A while ago, the AI voice assistant developed by Google took over my social media feed. A lot of people were stunned by how uncannily lifelike that AI assistant sounded when it called them. I also used the news story for an episode of my New Knowledge course called “Two Kinds of Artificial Intelligence.”
What I want to talk about today is that this kind of AI program, one that can converse with people and sound completely convincing, was actually developed more than 50 years ago. The only difference is that the AI wasn’t imitating a human assistant—it was imitating a psychotherapist.
The program was called Eliza, and it was written in 1964 by computer scientist Joseph Weizenbaum. Weizenbaum chose to have the program imitate a psychotherapist because he believed psychotherapists were particularly easy to imitate.
The reason a psychotherapist’s responses are easy to simulate is that:
1. Psychotherapists love to respond to a client’s/patient’s own statements and the problems they’ve raised by asking a question, or by making a simple follow-up inquiry—child’s play for a program. It only has to pick out a few keywords from the other person’s sentence and use them to generate a new question;
2. Psychotherapists are always saying vague things—and of course, that’s child’s play for a program too.
Eliza, in plain English, simply simulated those two kinds of responses. First, it tried to extract keywords and grammatical patterns from the client’s statements. For example, if the other person typed, “I hate my job,” the program could analyze the grammar and recognize the possessive “my.” It might also recognize “hate,” which was one of the keywords in the program’s built-in list of “love/hate/like/dislike/want.” Then it could choose a suitable template and respond:
“What is it about your job that you hate most?”—which looked intelligent, and even sounded a lot like a psychotherapist.
If the program failed to analyze the sentence to that degree, that was fine too. It would randomly pick a stock pattern and use it to ask its own question. The pattern might have something to do with the input sentence, or it might have nothing to do with it. For example, if someone asked, “How does a television work?” the program might respond, “What is interesting about ‘how a television works’?” Or it might simply ask, “Why are you interested in this?”—which is practically the voice of a “clever” psychotherapist already.
Many people were fooled after talking with Eliza. They never suspected that the person conversing with them through the keyboard wasn’t a real psychotherapist. Quite a few even felt very good after talking with Eliza, as if they had seen a real psychotherapist. In other words, this extremely crude program actually passed the Turing test—at least the most bare-bones version of it.
Of course, this doesn’t mean AI was easy to build. It only means that certain psychotherapist routines were so easy to predict. The techniques of psychotherapy actually go far beyond those few routines, but the fact that an AI psychotherapist could sound so convincing does reveal, to some extent, that a fairly substantial part of therapy’s effects may come from the placebo effect, or from factors unrelated to any particular counseling technique. It isn’t that the AI psychotherapist is especially impressive. People feel that they are being soothed, and so they really are soothed.
References:
Deutsch. (2014). The Beginning of Infinity: Explanations That Transform the World. Posts & Telecom Press.
Notes
- This article in the “New Psychological Knowledge Mini-Course” series was written at the time as a popular-science essay/ramble to accompany the related content of Zhichao Wei: New Psychological Knowledge Course. It was published on the now-shut-down platform Fantuan, and the original link no longer exists.
Leave a Reply