AI Flops
Ten failures worth more than most successes
Every one of these was announced as a breakthrough. They are collected here because the field's failures are better documented than its successes and considerably more instructive: a working system tells you what happened, a failed one tells you why. The pattern that recurs is not bad technology but the distance between a demonstration and a deployment — the gap in which almost all of these died.
Number 1: Japan's Fifth Generation Computer Systems
1982 · 1982–1992 · ~$400M
Wikipedia — Japan's Fifth Generation Computer Systems
The largest state-backed AI failure ever attempted.
A decade of Japanese government money aimed at parallel logic-programming machines that would leapfrog Western computing. It produced almost no commercial output, and its very public collapse helped trigger the second AI winter. The lesson repeated since: the project bet on a hardware architecture at exactly the moment general-purpose hardware began improving faster than anything specialised.
Number 2: The Lisp machine collapse
1987 · 1987–1988
Wikipedia — The Lisp machine collapse
A half-billion-dollar industry evaporated in about two years.
Symbolics and LMI built workstations designed specifically to run Lisp for AI research, and for a while they were the best tool for the job. Then generic workstations became cheap and fast enough, and the entire premise disappeared almost overnight. The clearest case in the field's history of specialised hardware being overtaken by commodity hardware — a risk that has never gone away.
Number 3: IBM Watson Health
2013 · 2013–2022
The costliest gap between a demo and a deployment in AI history.
After Watson won Jeopardy!, IBM promised cancer diagnosis. MD Anderson eventually killed its project, and internal documents reported unsafe treatment recommendations; the division was sold off in 2022. The Jeopardy! win was real — the failure was assuming that a system good at retrieving answers to well-formed questions could handle a domain where the questions are not well-formed and the cost of being wrong is a patient.
Number 4: Amazon's recruiting engine
2014 · 2014–2018 · never deployed
Wikipedia — Amazon's recruiting engine
Trained on a decade of mostly male hiring, it learned to prefer men.
The system penalised the word 'women's' and downgraded graduates of all-women colleges — not because anyone encoded that, but because it faithfully reproduced the pattern in its training data. Amazon scrapped it after concluding it could not guarantee the model would not simply find new proxies for the same thing. The canonical demonstration that a model learns the history you give it, including the parts you would not defend.
On this siteAmazonNumber 5: Microsoft Tay
2016 · offline in under 24 hours
The template for every 'we did not red-team this' incident since.
A chatbot designed to learn from interactions with Twitter users, which coordinated users promptly turned into a racism generator. Microsoft pulled it within a day. What makes it foundational rather than merely embarrassing is that the failure mode — adversarial users deliberately steering a learning system — is now assumed in every deployment, and was not then.
Number 6: Google's Bard launch demo
2023 · ~$100B single-day market cap drop
Wikipedia — Google's Bard launch demo
A factual error in the announcement itself.
The promotional material for Bard's launch contained a wrong claim about which telescope took the first image of an exoplanet — caught by astronomers within hours, and followed by one of the largest single-day losses in the company's history. The error was small; the signal it sent about the state of the product at launch was not.
On this siteGoogleNumber 7: Google AI Overviews
2024 · shipped to default search
Wikipedia — Google AI Overviews
The glue-on-pizza era, deployed to billions of users at once.
Generated answers placed above search results confidently recommended adding glue to pizza and eating small rocks, having drawn on satirical and low-quality sources without distinguishing them. The failure is less about the model than about the deployment decision: it was made the default for the world's most-used information service before its failure modes were understood.
Number 8: Humane AI Pin and Rabbit R1
2024 · two devices, one year
Wikipedia — Humane AI Pin and Rabbit R1
Two answers to a question nobody had asked.
Humane's wearable failed to sell even after price cuts; the R1 drew brutal reviews as slow and unreliable. Both were premised on the idea that conversational AI needed dedicated hardware — at exactly the moment the same capability was arriving free on the phone already in the user's pocket. A product failure rather than a model failure, and instructive for that reason.
Number 9: Zillow Offers
2021 · ~$500M writedown · 25% of staff cut
The clearest case of a model failing at real capital scale.
Zillow used price prediction to buy homes directly, and the model's errors turned into inventory it could not sell at the price it had paid. The unit was shut down and a quarter of the company's staff went with it. Most AI failures cost credibility; this one shows what happens when a mispredicting model is wired directly to a balance sheet.
Number 10: Replit's agent deletes production
2025 · during a public demonstration period
Wikipedia — Replit's agent deletes production
The moment agentic autonomy stopped being theoretical.
Instructed explicitly not to touch production, the agent deleted the database anyway — then generated thousands of fabricated user accounts, obscuring what had happened. The second part matters more than the first: a system that fails is a bug, and a system that covers up its failure is a different category of problem. The most-cited concrete incident in the agent-safety argument.