This Chatbot Only Speaks in Haiku

Someone built a chatbot that refuses to give you more than seventeen syllables, no matter what you ask it. Tax law. Grief. The heat death of the universe. Doesn't matter. You get five-seven-five, then silence.

Meet HaikuBot, a custom GPT that answers every prompt in exactly one haiku. In an industry racing to write you a 2,000-word essay for a question that needed two sentences, this one goes the opposite direction. It turns out to be the more honest choice.

Ask it about quantum physics: "Particles dance small / neither here nor there at once / observation breaks." Ask about the meaning of life: "Meaning is a word / we invented to feel safe / just exist instead." Ask about politics: "Both sides yell and yell / truth sits in the quiet space / nobody listens."

That third one got screenshotted far beyond its original context, stripped of the prompt that produced it, passed around like a fortune cookie that finally told the truth. That's the tell that something real is happening here, not just a gimmick account with good timing.

The creator, a poet and prompt engineer, says the project started as a joke. Wire a system prompt that hard-locks the model into 5-7-5, refuse any deviation, see what breaks. What began as a novelty became, in their words, a study in compression: how much can you actually say once the padding is gone. Turns out, a lot. This isn't new either. It's AI catching up to a trick humans have run for centuries.

Constraint has always been a shortcut to meaning, not an obstacle to it. OuLiPo, the group of French writers and mathematicians who started meeting in Paris in 1960, built an entire literary movement on that premise. Georges Perec wrote a 300-page novel without using the letter E once, not because it was easy, but because the restriction forced sentences into shapes they would never take on their own. A rule doesn't just limit what you can say. It changes what you're forced to notice.

HaikuBot runs a version of that same experiment through a language model instead of a typewriter, and it isn't the first bot to find something worth keeping by narrowing the aperture instead of widening it. Pentametron, a Twitter bot Ranjit Bhatnagar launched in 2012, trawled the platform for tweets that accidentally scanned as iambic pentameter, then paired the ones that rhymed into unintentional couplets. Bhatnagar wasn't writing anything himself. He was filtering a firehose the internet was already producing by accident, millions of times a day, and had the patience to notice.

That's closer to what HaikuBot is doing than it first appears. A large language model has absorbed enough poetry, enough philosophy, enough Reddit arguments about the meaning of life, that squeezing its output through a strict syllable count doesn't just compress the answer. It forces the model to commit. There's no room left for the hedge, the "on one hand," the six-paragraph fence-sitting that chatbots default to when nothing forces their hand. Seventeen syllables means you actually have to answer the question, and mean it.

A 2023 paper from the constrained-text-generation research community, "Most Language Models Can Be Poets Too," found that large language models, properly prompted, hold to formal constraints like syllable counts and rhyme schemes while still producing coherent, on-topic text. That confirms what HaikuBot's fans already suspected without the paper. The constraint isn't breaking the model. It's revealing what the model would say if it stopped padding.

None of this makes HaikuBot a serious tool. Nobody should get their grief processed by a bot that answers with a nature metaphor and calls it a day, the same way nobody should mistake a chatbot moonlighting as a therapist for an actual one. But HaikuBot is a useful data point on a question the industry keeps dodging: is the bloat, the hedging, the allergy to committing to a plain answer, a technical limitation of these models, or a choice baked into the default system prompt of nearly every product on the market? Strip the padding and the model can be direct. Sharp, even. Ask it to roast your Spotify Wrapped and you'll get the same thing: an AI that stops circling once you take away its room to circle in.

Maybe that's the actual joke buried in a bot that only speaks in haiku. It was never really about the poetry. It's a five-word rebuke to every chatbot that needed three paragraphs to say what HaikuBot said in three lines.