Every language you know was learned by children from adults who learned it from other children. Nobody designed the result, and yet it has grammar: parts that combine, so you can say a sentence you have never heard. A famous experiment in 2008 showed that a pile of random noise can turn into something like grammar just by being passed down a chain of learners who only ever see part of it.
Simon Kirby, Hannah Cornish and Kenny Smith, then at the University of Edinburgh, built a tiny world of 27 pictures. Each picture was a coloured shape with a motion arrow: three shapes (square, circle, triangle), three colours (black, blue, red) and three motions (horizontal, bouncing, spiralling). The first "alien language" gave each picture its own random label, made of two to four syllables such as kapimu. There was no pattern at all.
A volunteer studied 14 of the 27 picture-and-word pairs, then was shown pictures and asked to type the word the aliens would use. Those pictures included the 13 they had never seen, so they had to guess. Their answers, all 27 of them, became the language taught to the next volunteer. The study ran four separate chains of ten people, 80 students in total, each chain starting from a different random language.
Over the generations, the words went wrong less and less often, and the languages gained structure. In the main experiment the lists collapsed from 27 different words to only about two to five, with one word covering whole groups of pictures, such as everything that moves horizontally. When the researchers filtered the training data so that no two pictures shared a word, the languages kept more words (roughly 12 to 23) and became compositional: in one chain the words split into separate parts for colour, shape and motion. Nobody asked for either outcome.
This toy has nine meanings (three shapes by three colours), and every word is two syllables. Syllables are colour-coded so you can see them repeat. Each learner sees only the outlined words, then fills in the dashed ones using a simple habit: "the first syllable goes with the shape, the second goes with the colour." Press Pass it on and watch the language change. Tap words to choose what the next learner gets to see.
Rows are shapes: circle, square, triangle. Columns are red, blue, yellow. Outlined cells are what the next learner will see; tap to change them.
A language that must pass through a learner is only as good as what that learner can reconstruct. A word that belongs to no pattern has to be memorised one picture at a time, so when it falls among the unseen items it is simply lost, and a guess takes its place. A word that follows a pattern survives, because a learner who has met the pattern can rebuild it from other words. Over many generations, the patterns are what is left standing.
Try the slider. With all nine words visible there is no squeeze: a random language is copied perfectly, forever, and never gains any structure. Show learners five or six, and the language usually settles into a tidy rule within a handful of generations and then stops changing. Show them only two or three, and the language keeps shuffling, because there are too few examples to pin the pattern down.
Our learners come pre-loaded with the habit of splitting words into a shape part and a colour part, and they never forget a word they have seen. Real people forget, misremember, and bring their own biases, and in the experiment the structure had to emerge without the volunteers being told to look for it. The real study also scored structure statistically, comparing each language with 1,000 reshuffled versions of itself. What the toy keeps is the central mechanism: when a language is rebuilt again and again from partial evidence, the parts that can be predicted are the parts that last.
Languages are not only things we use. They are things we repeatedly learn, and the learning is a filter. Anything that cannot be rebuilt from a few examples gets dropped, and anything that can be rebuilt gets stronger. Grammar may be what is left after many generations of that filter.
Source: Kirby, S., Cornish, H. and Smith, K. (2008), "Cumulative cultural evolution in the laboratory: An experimental approach to the origins of structure in human language", Proceedings of the National Academy of Sciences 105(31).