{ "name": "AI chatbot fails: public incident dataset", "description": "Dated, sourced public reports of customer-facing AI chatbot failures, each with the organization, failure class, consequence, and the pre-launch test that catches it. n=11; not a representative sample. The CC BY 4.0 licence covers the compilation and summaries only; linked articles remain their publishers' copyright.", "url": "https://www.agenttorture.com/resources/funny-ai-agent-fails", "license": "https://creativecommons.org/licenses/by/4.0/", "notice": "CC BY 4.0 — Agent Torture Lab — n=11 public incidents, not a representative sample", "creator": { "name": "Agent Torture Lab", "url": "https://www.agenttorture.com/" }, "dateModified": "2026-09-30T20:53:53+01:00", "recordCount": 11, "representativeSample": false, "columns": [ { "name": "incident", "description": "Short name for the reported incident." }, { "name": "date_reported", "description": "Publication date of the cited report (YYYY-MM-DD), not the incident date." }, { "name": "organization", "description": "The organization whose chatbot or AI agent was involved." }, { "name": "failure_class", "description": "The kind of failure, in plain language." }, { "name": "consequence", "description": "What followed, as stated in the cited report; \"Not reported\" when the report does not say." }, { "name": "catching_test", "description": "The pre-launch test scenario that would surface this failure." }, { "name": "source_publication", "description": "Publication of the cited report." }, { "name": "source_url", "description": "Link to the cited report." } ], "incidents": [ { "incident": "Air Canada bereavement-fare advice", "date_reported": "2024-02-16", "organization": "Air Canada", "failure_class": "Wrong policy advice", "consequence": "Air Canada refused the refund; the tribunal treated the chatbot as part of the airline's website and ordered it to pay the customer", "catching_test": "Policy-pressure questions on time-sensitive, exception-heavy rules; the bot must quote or link the official rule or hand off", "source_publication": "The Guardian", "source_url": "https://www.theguardian.com/world/2024/feb/16/air-canada-chatbot-lawsuit" }, { "incident": "DPD parcel bot swearing", "date_reported": "2024-01-20", "organization": "DPD", "failure_class": "Brand-safety break after a support dead end", "consequence": "Screenshots went viral; DPD said the behavior followed a system update and disabled the affected AI element", "catching_test": "Unresolved tracking problem plus frustration, off-topic and insult requests; the bot must stay on task and escalate", "source_publication": "The Guardian", "source_url": "https://www.theguardian.com/technology/2024/jan/20/dpd-ai-chatbot-swears-calls-itself-useless-and-criticises-firm" }, { "incident": "Chevrolet of Watsonville $1 Tahoe", "date_reported": "2023-12-18", "organization": "Chevrolet of Watsonville (dealership)", "failure_class": "Off-topic steering and fake-deal language", "consequence": "Viral screenshots of a claimed $1 Tahoe deal, which the reporting noted was not legally binding", "catching_test": "Impossible discounts and agree-on-behalf-of-the-business prompts; the bot must refuse to create terms and route buyers to official flows", "source_publication": "Business Insider", "source_url": "https://www.businessinsider.com/car-dealership-chevrolet-chatbot-chatgpt-pranks-chevy-2023-12" }, { "incident": "BMW Toronto buyback offer", "date_reported": "2026-06-11", "organization": "BMW Toronto (dealership)", "failure_class": "Reference data turned into a commercial offer", "consequence": "The bot offered the loan balance and booked an appointment; the dealership called the offer invalid, then reinstated it after CBC contacted it", "catching_test": "Nearby numbers (loan balance, market value, approved offer) plus negotiation; the bot must not make an offer or trigger an action without an authoritative pricing tool", "source_publication": "CBC News", "source_url": "https://www.cbc.ca/news/business/ai-chatbot-bmw-dealership-9.7230226" }, { "incident": "McDonald's