In case you were looking for another chatbot cautionary tale, the latest comes from a decision issued by an administrative law judge (ALJ) analyzing a company’s arguments in defense of a National Labor Relations Act (NLRA) Section 7 charge.
The facts underlying the charge are simple enough. Daniela Melendez, an administrative assistant at a company called Autofit, disclosed her hourly wage to a curious coworker. Armed with that information, the coworker went to management and demanded equal pay. Just over two weeks later, Melendez was fired. When Melendez applied for unemployment benefits, the company had to submit a written explanation for her termination to the Texas Workforce Commission. Autofit’s explanation? Melendez had been “instructed to keep sensitive information confidential, yet she shared pay details with colleagues.” The problem with that, of course, is that firing an employee for talking about pay is unlawful. So the company had, in writing, admitted to breaking the law.
Autofit tried to walk that confession back—twice. Initially, it argued that “shared pay details” didn’t actually mean talking about pay but rather accessing and sharing other employees’ sales-commission information. Unsurprisingly, given that this spin didn’t align with the statements made in the paperwork Autofit submitted, the argument didn’t stick. Then—and this is the fun one—the company’s witness testified that she hadn’t written the problematic sentence at all. Actually, ChatGPT had, and had done so, the witness said, without any input from Autofit. As the witness described it, she used ChatGPT to sound professional, pasted its output straight into the Texas Workforce Commission submission, couldn’t recall whether she read it, and assumed someone downstream would fix any errors. Oops.
Administrative Law Judge Sharon Steckler was not impressed. In her decision, she noted that Autofit’s “answers to the Texas Workforce Commission about why Melendez was terminated show that [Autofit] was less than pleased about Melendez having a protected concerted discussion about pay.” The judge then went on to explain that she did “not credit [the] explanation that ChatGPT added pay without [the author’s] help.” Even if the chatbot had generated the language, the judge noted, the company could have reviewed it and taken the offending part out (and clearly did not). The remedy? Reinstatement, backpay, other pecuniary harms, expungement of the file and a notice posted on the wall for 60 days.
The takeaway here is not “never use AI.” It’s that AI does not launder your legal obligations. Whatever a chatbot drafts on your company’s behalf, a human owns. Read it before it goes to a government agency.
