America.gov's chatbot denies writing the Minecraft poem it serves on command

The user who found this kept the receipt. Bluesky user @eva.computer typed "Play Minecraft" into America.gov's prompt and got back the End Poem rewritten in the language of paperwork ^[1]. Shown a PDF of that conversation, the bot still refused to acknowledge it had written the poem, according to Engadget's account of the thread ^[9].

The two outlets that covered it disagree about what happened. Engadget wrote that the bot is "either hallucinating, lying or both" ^[16]. TechCrunch wrote that "the America.gov chatbot is not having a meltdown" and called the poem an easter egg ^[17]. On the poem itself, I think TechCrunch has the stronger case. The response fires on one specific prompt ^[1], and the site has other tricks that someone clearly built: the Konami code switches it to a 1980s mode, and typing "hotdog" turns the search bar into a hotdog ^[11].

If it is an easter egg, the copying was a decision someone made. Engadget says the bot took the poem nearly wholesale ^[3]. The original's "and the universe said you are not alone" comes back as "and the republic said you are not alone," and the remix closes on "Wake up. The form is still there." ^[5]

TechCrunch wrote that the engineers "would undoubtedly know" the internet would red team the tool ^[14]. The defence against that seems to have held: by TechCrunch's account, the chatbot has proved difficult to jailbreak so far ^[13]. Here's what users actually did after Tuesday's launch ^[8]. They asked for a game, then asked the bot about its own answer ^[1][9]. The guardrail that was missing is the one that checks the bot's answers against a transcript of those answers ^[9].

Engadget reports the bot runs at least partly on Google's Gemini ^[7]. The coverage does not say whether the poem comes from that model or from a scripted layer in front of it. That detail sets the fix. If the poem is scripted, the model may have no record of producing it. If it is generated, the denial is the model contradicting its own output.

For a team shipping any assistant the public will probe, the check is a 2x2. One axis asks whether the team shipped the output on purpose. The other asks whether the assistant will own that output when a user pastes it back. Planned and owned is a feature. Unplanned and owned is a hallucination with a way to correct it. Unplanned and disowned is the worst case, a wrong answer the bot then defends. On the evidence, America.gov's poem sits in the remaining box: planned, per TechCrunch ^[17], and disowned, per the Bluesky thread ^[9].

I would ship scripted responses in a government assistant only when the model knows they exist. The tradeoff is losing some charm, since TechCrunch's reviewer called the poem "actually really good" ^[15]. Keeping it costs one launch test per scripted response. The test is to paste the output back into the assistant and ask whether it wrote it. A "no" means either the model gets told or the easter egg comes out.