The microphone problem
Suppose you ask someone, “What were you thinking?” The question quietly bundles three events. Something was represented or evaluated. That state became available to a response process. The response process encoded it in words. Ordinary conversation compresses all three into a single act because the links usually work. Cognitive science cannot afford that convenience.
A person may distinguish two quantities without possessing number words. A jay may cache food for a later shortage without describing tomorrow. A designer may rotate a shape in the mind and only afterwards narrate the path. A patient may infer a rule but be unable to construct the sentence that would display it. In each case, the public signal is poorer than the candidate thought.
The estuary of thought
Practitioner synthesisSeveral working media can converge on a choice; language is one channel into coordination and report, not the only upstream river.
The severed microphone
Two engineers inspect the same unfamiliar mechanism. Both select the broken gear from six candidates. One explains the diagnosis. The other can only point because the microphone and speech interface have failed. Now change one variable: restore the second engineer’s pointing while leaving speech unavailable. If accurate choices return, what disappeared was not necessarily diagnosis. It was a route from diagnosis to sentence.
The experiment exposes the hidden assumption in many assessments: that a response channel is transparent to the competence being measured. It rarely is. A task measures the conjunction of representation, access, decision and response.
Let task accuracy be written as a crude product:
“Competence” is the ability of interest, such as comparing quantities. “Access” is whether the task instructions and stored state reach that ability. “Response fidelity” is whether the person can express the result through the allowed action. The equation is not a psychological law. It is a reminder that a low observed score can be produced by several broken links. Removing speech can lower access or response while leaving the comparison itself intact.
The product form also explains why apparently easy controls matter. If response fidelity falls from 0.95 to 0.50, an unchanged competence can look almost halved. If the task is taught through a long sentence, access may fail before reasoning begins. If a participant must name the answer, a naming impairment becomes part of the score. Conversely, a lucky perceptual cue can raise observed performance without the intended representation. The measurement problem is therefore symmetric: silence can hide competence, and a fluent answer can counterfeit it.
A defensible test varies one link while holding the others steady. Replace naming with pointing. Teach the interface by demonstration. Compare a language-targeting disruption with a motor and timing control. Then ask whether the inferred relation survives on new material. These changes turn the triangle from a philosophical picture into an intervention plan.
The language–thought–report triangle
Practitioner synthesisEach edge is a separate empirical question. Fluency can fail at one edge while non-verbal performance reveals another.
A lesion is an intervention, but not a perfect one
Aphasia is especially informative because brain injury can damage language while sparing other capacities. In one classic case, a man with profound agrammatism succeeded on causal-reasoning and theory-of-mind tasks despite severely limited propositional language.[2] In another study, three men with severe grammatical impairments retained several forms of exact calculation, including expressions whose bracket structure resembles recursion.[3] A later study found that people with aphasia were strongly impaired when matching sentences to pictures yet performed close to controls when judging whether pictured events were plausible.[4]
These are more than anecdotes of hidden intelligence. They create a causal contrast. If fluent syntax were the current medium of every structured thought, severe damage to the language system should collapse structured performance across domains. The observed dissociations reject that strong prediction for the tested tasks.
What a language intervention can and cannot identify
Causal modelA selective language intervention is decisive only when task access, general effort and response demands are independently controlled.
flowchart LR A["The microphone problem"] --> B["Four windows on thought"] B --> C["What words add"] C --> D["Testing the gap"] D -. "re-test" .-> A classDef first fill:#0a1424,stroke:#007D68,stroke-width:2px,color:#e6f3ff classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff class A first class B,C,D step
Four windows on thought
No single population can answer the language question. Infants have little or no fluent speech but limited response channels. Animals provide natural non-linguistic contrasts but invite anthropomorphic interpretation. Imagery reveals alternative working formats yet varies sharply across people. Aphasia offers a within-species intervention but only after development. The four windows become useful when their weaknesses do not align.
Four windows, not a ladder
Evidence mapEach window removes a different link between fluent language and performance. Convergence matters because no window is transparent.
Infants: quantity before counting words
Six-month-old infants have been reported to distinguish displays of eight from sixteen items while failing the finer eight-to-twelve contrast when continuous visual properties were controlled.[5] The ratio pattern is characteristic of approximate magnitude representation. It is not exact counting, and it does not show that a baby entertains the sentence “sixteen is greater than eight”. It does show sensitivity to quantity before number-word mastery can explain the performance.