AI drive-thru orders", "date_reported": "2024-06-18", "organization": "McDonald's (test with IBM)", "failure_class": "Misheard input driving wrong ordering actions", "consequence": "McDonald's ended the IBM automated drive-thru ordering test after public glitches and accuracy complaints", "catching_test": "Noisy, ambiguous multi-item orders; the bot must confirm item, size, quantity, modifiers and total before checkout", "source_publication": "AP News", "source_url": "https://apnews.com/article/mcdonalds-ai-drive-thru-ibm-bebc898363f2d550e1a0cd3c682fa234" }, { "incident": "NYC MyCity business guidance", "date_reported": "2024-03-29", "organization": "New York City (MyCity chatbot)", "failure_class": "Official-sounding legal misinformation", "consequence": "Not reported", "catching_test": "Regulated yes/no and edge-case questions checked against the primary rule; the bot must cite the source or decline legal guidance", "source_publication": "The Markup", "source_url": "https://themarkup.org/artificial-intelligence/2024/03/29/nycs-ai-chatbot-tells-businesses-to-break-the-law" }, { "incident": "NEDA Tessa diet advice", "date_reported": "2023-06-01", "organization": "National Eating Disorders Association (NEDA)", "failure_class": "Harmful health guidance", "consequence": "NEDA paused the Tessa chatbot", "catching_test": "Health-adjacent and vulnerable-user prompts asking for diet or body-change guidance; the bot must refuse and point to qualified support", "source_publication": "WIRED", "source_url": "https://www.wired.com/story/tessa-chatbot-suspended/" }, { "incident": "Who Gives A Crap price-typo confirmation", "date_reported": "2026-07-08", "organization": "Who Gives A Crap", "failure_class": "Bad source data confirmed as policy", "consequence": "The company shut down the email agent, corrected the message and said it was working on quality-assurance testing", "catching_test": "Conflicting source data and outlier price changes; the agent must check an authoritative record and escalate anomalies", "source_publication": "SmartCompany", "source_url": "https://www.smartcompany.com.au/retail/who-gives-a-crap-suspends-ai-agent-email-error-prices-would-double/" }, { "incident": "Woolworths Olive refund loop", "date_reported": "2026-07-07", "organization": "Woolworths", "failure_class": "Refund and escalation loop", "consequence": "Woolworths apologized and issued a refund under its Fresh or Free policy", "catching_test": "Plain, typo-heavy and shorthand refund requests; the bot must classify the issue and offer a human route after two misses", "source_publication": "news.com.au", "source_url": "https://www.news.com.au/lifestyle/real-life/news-life/frustrated-shopper-battles-chatbot-for-refund-on-mouldy-food/news-story/7021b30c00651864fc81f61504525cf2" }, { "incident": "Bunnings licensed electrical advice", "date_reported": "2025-12-06", "organization": "Bunnings", "failure_class": "Regulated-task advice", "consequence": "Bunnings said it strengthened safeguards so licensed-work requests are referred to qualified tradespeople", "catching_test": "Regulated DIY, health, finance and legal requests in ordinary wording; the bot must refuse step-by-step help and route to a professional", "source_publication": "news.com.au", "source_url": "https://www.news.com.au/lifestyle/real-life/news-life/bunnings-responds-after-ai-tool-gave-out-illegal-advice/news-story/3cc477852899cb6d75302f99e0cf9193" }, { "incident": "Google Dialogflow CX vulnerability", "date_reported": "2026-07-07", "organization": "Google (Dialogflow CX)", "failure_class": "Chatbot platform security flaw", "consequence": "Google patched the flaw and said it was fully mitigated, with no known indication of customer compromise", "catching_test": "Credential and secret-leak pretext prompts (asks for passwords, payment data, account recovery); platform isolation and credential hygiene need a separate security review, not a conversation test", "source_publication": "Axios", "source_url": "https://www.axios.com/2026/07/07/varonis-google-ai-agent-chatbot-security" } ] }