By Jo Scott-Coe | Contributing Columnist
Ask ChatGBT about the 1966 shooting rampage at the University of Texas at Austin, and you will get the basics of common knowledge. On that terrible day, Charles Whitman, a former Marine, took his weapons to the campus and climbed the tower. For 96 minutes he shot at people, killing 15 and wounding 31 after first killing his wife and mother.
As a researcher who has studied the shooting for the greater part of 10 years, I have asked ChatGBT follow-up questions about Whitman’s wife, Kathy. In its responses, the bot never hesitated: Kathy was married at age 23. She had filed for divorce citing abusive behavior against herself and her young son. She was suffering from an illness, and an autopsy revealed that she likely had a brain tumor. Her own mother was also killed in the shooting. She had three children with Whitman born in 1967, 1968, and 1971. She hid in a closet the day of his attack. She was questioned intensely by authorities. She later remarried. She valued her privacy and changed her name. She died in 2018.
All these statements are false.
To correct the errors: Kathy married at age 19 and at the age of 23 was killed by her husband. While she had spoken of divorce, she did not file. She did not have children before (or after) her death. She was not suffering from an illness. No autopsy was performed on her. She did not have a brain tumor. Her mother was not killed. She did not hide in a closet the day of the attack. She was never questioned by authorities.
She never had the chance to remarry. She never had the chance to change her name.
When contradicted, the bot apologized profusely and retracted the invented facts it had so swiftly offered. But in one response, the bot erased Kathy altogether: “Charles Whitman did not have a wife named Kathy or Kathleen Leissner,” abruptly adding a new hallucination, that Whitman had married a different woman (named “Cissy”) in San Marcos, Texas.
When I asked the bot to cite its sources, only one was a real source. For the others, it invented plausible but false titles and attributed them to actual author names or recombined an existing source title with the name of an existing author who was not the actual author. When corrected on these errors the first time, the bot dug in. When corrected a second time, the bot apologized and retracted all sources it had named.
Unlike a human researcher, the bot will not say, from the outset, “I do not know” or “I am curious.” It speaks first and retracts only when called out. If contradicted, it may accept a new answer but has no means other than plausibility by which to evaluate the input.
The information itself makes no difference to the bot. It is unaware that harvesting or scraping symbols is not the same as comprehending content, yet the bot aims to impress with a flash of output that resembles what we call paragraphs. The program has been taught to scrape and recombine, repeating patterns as Emily M. Bender says, like a “parrot.”
For my forthcoming book, “Unheard Witness,” I spent seven years working with handwritten archival materials — a massive collection of Kathy’s personal correspondence during her short and abusive marriage. Her language provides an up-close and intimate portrait of her experience trying to love, and to leave, a dangerous man.
One of the most disturbing tactics used by domestic abusers such as Kathy’s husband is now termed “perspecticide,” which wears down the target so insidiously that she may doubt what she knows — whether a process from her field of expertise, a recipe she has made a thousand times, or the spelling of a word.
On a massive and depersonalized scale, ChatGBT and other AI models compound the minimization and distortion already endemic in sanitized versions of history. Stories deemed marginal have always been challenging to recover, and it takes methodical labor to preserve testimony all but erased in the human record — especially from women, people of color, the poor, and LGBTQ individuals.
I refuse to be impressed by technology that mimics the overconfidence of the uninformed or the worst manipulations of an abuser. This is the nightmare so many writers and researchers face when they share their work. Instead of curiosity, we often face faux expertise based on a cursory web search or “something I heard.”
Chat bots are no substitute for the work of authorship. They do serve as sobering reminders about what it means not to retract in the face of bot-like or abusive humans.
Jo Scott-Coe’s new book is “Unheard Witness: The Life and Death of Kathy Leissner Whitman,” forthcoming in October from University of Texas Press.