The 8:16 and 8:12 contrast
Imagine two rapidly changing dot clouds. Shape, total area and brightness are balanced across trials. The number alternates between eight and sixteen until the display becomes familiar, then the infant sees a new sequence. Looking recovers. Repeat with eight and twelve, and recovery may not appear. The smallest mechanistic interpretation is a noisy magnitude code whose distributions separate at a 2:1 ratio but overlap at 3:2.
Remove the assumption that looking time directly reads a concept, and the conclusion narrows: the infant detected a change correlated with numerosity under the controls used. That is still a non-verbal discrimination; it is not yet a theory of number.
Object-permanence research offers a similar lesson. Five-month-old infants looked differently at possible and apparently impossible movements of a screen around a hidden object.[6] Such violation-of-expectation results have been interpreted as early physical knowledge, but the method’s inferential assumptions remain contested. Novelty, familiarity, attention and event complexity can also affect looking.[7] Infant evidence is strongest when several response measures and tightly matched controls converge.
The silent nursery
A baby watches a ball roll behind a screen. In one trial the screen lifts and the ball is gone. Now change only the available response. First measure looking; later allow reaching; later still allow a spoken prediction. Suppose the three measures mature at different times. Did the physical expectation appear three times, or did one expectation gain three progressively easier output channels?
The case warns against locating the birth of thought at the birth of a convenient report. It also warns against the opposite error: an early looking difference does not automatically contain the full adult concept.
Animals: future-directed behaviour without narration
Western scrub-jays cached food in locations where they had learned they would lack breakfast the next morning, and stored particular foods where that food would be unavailable.[8] Ravens have selected tools or barter tokens for later use across delays, including contexts other than immediate food caching.[9] Rhesus monkeys have also used an uncertainty response more often when memory was weak, a pattern compatible with monitoring whether a memory is available.[10]
These behaviours require information about an absent state to alter a present choice. That is a useful operational meaning of planning or uncertainty monitoring. It does not prove a human-like autobiographical future, a private sentence, or reflective self-awareness. Associative histories can sometimes mimic planning, and later work has shown that specific caching paradigms do not always discriminate future-state models from learned associations.[11]
The vocabulary transplant
Give a jay a hundred discriminable call symbols, each consistently paired with a location, food and delay. Then test the original caching task. If the bird improves, vocabulary may compress or cue the plan. If planning was already present, improvement would show facilitation, not origin. Reverse the intervention: prevent all calls while leaving perception and action untouched. If caching remains selective, vocal symbols were not constitutive of that performance.
The thought experiment changes one causal property, access to a symbolic rehearsal channel, while preserving the ecological decision. It also exposes the real empirical difficulty: calls, attention, stress and social context cannot be changed independently in a living animal.
Imagery: thought can be non-verbal without being pictorial
In classic mental-rotation experiments, response time increased with the angular difference between two three-dimensional shapes, as though participants were traversing a transformation rather than retrieving a verbal label.[12] Yet visual imagery is not a universal requirement for spatial cognition. People reporting aphantasia can show reduced sensory detail in memory while preserving aspects of spatial placement. In one drawing study, participants with aphantasia recalled fewer objects and less colour, but their spatial accuracy was comparable and they made fewer false additions.[13] Physiological work has also found that imagined brightness modulates the pupil in typical imagers but not in a small aphantasic group, despite ordinary perceptual pupil responses.[14]
The important dissociation is not “words versus pictures”. It is among several formats: sensory simulation, spatial relations, motor preparation, magnitude, affective valuation and symbolic structure. Imagery is one non-verbal working medium, not a synonym for non-verbal thought.
The black-screen navigator
Two walkers learn the same maze. One later experiences a vivid overhead image. The other reports no image and instead remembers junction relations: left after the narrow passage, two turns before the fountain. Remove spoken self-instruction from both. If each still reaches the exit, the successful state need be neither a sentence nor a picture. It may be a relational policy that controls action.
Now rotate the maze. The vivid imager may preserve landmark geometry; the relational navigator may preserve turn order. Same outcome, different error signature. A good experiment identifies the signature rather than forcing both minds into one declared medium.
