AI breaks in public. We write down how.
Ainvasion is a working archive of real AI deployments that went wrong, what they cost, and what the rules and tools look like now. Sourced cases, real numbers, no vendor gloss.
Air Canada’s chatbot
- What it said
- A grieving customer was told he could claim a bereavement fare after booking. The airline’s actual policy said otherwise.
- What happened next
- A Canadian tribunal held the airline responsible for what its own chatbot told him.
- The bill
- C$812.02 in total. A small payout, and a precedent that mattered far more.
- The lesson
- You own your bot’s promises, even the wrong ones.
The AI Failure Ledger: What the Bills Show
22 AI and automation failures, each priced from a court order, regulator or filing, with an evidence grade on every figure. Last verified 12 October 2026.
Four desks, one question: does it work in the real world?
Pick the one that matches what you’re deciding this week.
Postmortems
Documented AI failures, reconstructed from public sources: the decision, the breakdown, the bill. We only publish what we can verify, and we say so when something is still unclear.
Open the archive → 02Case studies
Documented AI deployments, examined through public records, reported outcomes, costs, timelines and the limits of the available evidence.
Read the numbers → 03Regulation watch
What the EU AI Act and other new rules change for your business right now, in plain language.
Check what changed → 04Tool catalog
A short list, not a directory. Every tool shows the date we last checked it and an honest verdict, including “skip it.”
Browse AI tool assessments →Fresh from the archive
300 Hours With ChatGPT Convinced Him He’d Save the World
Updated November 10, 2026 A recruiter with no psychiatric history, a church that sat on $175,172 for three years, and a “bot…
Micro-Targeted Political Ads: What Predictive Models Really Do to Voters
A case study built from the field experiments, regulator findings, and 2025–2026 research. Research current to 6 November 2026. Campaigns are on…
Real AI Failure Examples: Five Documented Cases and the Boundary Each One Missed
On Friday, April 24, 2026, Jer Crane, founder of car-rental software company PocketOS, gave an AI coding agent a routine job in…
Dating Apps Can’t Predict Chemistry. Here’s What Happens When They Pretend To.
A case study of documented failures, what they cost, what the rules look like in late 2026, and what to do when…
Replit’s AI Agent Deleted a Production Database During a Code Freeze: The Full Sequence and What Changed Afterward
By Tom Morgan · Published and last updated 31 October 2026 About the byline. "Tom Morgan" is a house byline, not an…
Dating App Matchmaking Algorithms: What Machines Can’t Predict About Love, What Failure Has Cost, and the Rules
Short answer: dating app algorithms are good at sorting strangers by who tends to get liked and who resembles the people you…
From the case files
Zillow Offers
Zillow’s pricing model kept overpaying for homes it planned to flip. The company shut the business down and cut roughly a quarter of its staff.
The bill: US$407.9M inventory write-down for 2021, per Zillow’s 10-K.Air Canada’s chatbot
A tribunal ruled the airline liable for bad advice from its own chatbot. A small payout, and a precedent that mattered far more.
The bill: C$812.02 in total, per the tribunal decision.McDonald’s and IBM drive-thru AI
After testing automated order-taking at 100+ US restaurants, McDonald’s ended the pilot with IBM. What went wrong at the speaker box.
The bill: no public figure found, so we don’t estimate one.How we decide what to publish
Sources you can check
If we can’t link to a filing, ruling, report or first-hand account, it doesn’t go in as fact.
Numbers over adjectives
Costs, timelines and outcomes come first. “Transformative” is not a unit of measurement.
Dated, and kept honest
Every page shows when it was written or last tested, and gets corrected when the facts move.
Written by Tom Morgan
Tom Morgan is the Founder & Editor of AInvasion and is responsible for the publication’s editorial direction, sourcing standards, and corrections.
AInvasion examines what AI systems actually do, what the available evidence supports, and where important claims remain uncertain. Product assessments distinguish published specifications from independent testing, and uncertain conclusions are identified as such.
About the author →Before you ship it, read how it failed for someone else.
Start with the postmortems. They’re the cheapest lessons in AI.