A Texas company’s attempt to blame an AI chatbot for incriminating language in an official document has failed spectacularly before a federal labour judge.
Daniela Melendez worked as an administrative assistant for Autofit, an aftermarket auto parts company, earning $22 an hour before her abrupt dismissal.
Just two or three days into the job, a colleague asked Melendez what she was earning, and Melendez told her the figure.
The other employee then printed out Melendez’s job posting, marched into a manager’s office, and demanded equivalent pay, which management refused.
Sixteen days after that conversation, Melendez was fired, a sequence of events that would later draw scrutiny from the National Labour Relations Board.
When Melendez applied for unemployment benefits, Autofit was required to explain its reasoning in writing to the Texas Workforce Commission.
The company’s explanation, submitted via Chief Logistics Officer Elham Emadi’s administrative assistant Esperanza Guevara, stated that Melendez “shared pay details with colleagues” as a serious breach of confidentiality.
Under US labour law, discussing pay with colleagues is explicitly protected activity, and firing an employee for doing so is illegal.
Autofit first attempted to redefine the phrase during a sworn statement, arguing “shared pay details” referred to misappropriating others’ sales commissions rather than discussing wages, a reading Judge Sharon Steckler did not accept.
The company’s lawyer then introduced a second defence at trial, arguing the damaging language was not written by Guevara at all but was generated by the free version of ChatGPT.
Guevara testified she had used ChatGPT to produce a more professional-sounding response and then copied and pasted the output directly into the official form without thoroughly reviewing it.
She stated she had assumed ADP would correct any errors, and notably had not mentioned the AI involvement to anyone until witness preparation ahead of the trial.
Judge Steckler was unconvinced, writing: “Guevara would have us believe that ChatGPT made up the language that sensitive information includes pay.”
The judge further noted: “I therefore do not credit Guevara’s explanation that ChatGPT added pay without Guevara’s help.”
Steckler also observed that even if ChatGPT had generated the language, Guevara could have reviewed the text and removed the reference to pay before submitting it.
Autofit lodged 16 affirmative defences, including a Seventh Amendment jury trial demand and a constitutional challenge to the NLRB’s authority, none of which succeeded.
The judge ordered reinstatement, backpay, compensation for pecuniary harms, expungement of Melendez’s file, and a notice to be posted at the company for 60 days.
A pointed irony runs through the case: Autofit’s witnesses testified the company prizes trustworthiness above all else, yet outsourcing a state agency filing to a free chatbot apparently posed no problem for management.