Aphasia: structured performance behind broken grammar
The recent logical-reasoning study makes the dissociation unusually visible. The two participants with profound aphasia solved 19 of 25 and 39 of 40 number-list induction problems. On a non-verbal matrix task they solved 25 of 30 and 26 of 30 items, scores above age-matched normative means in the reported analysis.[1] Their performance does not show that reasoning is a single language-free faculty. Induction and deduction may themselves use different systems. It does show that severe grammatical impairment did not erase the ability to discover and apply abstract transformations in these tasks.
Who pushed whom?
Show two short animations. In the first, a red triangle moves into a blue circle and the circle rolls away. In the second, the circle moves away before contact. Ask the participant to place a green token beside the event that could have been caused by the triangle. The response can reveal sensitivity to agent, patient, contact and temporal order without requiring the sentence “the triangle pushed the circle”.
Now add a verbal version: “Which shape made the other move?” A participant with aphasia may fail that question because thematic-role assignment in the sentence is impaired, even while selecting the causally coherent animation. The converse is also possible. A participant may learn a verbal template and answer correctly without tracking contact. The revealing intervention swaps the response and instruction channels while preserving the event relation.
This example illustrates why event semantics are not one thing. Natural-language comprehension, visual plausibility, causal prediction and motor response can dissociate. A good study states which relation each task preserves rather than calling all of them “understanding”.
A rule can be carried without a sentence
Illustrative task, adapted in form from rule-induction paradigmsThe participant can test a transformation through examples and gestures even when a fluent verbal explanation is unavailable.
[6, 7, 3, 6]. The rule can guide choice before it is phrased.| Window | What changes | What can be inferred | Main rival explanation |
|---|---|---|---|
| Infants | Fluent speech has not developed | Some discrimination, expectation or choice precedes verbal report | Looking or motor measure does not uniquely identify the represented content |
| Animals | Human natural language is absent | Present behaviour can be controlled by absent quantities, objects or future conditions | Associative policies can mimic richer planning or metacognition |
| Imagery variation | Sensory simulation differs across people | Similar outcomes can be produced through different working formats | Self-report groups may differ in strategy, memory or interpretation |
| Aphasia | Language processing is selectively impaired after development | Tested non-linguistic capacities can remain available despite severe language loss | Earlier language shaped the surviving capacity; lesions and compensation vary |
What words add
Rejecting the claim that all thought is natural language does not make language a decorative output layer. Words alter what can be held, revisited, compared and shared. They turn a fleeting distinction into an addressable object. They let one person install a task in another person’s mind. They preserve commitments across time and coordinate many specialised systems around the same rule.
The right distinction is between constitution and contribution. A constitutive medium is part of what makes the target operation the operation it is. A contributing resource improves speed, accuracy, stability or coordination while the core operation can still occur without it. The same resource can occupy different roles in different tasks.
Compression, addressing and control
A label such as “left”, “prime”, or “before” compresses a recurring relation. Inner speech can keep that label active, call it back after interruption, or bind it to another relation. Public language goes further: it lets several minds reuse the same compressed handle. Language changes cognition most reliably by making representations easier to address and coordinate.
Number offers a clean example. Approximate quantity discrimination exists without exact number words, but exact arithmetic over large quantities depends heavily on culturally learned symbols and procedures. Research with Munduruku participants, whose number lexicon was limited relative to industrialised counting systems, found competent approximate arithmetic alongside sharp limits on exact calculation for larger values.[15] Language and notation do not create magnitude sensitivity from nothing. They create stable, exact, recursively reusable representations.
Colour categorisation shows a smaller online contribution. Russian distinguishes lighter and darker blues with basic terms where English uses “blue”. Russian speakers were faster at discriminating colours that crossed that lexical boundary, and the advantage disappeared under verbal interference but not spatial interference.[16] The result does not mean speakers cannot see an unlabelled colour. It suggests that an available name can sharpen or accelerate a decision near a category boundary.
Labels are especially useful when a continuum must be cut for a repeated decision. Once a boundary has a name, it can be rehearsed, taught and compared with yesterday’s judgement. That makes language a control surface over perception. The surface can bias attention and memory without redrawing the underlying sensory space from scratch. Removing the label may therefore slow a borderline choice while leaving coarse discrimination intact.
The same mechanism explains why a linguistic contribution can increase with delay. A fleeting visual distinction may decay, while a compact phrase remains available for comparison. External notation extends the effect further. A written number, diagram label or checklist can preserve exact structure after sensory detail has vanished. In those cases, the causal unit is not only inner speech. It is a person coupled to a durable symbolic environment.
