The Classic

The ELIZA effect is a classic example from 60 years ago, where a simple program simulates a psychotherapist by just playing back simple sentences to the user based on the user's latest input.

There was no intelligence, no language model, nor any other kind of sophistication involved. Just a sentence-to-sentence mirroring.

However, many users reported feelings of connection with the imaginative counterpart even though they knew that it was a computer and not a person. They started to share private secrets.

It was reported that the secretary - who was part of the development team and knew the technical setup - asked the inventor Mr. Weizenbaum to leave the room when she conversed with the system. For privacy reasons.

Note that knowledge of the system does not prevent us from falling for the effect.

The Barnum Amplifier

The Barnum effect makes us assign credibility to information that is sufficiently vague but somehow matching to our profile or personality. It partly explains the popularity of horoscopes and fortune cookies and anecdotal evidence.

Large Language Models are configured from huge amounts of data. That data gets interpolated and extrapolated and compressed, which leads to the infamous "trendslop": generalized, average, vague, mainstream answers that fit the Barnum characteristic.

We talk to ourselves

It might be counterintuitive but when using an AI chatbot we essentially talk to ourselves.

The whole setup makes us think that we use a tool. The ELIZA effect makes the tool appear like a person. The Barnum effect makes the answers appear personally relevant to us. This combination hides the fact that it was our very input that in the course of the conversation provided all the perceived "intelligence":

• the conversation paths that we followed,
• the additional questions that we asked, and
• the constraints that we imposed.

In short: we essentially talk to ourselves, augmented with some additional details that are mixed in from public and stolen sources.