Why do we think LLMs and other Chatbots are thinking beings? Or why do we attribute them thought? For a long time we’ve called the ELIZA Effect, but lately I’ve been thinking that we should call this the Henry Higgins Effect.
(I’ve written up some of my thoughts over on Medium and am not convinced that articles should go here and there.)
The TL;DR version of the argument takes us back to the source material, the ELIZA chatbot and the inspiration for its name, Pygmalion and specifically its movie musical adaptation, My Fair Lady.
In that story, Henry Higgins is an early social linguist, who conducts and experiment (wager) on a lowly flower girl, Eliza Doolittle, who he claims he could transform into a upperclass lady just by teaching her proper diction. (Yep, the story is fraught, and GB Shaw knew it.)
Joseph Weizenbaum took Eliza’s name for his chatbot system (1964-1966), the first of its kind, as a system that could be “taught” to perform talking … and by extension pass itself off as a talking person. If you don’t think Weizenbaum was aware of the gender implications of using that name, just read the sample dialogue he published where the patient, talking to the ELIZA system running a DOCTOR script, declared “Men are all alike.”
As the years went by, the phrase ELIZA effect came to be used for anytime we treat a piece of software (and sometimes other objects) like they have they are people. Yes, lots of class implications here — gets kind of messy. The phrase is often attributed to Sherry Turkle, but predates her usage.
However, in my article, I make the case that we should be paying attention to the Henry Higgins effect, where we attribute “intelligence” to software just because it can produce passible prose. The dangers are as great to a world run by Henry Higginses. (Imagine that!) And to steal the ending of that article, if you’re not concerned about the implications of that, just you wait.
Again, check out the full article and let me know what you think.
BTW, a team of scholars and I are writing a book about ELIZA. More on that here: https://findingeliza.org