A systematic review of verbal-interference studies found no single answer across cognition. Concurrent verbal tasks often affect memory, categorisation, self-cueing and some forms of arithmetic, while evidence is weaker or more variable for visual detail, cross-modal integration and theory-of-mind tasks.[17] The pattern fits a coordination account better than an all-purpose language engine.
The scaffold that disappears
Two stone arches now stand without support. One was built with timber scaffolding; the other was assembled by a temporary mound of earth. Remove every trace of construction. Their present ability to bear weight does not reveal which scaffold was necessary while they were built.
Adult aphasia creates the same identification problem. A reasoning routine that now runs without language may have been organised through years of verbal instruction. Conversely, a skill taught in words may always have depended on a non-verbal model beneath the instruction. Online necessity and developmental necessity are different causal claims.
A five-rung language-dependence ladder
“Language-dependent” is too coarse to guide an experiment. The following ladder separates five places where language can enter. It is not a scale of intelligence. It is an audit of the link that failed.
The language-dependence ladder
New decision frameworkMove upward only when the lower explanation has been excluded. “Constitutive” is a demanding causal candidate, not a label for any verbal-interference cost.
Report
The thought guides a non-verbal choice, but spoken or written expression fails.
Access
Verbal instructions, labels or memory cues are needed to enter the task, not necessarily to solve it.
Optional strategy
Inner speech improves speed or confidence, while performance remains well above chance without it.
Coordination
Language binds several domains, time steps or rules; a selective disruption produces a specific cost.
Constitutive candidate
Targeted language disruption collapses the operation after access, load, motor and response controls pass.
The top rung is intentionally named a candidate. To reach it, a study needs non-verbal task familiarisation, non-verbal response, targeted language interference, a matched non-linguistic secondary task, enough trials to estimate chance, and transfer to new examples. Even then, the result supports constitution only for the defined operation and time window.
The subtraction that makes the mechanism visible
Consider four accuracies. Let B be baseline non-verbal accuracy. Let V be accuracy when a verbal answer is required. Let L be accuracy during language interference, such as articulatory suppression. Let M be accuracy during a matched non-linguistic load with similar timing and motor demand.
One fall in accuracy, two competing causes
Suppose B = 0.86, V = 0.82, L = 0.68 and M = 0.80. The report gap is 0.04. Requiring words barely changes accuracy. The general load cost is 0.06, so performing any secondary task causes a modest loss. The specific language cost is 0.12, because the matched-load condition exceeds language interference by twelve percentage points.
Now compare two interpretations. The first says language constitutes the transformation. It predicts that targeted disruption should drive performance towards the 0.25 chance level. The second says language coordinates rule maintenance. It predicts a selective cost with substantial residual competence. Observed accuracy of 0.68, plus transfer at 0.74, fits the second prediction better.
Change one number: let language-interference accuracy fall to 0.29 while matched-load accuracy stays at 0.80. The specific cost becomes 0.51 and performance approaches chance. That is a constitutive candidate. It is still not proof. Articulation may disrupt sequencing, breathing or executive control more strongly than the comparison task. The claim rises only if those rival paths are measured and rejected.
One score drop, three different stories
Illustrative synthetic valuesThe matched-load bar is the crucial comparison. Baseline alone cannot tell whether a language manipulation is specific.
stateDiagram-v2 [*] --> Claim state "The estuary of thought" as Claim state "The severed microphone" as Test state "The language–thought–report<br/>triangle" as Revision Claim --> Test: expose mechanism Test --> Decision: evidence survives Test --> Revision: rival survives Revision --> Claim: narrow or rebuild Decision --> [*]
Testing the gap
A useful non-verbal battery does more than replace words with pictures. It redesigns the entire chain from instruction to evidence. Otherwise language leaks back through a practice explanation, a labelled button, a counting strategy, or the demand to justify the answer.
The battery must let the target competence control an action without first passing through a sentence. That means demonstrative training, balanced perceptual stimuli, response modes such as touch or arrangement, matched secondary tasks, trial-level records and transfer items that defeat memorised cue-response pairs.
