Anthropomorphism

People shout at printers, name their cars and say that the thermostat “thinks” the room is cold. Anthropomorphism is the habit behind all three: describing and treating a non-human thing as though it had a human inner life. With conversational AI the habit has met an object built to invite it.

Why we do it

Nicholas Epley, Adam Waytz and John Cacioppo proposed three factors that make anthropomorphism more or less likely.

• Available knowledge. Knowledge about humans, and about oneself, is the most accessible model anyone has for an agent, so it is applied first and corrected only with effort.
• The wish to understand and predict. Treating a thing as having goals makes its behaviour easier to anticipate, so the tendency grows when the thing is unpredictable or the stakes are high.
• The wish for social connection. People who are lonely are more inclined to find a companion in a pet or a device.

A chatbot scores on all three. It uses language, the most human signal there is; its behaviour is hard to predict otherwise; and it is available at any hour. The ELIZA effect is the special case in which a text-producing program is credited with understanding.

Anthropomorphism in the science itself

Arleen Salles, Kathinka Evers and Michele Farisco point out that the habit is not confined to users. The vocabulary of the field, with its “neurons”, “learning”, “attention”, “hallucination” and “reasoning”, describes machines in human terms before any argument has been given that the terms apply. This shapes what the public expects and what researchers take themselves to have shown.

The opposite error

Rejecting every mental description of a machine as anthropomorphism begs the question the other way. Whether a system has something worth calling a belief or a goal is a substantive question, and it is not settled by noting that the words were first used of people.

Where I stand

I try to keep two questions apart. One is what a system is: whether it understands, believes or intends anything. The other is what a user is entitled to do with its outputs. Much of my own work argues that the second can be answered without the first, because what matters for knowledge is how the output is connected to the truth, not whether there is a mind behind it. Anthropomorphism does its damage by answering the second question through a guess at the first: it sounds like a person who knows, so I may believe it as I would a person who knows.