An administrative law judge has rejected an employer's claim that ChatGPT — not a company employee — wrote the sentence in which the employer admitted to firing a worker for discussing pay. Administrative Law Judge Sharon Steckler found Autofit violated the National Labor Relations Act and ordered reinstatement, backpay, other pecuniary damages, expungement of the employee's file, and a 60-day posted workplace notice.

The decision lands squarely on obligations HR teams touch every day: what a company writes to a state unemployment agency about a termination can become evidence in an NLRA case, and offloading that drafting to a chatbot does not transfer responsibility for its contents.

What happened

The facts are simple. Daniela Melendez, an administrative assistant at Autofit, told a coworker her hourly wage. The coworker took that information to management and demanded equal pay. Just over two weeks later, Autofit fired Melendez.

The trouble deepened when Melendez applied for unemployment benefits. Autofit had to submit a written explanation of the termination to the Texas Workforce Commission. The company wrote that Melendez had been "instructed to keep sensitive information confidential, yet she shared pay details with colleagues." Firing an employee for talking about pay is unlawful under NLRA Section 7, which protects concerted discussions about wages. In effect, the company had confessed to an unlawful motive in writing.

Autofit tried to walk the confession back twice. First, it argued that "shared pay details" did not mean discussing pay, but rather accessing and distributing other employees' sales-commission information. That argument failed because it contradicted what the company had already put in its own paperwork.

Second — the argument at the heart of the case — the company's witness testified that she had not written the problematic sentence at all. ChatGPT had, without any input from her. According to her testimony, she used the chatbot to sound professional, pasted its output directly into the Texas Workforce Commission submission, could not recall whether she had read it, and assumed someone downstream would fix any errors.

What the judge said

Judge Steckeler was unpersuaded on both counts. Her decision states 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."

On the AI defense specifically, the judge wrote that she did "not credit [the] explanation that ChatGPT added pay without [the author's] help." Even if the chatbot had generated the language, Steckler noted, the company could have reviewed the output and removed the offending passage. It did not.

The remedies: reinstatement, backpay, other pecuniary harms, expungement of the file, and a notice posted in the workplace for 60 days.

What it means for employers

The ruling does not hold that HR teams must stop using generative AI for drafts, correspondence or agency submissions. It holds that AI does not launder legal obligations. Whatever a chatbot drafts on a company's behalf, a human owns — and that ownership extends to documents filed with government agencies, where admissions can carry direct legal consequences.

For people-process owners, the practical exposure points are familiar ones. Separation paperwork and unemployment responses increasingly circulate through multiple hands and, now, multiple tools. This case shows what happens when no one in that chain reads the final text. Pay discussions are protected under Section 7, and any written rationale that ties a termination to wage transparency creates liability the company then has to explain — to an ALJ, in this instance, without success.

The decision also signals how tribunals are likely to treat AI-assisted filings going forward: as the company's own statements, reviewable and attributable, regardless of who or what produced the first draft. Employers that embed AI drafting into unemployment responses, disciplinary documentation and other compliance-adjacent workflows should expect judges to apply the same standard Judge Steckler applied here — the output went out under the company's name, so the company answers for it.