An investigation across 45 languages and 12 language families reveals a universal language network

Nature Neuroscience 25(8):1014–1019, doi 10.1038/s41593-022-01114-5; NIH author manuscript, PMC10414179. Read by researcher R4e (innate against learned) for research batch R4, 5 October 2026. Provenance: papers/base/malikmoraleda2022_universal_language_network.provenance.json.

What was read

• Read in full: every line of the PMC author manuscript, converted from XML: abstract, main text, Methods (participants, design, localizers, tasks, acquisition, preprocessing, fROI definition, statistics), all ten Extended Data figure legends (which carry several of the key statistics), acknowledgements, references and the three main figure captions.
• Not read: the Supplementary Information (Supp. Tables 1–4 with prior studies, participants' language profiles and mixed-model results; Supp. Figures 1–8, including breakdowns by language family and fROI); figures as images.

What they did

• Participants: 86 adults (91 recruited, 5 excluded for motion or sleepiness; 43 male; aged 19–45, mean 27.5), right-handed, recruited at MIT and in Boston. Native speakers of 45 languages from 12 families: Afro-Asiatic, Austro-Asiatic, Austronesian, Dravidian, Indo-European, Japonic, Koreanic, Atlantic-Congo, Sino-Tibetan, Turkic, Uralic, and Basque (an isolate). Two speakers per language (one man, one woman), except one each for Tagalog, Telugu, Slovene and Swahili. 31 of the 45 languages are Indo-European. All participants were also fluent in English.
• "Shallow sampling" by design: few speakers, many languages. The aim was to test whether the properties established in English speakers generalise, not to find differences between languages.
• Tasks (one 2-hour fMRI session):
  • an English reading localizer (sentences against nonword lists), used to define each person's language regions;
  • the critical localizer: passages of Alice in Wonderland in the native language, against acoustically degraded versions and against an unfamiliar language (Tamil, or Basque for 11 people who had heard Tamil);
  • non-linguistic tasks: spatial working memory (everyone) and arithmetic (67 of 86);
  • about 5 minutes of story listening in the native language, and 5 minutes of rest.
• Analysis: language regions defined in each individual (top 10% of voxels inside six standard left frontal, temporal and parietal parcels, plus right-hemisphere homologues); the domain-general "multiple demand" (MD) network as a comparison. Statistics are across languages (n = 45).

Main results (verified)

• The same network, with the same properties, in every language tested. In left-hemisphere language regions (percent signal change):
  • native language 2.13 against degraded speech 0.84, and against an unfamiliar language 0.76 (both t(44) = 21.0, p < 0.001);
  • across languages, effect sizes range from 0.49 to 2.49 (native against degraded) and 0.54 to 2.53 (native against unfamiliar).
• Left-lateralised everywhere: stronger on the left than the right (2.13 against 1.47, t(44) = 7.00) and more extensive (318.2 against 203.5 voxels, t(44) = 6.97).
• Functionally integrated: regions correlate during story listening (r = 0.52) and rest (0.41), more on the left than the right (0.52 against 0.35; 0.41 against 0.28), and more during stories than rest.
• Selective for language: native language 2.13 against spatial working memory −0.01 (t(44) = 20.7) and arithmetic 0.03 (t(40) = 21.5). Language and MD networks are each internally coupled (story: 0.43 and 0.40) but not with each other (−0.01; rest: 0.34, 0.43 and −0.03). The MD network, in contrast, responds more to degraded and unfamiliar speech than to the native language.
• Language does not shape the map more than the individual does.
  • Overlap of individual maps between two speakers of the same language (Dice 0.17, SD 0.07) was no higher than between speakers of different languages (0.16, SD 0.06; t(40.7) = −0.52, p = 0.61). Same language family against different families: also no difference (p = 0.49).
  • The overlap atlas from 45 languages correlates with an atlas of 629 English speakers at r = 0.83, and of 19 Russian speakers at r = 0.85, with a comparable spread of overlap values.
  • The authors' summary: variability across languages is comparable to, or lower than, variability between individuals who speak the same language.

The authors' interpretation and caveats

• The network's position relative to other systems, its left-lateralisation (in most individuals), its integration and its selectivity "make it well-suited to support the broadly common features of languages, shaped by biological and cultural evolution."
• Differences between languages may exist but are likely subtle: detecting them needs deep sampling (many speakers per language), finer paradigms, multivariate methods or high temporal resolution. Example hypotheses: more lateralisation for strict word order; stronger pitch-to-language links in tonal languages; more inferential (right-hemisphere, mental-state) processing for languages like Riau Indonesian whose utterances underdetermine meaning.
• Limitations they name: every participant was bilingual (fluent in English), which may or may not affect native-language processing (disputed); Indo-European languages are over-represented (31 of 45).
• The localizers are public for 46 languages.

Limits (reader's)

• Adults living in the US, all bilingual in English. Not tested: monolinguals, people living in their own language's society, children.
• Two people per language, so the study can confirm shared properties but has no power to detect language-specific ones.
• Group statistics over languages; no individual-level reliability is reported here (cited from earlier work).

What it means for Amadeus (inference)

• The language system is universal; the language is not. Every speaker has the same left fronto-temporal network with the same properties. That network responds to the person's own language (2.13), not to an unfamiliar one (0.76, about the level of unintelligible degraded speech, 0.84). The machinery is shared by all humans; what it is tuned to is acquired from the speaker's community.
• For the base: the architecture and properties of a language system (selectivity, separation from general problem-solving machinery) qualify as "what we all share". A particular language does not: it is acquired from a shared cultural environment. A base trained on one language is therefore already partly culture-specific (an inference about our design, not a claim of the paper).
• The individual differs more than the language does. Speakers of the same language differ as much as speakers of different languages in where the network lies. Person-level variation is at least as large as culture-level variation in this measure.
• Language and thinking are separate systems here. The language network did not respond to arithmetic or spatial working memory, while the MD network did. A vocabulary engine (B1) is not, by this evidence, the whole of reasoning.