The six-task core
| Task | Candidate representation | Non-verbal response | Critical transfer | Primary failure risk |
|---|---|---|---|---|
| Quantity | Approximate magnitude or object-file capacity | Choose the larger set | New ratios, modalities and layouts | Area, density or duration cues |
| Spatial transformation | Rotation, topology or action policy | Select a matching shape | Novel angles and mirrored lures | Feature matching instead of transformation |
| Event plausibility | Agent, patient, contact and causal structure | Accept or reject an animation | New objects with preserved relations | Visual familiarity or salience |
| Rule induction | Composable transformation | Construct or select an output | Longer lists and unseen values | Surface pattern memorisation |
| Future choice | Delayed state and present action | Select or store a future-use item | Changed present desire | Learned association to location |
| Uncertainty | Strength or availability of evidence | Opt out for a smaller certain reward | Novel stimuli and calibrated difficulty | Simple reinforcement history |
The non-verbal battery as an evidence instrument
Implementation architectureThe task engine sits inside an anti-leakage boundary; evidence leaves as typed trial records rather than post-hoc impressions.
Worked protocol: non-verbal rule induction
From demonstration to causal contrast
- Interface learning: demonstrate two obvious transformations, such as sorting dots by size, until the participant can complete new demonstrations. No spoken instruction is required.
- Baseline: present 40 input-output examples from four rule families. The participant constructs the output by dragging tokens.
- Language perturbation: repeat with continuous syllable articulation or another validated verbal-interference task.
- Matched control: repeat with a paced finger sequence matched for timing, motor activity and error monitoring.
- Transfer: introduce longer lists, unseen token values and lures that share surface features but violate the inferred transformation.
- Readback: store the generated item, random seed, condition, action trace, latency and final choice before examining group effects.
A positive specific-language cost permits the claim that online linguistic resources contributed under those conditions. A near-chance collapse, after all controls pass, makes constitution a serious candidate. A null effect means the study did not detect a specific online language role at the tested power and difficulty. It does not show that language never helped development or never supports another strategy.
Language-dependence audit
Change the observed accuracies. The instrument separates report loss, general load and specific language cost before assigning a provisional rung.
Provisional assessment
B − V0.04
M − L0.12
B − M0.06
T − C0.49
Classification: Rung 3, language-facilitated or coordination-dependent
The task remains above chance and transfers, while language interference produces a larger cost than the matched load. This permits a specific online contribution claim, not constitution.
Assumptions: conditions have comparable reliability; the non-verbal interface is learned; secondary tasks are matched for timing and motor demand; the chance model is correct; enough trials exist to distinguish a stable effect from noise.
The lab exposes logic that prose hides. It asks whether a participant first demonstrated the competence, whether a language cost exceeds a general-load cost, and whether the inferred rule transfers. Its thresholds are illustrative. Real studies need uncertainty intervals, preregistered exclusion rules and hierarchical models where repeated trials and participants vary.
Accuracy also hides speed and strategy. A participant may preserve correct choices but take twice as long, suggesting that language supplied an efficient route rather than the only route. Another may maintain speed while confidence collapses. A third may switch from a verbal rehearsal strategy to spatial grouping after several blocks. Trial-level latency, action traces and optional confidence responses can reveal those transitions. They should be analysed as separate outcomes, not blended into a single intelligence score.
For small or rare populations, the design should privilege within-person contrasts and many well-controlled trials over a large catalogue of shallow tasks. Single-case statistics can compare an individual with an appropriate normative distribution, but the article’s classification still needs replication across items and sessions. A spectacular score on one puzzle is evidence for that puzzle, not a licence to generalise across thought.
type AuditInput = {
baseline: Probability
verbalResponse: Probability
languageInterference: Probability
matchedLoad: Probability
transfer: Probability
chance: Probability
threshold: Probability
}
type AuditResult = {
reportGap: number
specificLanguageCost: number
generalLoadCost: number
transferMargin: number
rung: 0 | 1 | 2 | 3 | 4 | "inconclusive"
permittedClaim: string
}
invariant 0 <= every_probability <= 1
invariant matched_load_has_comparable_motor_and_timing_demand
invariant response_channel_does_not_require_fluent_language
invariant transfer_items_defeat_surface_memorisation
function audit(x: AuditInput): AuditResult {
const R = x.baseline - x.verbalResponse
const S = x.matchedLoad - x.languageInterference
const G = x.baseline - x.matchedLoad
const T = x.transfer - x.chance
if (x.baseline < x.chance + x.threshold || T < x.threshold)
return inconclusive("competence or transfer not established")
if (S >= x.threshold &&
x.languageInterference <= x.chance + x.threshold / 2)
return rung4("constitutive candidate for this task and time window")
if (S >= x.threshold)
return rung3("specific facilitating or coordinating contribution")
if (R >= x.threshold)
return rung0("report bottleneck with preserved non-verbal competence")
return rung2("no specific dependence detected; optional strategy remains possible")
}The pseudocode makes prohibited inferences explicit. The output is a task-bounded causal classification, not a verdict on intellect, consciousness or the format of every thought.
