, and considered thought all depend: not case endings, not memory palaces, not even the sentence as such, but the willingness to hold thought open until form arrives to discharge it. A civilization that externalizes this practice across all three dimensions at once has not lost Latin. It risks losing the condition under which Latin was worth having.