Official portrait of George Santos
22 February 2023
I asked ChatGPT to write my biography. In doing so, it showed itself to be a worthy rival of George Santos.
Things I didn’t know about myself:
My birth certificate is wrong. Evidently, I was born on 28 November 1955 and not on a different day in 1963.
I was born in Chelmsford, Essex in the UK, and not in New Jersey, USA as my parents told me. (I don’t remember the event, myself.)
I attended the University of Sheffield, earning a bachelor’s degree in French and German. Although I do seem to vaguely recall attending Lafayette College in Pennsylvania where I did take a few semesters of German while majoring in Political Science. Don’t remember any French though.
I have a master’s degree in applied linguistics from the University of Edinburgh. Oddly, I always thought my master’s was in National Security Policy from George Washington University and that I have a PhD in medieval English from the University of Toronto. I do recall visiting Edinburgh, however.
Evidently, I worked as a lexicographer for both Chambers Harrap and Oxford University Press, and I was a senior editor for the second edition of the OED. Strangely, I always thought I was in the US Army stationed in Germany at the time the OED2 was published.
I was made an Officer of the Order of the British Empire (OBE) in 2016. I think would remember having met the queen, but maybe not.
Things I did know about myself:
I do work in the fields of linguistics and lexicography.
I’m the author of Word Myths: Debunking Linguistic Urban Legends.
I write the Wordorigins.org blog.
When I typed the request, I thought this would be a neat little experiment in the limits of what ChatGPT knows. I’m essentially a nobody, and I assumed—correctly as it turns out—that there would be little information about me in ChatGPT’s training data. I wanted to see what the AI did with just that kind of limited info. I suspected it would produce a couple of lines, mentioning the book and website, and that’s all. What I got surprised me.
First, it heaped on the praise. Evidently, I’m a “beloved figure in the world of linguistics and lexicography.” While I do like to think that’s the case, even I have to admit that the AI was laying it on thick.
But that’s not the worst of it. The AI spit out 334 words in 6 paragraphs of almost complete fiction. It got only two things right, the book and blog, just what I figured might be somewhere in its training data. What astounded me was the extent to which it fabricated “facts.” (An OBE?! Really?!)
I asked it to regenerate the biography listing its sources, and it cited three, giving URLs for each. The first was a non-existent Wikipedia article. The second was to a UK National Archives page on Brook Walton, an eighteenth-century English baronet and MP. (I had never heard of him, either.) And the third was to a non-existent page on the OED site. Not only did it create “facts,” it created fictional sources for them.
This was truly disturbing and an indicator of how the AI cannot be relied upon, at least not in its present state.
The fabrication of data is not simply a result of its training data being limited. The problem is in the very nature of its programming. When it has little data, it tries to generate plausible information rather than admitting that its knowledge is limited. It is easy to say that the AI will improve with more and better training data—I suspect a ChatGPT biography of Barack Obama, for instance, would be pretty darn accurate. While that’s undoubtedly true, it misses the point. The purpose of the AI is make connections between disparate data; it engages in induction. But an essential requirement of induction is recognizing the limits of one’s data and not to exceed them. ChatGPT is clearly terrible at this. More and better training data will reduce the frequency of this occurring, but it will always happen when its training data does not cover the subject at hand.
In this case, the fabrications are obvious (but perhaps not to a person who stumbles upon Wordorigins.org and wants to see my credentials, which might be the only reason someone, other than me, would want to ask an AI about me), but this is an insidious problem that undoubtedly occurs often but goes unnoticed in output for which the AI has good training data but fills in minor gaps with plausible-sounding fabrications.
Unless the AI can be taught to recognize the difference between induction and a lie, it is pretty much useless as a general tool. And those in Silicon Valley should think long and hard before releasing this or any similar AI into the wild.
Image credit: U.S. House Office of Photography, 2022. Wikimedia Commons. Public domain image.
