Unconscious bias

To utter a sentence is to prefer one model of the world; one “narrative” over the infinity of alternatives one could construct. To construe one is to do likewise: the success of the communication such an utterance and construal represents depends on how far the speaker’s and listener’s respective narratives coincide — how far they share the cultural conventions on which the language is founded. To understand a language is to understand the cultural conventions which it represents.

Philosophy Go Premium A “child of a child,” yesterday. The JC looks deep into the well. Or abyss. Click ᐅ to expand: Philosophy Comments? Questions? Suggestions? Requests? Insults? We’d love to 📧 hear from you. Sign up for our newsletter.

Un petit d’un petit
S’étonne aux Halles;
Un petit d’un petit
Ah! degrés te fallent.
Indolent qui ne sort cesse
Indolent qui ne se mène
Qu'importe un petit
Tout gai de Reguennes.

—Luis d’Antin van Rooten, Mots D’Heures: Gousses, Rames, (1967)

If speaker and listener do not share cultural conventions, the result is usually incoherence: a monolingual English speaker cannot understand a French sentence, and that is that. Not necessarily, though: ask a French speaker to read van Rooten’s poem to an English speaker and watch what happens.

These cultural conventions are buried deep: the very beauty of language is that it allows us carry on without having to prove out all the assumptions, linguistic architecture, vocabulary, before communicating. It is to pre-agree a set of filters, switches and conventions. It allows us to take certain things as read.

These filters, switches and conventions are “biases”. We need them, to put one foot in front of another. An operating condition to meaning is shared, unstated, unconscious bias.

So, all language is biased. It’s part of the beauty, intractability, ambiguity and humanity of language. It is bias that sculpts the overwhelming white noise of the universe into meaningful figures; bias limits and makes manageable the infinite; bias makes the universe coherent to us.

We see well-meant attempts to apply neural networks to the existing corpus of human communication, such as it is, to train artificial intelligence discover undesirable biases in the canon — not always apparent on the surface of a given communication, but that emerge from the statistical aggregation of billions of lines of text. “Men” are associated with money, power, and status; while women are associated with “home” and ”mum”.

Now the written canon, for all its flaws and biases, is the nearest account of how humans do communicate. But it is biased in obvious ways: it favours those who write over those who do not; the loudmouth over the introvert, those who have the time, resources and inclination to publish over those who don’t; those who have a platform over those who don’t.

The same thing that is uncomfortable about that is democratic about that. It is democracy’s prerogative to deliver a result any voter, personally, might not like. That a small group of well-attuned, influential people — such a group you might call an “elite” — can make value judgments to correct for the unmediated, actual speech of the population – to second guess what the population did say with what, in that elite’s opinion, it ought to have said, is a political act. It is to rewrite the record. To erase the past.

Now statistical correlations between words either do say something profound about the values embedded in a text, or they don’t. But either way one should tread carefully before remediating or correcting text for “unwanted” biases. Unwanted by whom?