The world’s first AI chatbot? Learn about a 60s robot therapist named Eliza from this 99% Invisible podcast episode

What makes a podcast episode stick like that? I wish I could say it is because of the host, Roman Mars, who has one of the most celebrated voices of the medium as well as surely the most antiquarian name. It is not the production values or the musicality, though both are exceptional. For me, the poignant storytelling grooves a kind of moving picture vignette in my mind, possibly in black and white. With captions in Garamond.

Every episode is about a small detail from the built world that does not seem remotely of interest but ends up being fascinating. In hot weather, pummelled by artificially cooled air, I often think of the episodes on vernacular architecture and the one on air conditioning in particular. Others come to mind because of triggers in pop culture. I can’t hear the (amazingly ubiquitous) line “Who let the dogs out?” without thinking of the episode about someone who lost a decade trying to find the song’s provenance.

The Eliza Effect harks back to Joseph Weizenbaum, an MIT professor who in the 1960s created one of the first computer programs that could simulate human conversation, a simple chatbot called Eliza. He came to realise that one way to avoid having to input too much data was by having the program mirror speech, much as a therapist might. This first example of a robot therapist was an instant hit, but Weizenbaum was horrified by how people reacted to the simulated empathy, instinctively and foolishly treating the machine like a conscious being. As he put it, “a relatively simple program could induce powerful, delusional thinking in quite normal people”.

The episode ends, as they typically do, with a present-day tie-in, in this case wondering what Weizenbaum, who died in 2008, would have thought of this new neural network called Generative Pre-trained Transformer 2 created by a company called OpenAI.