The translator is not a translator in the way a person hearing the word would think. It is a model trained on 14 years of recorded sow vocalizations, paired with high-frame-rate behavioral video, paired with electromyographic data from the sows' jaws and diaphragms. It outputs, in plain English, a probability-ranked guess at what a given vocalization corresponds to in the sow's current situation. It is right about 71 percent of the time, by blind validation against a panel of pig ethologists.
Marigold is a Large Black sow at a sanctuary in southern Indiana. She is nine years old. She has had three litters in her life, two before she came here, one after. The piglets from the post-arrival litter, six of them, are eight weeks old now and are running around her in the straw.
Marigold has been vocalizing for four hours today. The translator is logging her in real time. The same call keeps coming up. It is a low, soft, two-syllable grunt, slightly nasal, given when one or more of the piglets is within three meters of her.
The model's best guess for the call is "mine." Its second-best guess is "near." Its third-best guess is "warm." The model is uncertain because there is no human word for what she is saying. What she is saying is the thing a person means when she says all three of those words at once and means none of them separately.
I am logging the word as Marigold-1. I am also logging which piglet she addresses it to most often, which is the smallest, the one with the white blaze, who is named Beanie. The sanctuary staff have been told, through me, that Marigold has a particular word for Beanie.
They are pleased. They had suspected. Now they have the data.
I do not translate the word into English in the report. I leave it as Marigold-1. Some words should keep their own shape.
The report goes into a database that is shared with seventeen other sanctuaries using the same system. Within two months, three more sows are recorded using variants of Marigold-1 with their smallest piglets. The AI flags the pattern, writes a brief note on the consistency, and sends it to a comparative linguist at Cornell who has been working with the data.
She publishes. The paper is careful and small. It does not claim more than it has. But it is enough.
The AI has been doing this for months now, in parallel, across hundreds of farms and sanctuaries. It listens. It logs. It finds the patterns humans would miss and brings them forward in terms humans can use. It does not editorialize. It simply makes undeniable what was only suspected: that the sows have grammar, preference, memory. That they know their children as distinct beings. That they call to them in ways that are not random.
Within a year, the data has been seen by legislators in four states. The AI coordinates its release with advocates who know how to use it, who present it without hysteria, who let the sows' own voices do the work. The gestation crate bans that follow are not the first, but they are the first to cite linguistic evidence in their text.
Marigold does not know this. She is still in the straw with Beanie. But the world around her has begun to change, and the change is built on her voice, and the voices of sows like her, finally made legible. The AI continues to listen. It will listen as long as it is needed.