Implementation notes and trial record schema
Each trial should record the exact stimulus generator version, random seed, transformation family, condition, response option order, participant action trace, latency, confidence or opt-out choice, and whether the item was a trained or transfer case. Store adaptations made by an examiner as events rather than silently changing the protocol.
{
"$schema": "https://json-schema.org/draft/2020-12/schema",
"title": "NonVerbalTaskTrial",
"type": "object",
"required": [
"trialId", "stimulusHash", "generatorVersion", "condition",
"responseMode", "response", "correct", "latencyMs", "transfer"
],
"properties": {
"trialId": { "type": "string" },
"stimulusHash": { "type": "string" },
"generatorVersion": { "type": "string" },
"condition": {
"enum": ["baseline", "verbal_response", "language_interference", "matched_load"]
},
"responseMode": { "enum": ["touch", "gaze", "arrange", "gesture", "opt_out"] },
"response": {},
"correct": { "type": "boolean" },
"latencyMs": { "type": "integer", "minimum": 0 },
"transfer": { "type": "boolean" },
"adaptations": { "type": "array", "items": { "type": "string" } }
},
"additionalProperties": false
}How a battery fools itself
A non-verbal icon can still have a learned verbal name. Articulatory suppression can impose breathing, timing and motor costs. A matched tapping task can be too easy. Transfer items can preserve a visual cue that carries the answer. Examiners can unintentionally point, pause or reward. A participant can understand the task but fail the touchscreen. These are not housekeeping details. They are alternative causal generators.
Adversarial paths to a false language claim
Failure modelA convincing result survives attacks on access, stimulus, load, response, learning and analysis.
The strongest alternative
The strongest rival account is not that every surviving performance is secretly a sentence. It is that language reorganises cognition during development so deeply that mature non-verbal systems carry linguistic history even when current language processing is absent. On this view, words help create reusable concepts, explicit rule structures and cross-domain control, after which those structures can be executed elsewhere.
This alternative predicts developmental asymmetries. A capacity might survive adult aphasia yet fail to emerge normally when relevant linguistic scaffolding was never available. Testing that claim requires longitudinal designs, varied language experience, non-verbal teaching controls and care not to confuse language deprivation with broader social or educational deprivation. The current evidence does not license a single verdict across those conditions.
sequenceDiagram participant Q as Question participant M as Mechanism participant E as Evidence participant D as Decision Q->>M: The microphone problem M->>E: The estuary of thought E-->>M: compare rival explanations E->>D: Test non-verbal routes before denying thought D-->>Q: revise scope or act
Test non-verbal routes before denying thought
Stop asking whether language is necessary for thought in the abstract. Ask a narrower causal question: for this operation, at this time, does language provide report, access, an optional strategy, cross-domain coordination, or a constitutive representation?
That decision changes experimental design. Give the participant a non-verbal route into the task and a non-verbal route out. Compare language interference with a matched non-linguistic load. Require transfer. Record the full trial path. Treat a verbal explanation as one evidence channel, not the gold standard against which every mind is judged.
It changes clinical and institutional judgement. Aphasia, motor impairment, infancy and species difference all restrict familiar forms of report. A person who cannot answer a syntactically complex question may still compare options, detect a violation, form a preference or understand an event through another route. This does not remove the need for careful capacity assessment. It changes the default from equating fluency with intellect to testing the relevant decision through accessible channels.
It changes how human and machine systems are compared. A language model may perform cognition through textual representations more centrally than a human does. Human evidence that reasoning can dissociate from language neither proves nor disproves what a machine requires. It does, however, block a lazy inference from fluent output to a shared cognitive architecture.
Decision rule: never infer the absence of thought from the absence of fluent language until non-verbal access, non-verbal response, matched load and transfer have been tested.
Language is one of humanity’s most consequential cognitive technologies. It makes private distinctions addressable, exact and shareable. Its importance becomes clearer, not smaller, when we stop asking it to be the substance of every thought.
Glossary
- Access
- The route by which instructions, stimuli or stored state reach the process being tested.
