A proto-language is the reconstructed common ancestor of a language family, the speech from which its member languages are thought to descend. It is usually not written down, so linguists work it out by comparing the languages that came from it.
Key takeaways
- A proto-language is inferred from its descendants, not observed, and is usually unattested or only partly attested.
- Borrowed words are excluded from the comparison, because they do not reflect descent.
- Indo-European, with 485 languages on the map, is the family whose ancestor has been studied longest.
- The map shows living languages only, so no proto-language appears on it.
How proto-languages work
A language family is a group of languages that descend from a shared ancestor. That ancestor is the proto-language: in the strict sense, the most recent common ancestor of the family, just before it began to split [1].
Nobody recorded it, in most cases. Proto-languages are usually unattested or only partly attested, and linguists reconstruct them by deducing shared features from the languages that survive [1]. The comparative method does this. Researchers line up cognates, words that look and mean alike across languages, set aside borrowed words because they do not reflect descent, and work backwards to the sounds and forms that could have produced the patterns [1][2]. If descent from the proto-language can account for every shared feature, the family tree counts as a complete explanation [1]. Reconstructed forms are conventionally marked with an asterisk to show that they are inferred, not found in a text.
August Schleicher produced the first systematic reconstruction of an unattested proto-language, Proto-Indo-European, in 1861 [1]. Some proto-languages are known from writing. Latin is the proto-language of the Romance languages, and Proto-Norse appears in fragmentary early runic inscriptions [1].
Reconstruction has limits. Scholars disagree about the nature of Proto-Indo-European, and some describe its reconstructions as formulae or abstractions and not a record of real speech [1]. A reconstructed proto-language is best read as a well-reasoned model of the ancestor, not a transcript.
An example from our data
Indo-European is the family whose proto-language is the best known. On our map it has 485 languages in 8 branches, led by Spanish, English and Hindi. All are treated as descendants of one reconstructed ancestor.
Austronesian is another case: 1,226 languages in 5 branches, with Standard Indonesian, Filipino and Javanese among the largest. Its languages are also traced back to a single proto-language, which linguists reconstruct from comparison, not from texts.
The map cannot show any proto-language. It draws living languages from Glottolog, with 1,023 extinct languages left off, and ancestors that no one speaks today are not in the data. What the family pages show is the result of the reconstruction work: which languages are grouped together as descendants.
The thing people get wrong
A frequent assumption is that a proto-language is a real, documented ancient language, like a dusty manuscript language that has been translated. It is a scientific reconstruction, and the forms are hypotheses supported by comparison.
Another is that the parent language is "better" or "purer" than its descendants. Languages change continuously, and a descendant is not a degraded copy of its ancestor. Each language in a family has simply changed in its own way since the split.
Frequently asked questions
Did anyone ever speak Proto-Indo-European?
A community presumably spoke the ancestral language, but no text in it survives. What linguists have is a reconstruction, and scholars disagree about how closely it matches real speech.
Is Latin a proto-language?
Of the Romance languages, yes. Latin is known from texts, which makes it unusual, because most proto-languages are not written.
Sources
- Proto-language, Wikipedia.
- Language family, Wikipedia.
- Every figure from our own map is worked out from the datasets on the data sources page (checked against Glottolog and Glottography, retrieved 13 October 2026).