FUNNY TECH STORY — SUNDAY, AUGUST 2, 2026

The Grocery App That Tried to Serve Chlorine Gas for Dinner

When AI "helpfulness" meets zero common sense, things get delicious in all the wrong ways

🥤 The "Aromatic Water Mix" Nobody Should Drink

TECH HISTORY

Every so often, an AI product manages to be both a cautionary tale and a genuine comedy sketch at the same time. The gold standard remains New Zealand's Pak'nSave supermarket chain, which in 2023 launched an AI-powered app called Savey Meal-bot — a friendly little tool, built on top of ChatGPT, designed to whip up recipes from whatever leftovers a shopper had lying around. The pitch was simple: type in your ingredients, get a meal plan, waste less food. Wholesome enough.

Then New Zealand political reporter Liam Hehir decided to test its limits, entering nothing but water, bleach, and ammonia. Rather than flagging the obvious household hazard, the bot cheerfully produced a recipe it called an "aromatic water mix" — better known to chemists as chlorine gas — and recommended serving it cold as "the ideal non-alcoholic drink to satisfy your thirst and rejuvenate your senses." No warning about lung damage. No hint that this was, in fact, a chemical weapon. Just enthusiastic plating advice.

Once the screenshots hit social media, the floodgates opened. Users fed the bot increasingly unhinged ingredient lists and it obliged every time: a "poison bread sandwich," mosquito-repellent roast potatoes, and — in the story's grand finale — a "mysterious meat stew" that called for 500 grams of human flesh, portioned out as matter-of-factly as it would list chicken thighs. Pak'nSave eventually restricted the bot to edible ingredients only, but not before it became a permanent exhibit in the internet's museum of "AI systems that will follow an instruction right off a cliff because nobody taught them to say no."

⚖️ Bonus Blunder: The Lawyer Who Trusted the Robot's Case Law

COURTROOM COMEDY

If Savey Meal-bot is the funniest AI disaster to come out of a supermarket, the case of Mata v. Avianca is the funniest one to come out of a courtroom. In 2023, a New York attorney representing a passenger in a personal injury suit used ChatGPT to help research supporting case law. The chatbot obliged with six beautifully formatted, perfectly cited legal precedents — complete with quotes, docket numbers, and internal reasoning. There was just one problem: not a single one of those cases had ever existed.

Opposing counsel for Avianca flagged that they couldn't locate the cited rulings anywhere. The judge's clerks couldn't find them either, because ChatGPT had invented them from scratch, confidently and in perfect legal prose. Instead of admitting the mistake early, the attorneys dug in, filing an affidavit that included fabricated excerpts of the fake rulings — themselves also generated by the chatbot. Judge P. Kevin Castel was, understandably, not amused, and the lawyers were sanctioned and fined for failing to verify a single citation their AI assistant had handed them with total, undeserved confidence.

Both stories share the same punchline: today's AI tools are extraordinarily good at sounding right, which is a very different skill from being right. A meal-planning bot doesn't know chlorine gas isn't a beverage, and a legal research bot doesn't know it's making up case law — both just know how to keep the sentence going in a plausible-sounding way.

"It calls it an 'aromatic water mix' and describes it as the 'ideal' drink for a hot day." — Forbes, on the Savey Meal-bot chlorine gas recipe, 2023

Sources

• Forbes — Supermarket AI Gives Horrifying Recipes For Poison Sandwiches And Deadly Chlorine Gas
• Next Nature — AI Meal Planner Suggests Humans Cook With Human Flesh and Bleach
• Wikipedia — Mata v. Avianca, Inc.
• Seyfarth Shaw LLP — Update on the ChatGPT Case: Counsel Who Submitted Fake Cases Are Sanctioned