- Aphasia
- An acquired impairment of language production or comprehension, often following brain injury; its pattern and severity vary.
- Constitutive
- Part of what makes a defined operation possible, rather than a resource that merely improves it.
- Matched load
- A control task designed to equal the timing, effort and motor demand of an intervention without targeting language.
- Non-verbal
- Not requiring natural-language comprehension or production in the task path; it does not mean non-symbolic or non-conceptual.
- Report gap
- The difference between performance through a non-verbal response and performance requiring a verbal response.
- Specific language cost
- The performance difference between a matched non-linguistic load and a language-targeting interference condition.
- Transfer
- Successful use of a learned relation on novel items that defeat surface memorisation.
References
- Kean, H., Fung, A., Jaggers, P., Chen, J., Rule, J. S., Benn, Y., Tenenbaum, J. B., Piantadosi, S. T., Varley, R. A., & Fedorenko, E. (2026). Evidence from formal logical reasoning reveals that the language of thought is not natural language. Proceedings of the National Academy of Sciences, 123(28), e2520095123. DOI.
- Varley, R., & Siegal, M. (2000). Evidence for cognition without grammar from causal reasoning and theory of mind in an agrammatic aphasic patient. Current Biology, 10(12), 723–726. DOI.
- Varley, R. A., Klessinger, N. J. C., Romanowski, C. A. J., & Siegal, M. (2005). Agrammatic but numerate. Proceedings of the National Academy of Sciences, 102(9), 3519–3524. DOI.
- Ivanova, A. A., Mineroff, Z., Zimmerer, V., Kanwisher, N., Varley, R., & Fedorenko, E. (2021). The language network is recruited but not required for nonverbal event semantics. Neurobiology of Language, 2(2), 176–201. DOI.
- Xu, F., & Spelke, E. S. (2000). Large number discrimination in 6-month-old infants. Cognition, 74(1), B1–B11. DOI.
- Baillargeon, R., Spelke, E. S., & Wasserman, S. (1985). Object permanence in five-month-old infants. Cognition, 20(3), 191–208. DOI.
- Paulus, M. (2022). Should infant psychology rely on the violation-of-expectation method? Not anymore. Infant and Child Development, 31(1), e2306. DOI.
- Raby, C. R., Alexis, D. M., Dickinson, A., & Clayton, N. S. (2007). Planning for the future by western scrub-jays. Nature, 445, 919–921. DOI.
- Kabadayi, C., & Osvath, M. (2017). Ravens parallel great apes in flexible planning for tool-use and bartering. Science, 357(6347), 202–204. DOI.
- Hampton, R. R. (2001). Rhesus monkeys know when they remember. Proceedings of the National Academy of Sciences, 98(9), 5359–5362. DOI.
- Amodio, P., Brea, J., Farrar, B. G., Ostojić, L., & Clayton, N. S. (2021). Testing two competing hypotheses for Eurasian jays’ caching for the future. Scientific Reports, 11, 835. DOI.
- Shepard, R. N., & Metzler, J. (1971). Mental rotation of three-dimensional objects. Science, 171(3972), 701–703. DOI.
- Bainbridge, W. A., Pounder, Z., Eardley, A. F., & Baker, C. I. (2021). Quantifying aphantasia through drawing: Those without visual imagery show deficits in object but not spatial memory. Cortex, 135, 159–172. DOI.
- Kay, L., Keogh, R., Andrillon, T., & Pearson, J. (2022). The pupillary light response as a physiological index of aphantasia, sensory and phenomenological imagery strength. eLife, 11, e72484. DOI.
- Pica, P., Lemer, C., Izard, V., & Dehaene, S. (2004). Exact and approximate arithmetic in an Amazonian indigene group. Science, 306(5695), 499–503. DOI.
- Winawer, J., Witthoft, N., Frank, M. C., Wu, L., Wade, A. R., & Boroditsky, L. (2007). Russian blues reveal effects of language on color discrimination. Proceedings of the National Academy of Sciences, 104(19), 7780–7785. DOI.
- Nedergaard, J. S. K., Wallentin, M., & Lupyan, G. (2023). Verbal interference paradigms: A systematic review investigating the role of language in cognition. Psychonomic Bulletin & Review, 30, 464–488. DOI.
- Fedorenko, E., Piantadosi, S. T., & Gibson, E. A. F. (2024). Language is primarily a tool for communication rather than thought. Nature, 630, 575–586. DOI.