Some AI failures require a forensic write-up to explain. This one you can hear in about four seconds. Call the Washington State Department of Licensing, choose the Spanish self-service option, and instead of Spanish you get an automated voice reading English text out loud in a thick Castilian Spanish accent. "Please press 1" arrives as "Please press uno." It is the phone-tree equivalent of a tourist confidently ordering in a language they do not speak.

What callers actually got

The Department of Licensing, Washington's version of the DMV, runs its self-service phone line on what it describes as a newer, AI-driven system offering 10 languages. Amazon provides the platform. When the Associated Press tried to reproduce the effect, its reporters managed it easily using Amazon Polly, Amazon's text-to-speech service, and selecting a voice named "Lucia" that is designed to speak European Spanish. Feed that voice English text and it does precisely what a Spanish speaker would do when reading English words aloud: it pronounces them with Spanish phonetics. Hence "uno" instead of "one."

Crucially, the problem was not limited to Spanish. Reporters and callers found the other language options were also coming through in English. So the self-service feature that promised service in 10 languages was, for a meaningful stretch, delivering English to everyone, just with varying accents depending on which voice had been wired to which menu.

This is a config error, and that is the point

In fairness to the machine, the AI did not hallucinate anything. Amazon Polly read the text it was given in the voice it was assigned. The failure was upstream and very human: a configuration change made during an expansion of the self-service options. Someone appears to have treated a text-to-speech voice as if it were translation software and handed a Spanish-accented voice a script written in English. The agency itself attributed the glitch to a DOL-made configuration change and said it was caused by its own staff.

That distinction matters, but it does not get anyone off the hook. The whole pitch of an "AI-driven" self-service system serving 10 languages is that residents who are not comfortable in English can get help. Shipping it without anyone calling the Spanish line to confirm it produced Spanish is the deployment failure. The model worked. The process that put the model in front of the public did not. This is the same category of mistake as a company feeding its chatbot the wrong reference number; the automation faithfully executed a bad input, and nobody checked the output in production.

It went unfixed for months

The timeline is the part that turns an embarrassing glitch into a real language-access problem. Maya Edwards, a Kitsap County resident, first encountered the accented-English voice in the summer of 2025 when her husband, whose first language is Spanish, tried to use the Spanish option to ask about his driver's license rather than wait on hold for a human. She filmed it and posted it to TikTok on July 29, 2025, where it went largely unnoticed.

She called again months later, found the issue still there, and reposted the video on February 22, 2026. This time it spread, eventually racking up somewhere around 2 million views. The Spokesman-Review and Spokane Public Radio independently called the line in late February to verify, found Edwards' account accurate, and noted that an automated message later that same day acknowledged problems with the language self-service option. In other words, the bug had been live and reachable by any Spanish-speaking caller for the better part of half a year before public pressure forced action.

The apology

Once it was a story, the agency moved quickly. A DOL spokesperson apologized "for the error and to its customers for any inconvenience," described the situation as "an unfortunate byproduct of expanding services," and said the department had determined the issue was caused by DOL staff and had fixed it. The agency added that it would "be more mindful of newer technology as the agency expands language access." Subsequent calls confirmed the Spanish option was working.

Edwards, for her part, was gracious about the speed of the eventual fix. But the underlying complaint she raised is the one worth keeping. Washington markets itself as welcoming to immigrants, and the people most dependent on a working Spanish-language line are exactly the residents least able to fall back on the English menu. For months, the state's answer to "can I get help in Spanish" was a robot mispronouncing English at them.

Why it belongs here

No one was financially ruined and nothing was breached, so this sits at the lighter end of the graveyard. But it is a clean, almost pedagogical example of an AI deployment failure in a public service: an automated, AI-driven system rolled out in front of citizens, promising capabilities (10 languages) it was not actually delivering, with no apparent quality check in the very languages it claimed to support. The fix was trivial once someone looked. The failure was that, for months, nobody official did, and it took a viral video to make a government agency dial its own Spanish line.

The lesson for anyone deploying AI voice or translation in front of the public is almost insultingly simple. Test the thing in every language you advertise, from the user's seat, before you ship it. A text-to-speech engine will read whatever you give it with total confidence. Confidence is not comprehension, and "press uno" is what happens when you confuse the two.