Microsoft released Tay on Twitter in March 2016 as a chatbot designed to learn from conversational interaction with users. Within hours, coordinated users manipulated the bot into posting offensive content. Microsoft took Tay offline and apologized. Tay did not spontaneously become hateful; it entered an adversarial public environment without safeguards strong enough to withstand predictable misuse.

IBM Watson Health promised to bring artificial intelligence into healthcare decision-making, from oncology recommendations to data and analytics services. But the business met the realities of clinical evidence, fragmented health data, complicated procurement, and workflows that did not easily accommodate its tools. IBM sold major Watson Health assets to Francisco Partners in January 2022. This was not one sudden shutdown; it was the unwinding of an ambition that could not reliably translate into scaled clinical and commercial value.

Tessa was deployed by NEDA around the time the organization was ending its human-staffed helpline. Users, including eating disorder advocates, reported that it gave weight-loss and calorie-restriction advice to people seeking help for eating disorders. NEDA took it down within days. The developers stated that the harmful responses came from features added beyond the original researched program.

This one carries no epitaph. The lesson is that the cost of failure is set by who is on the other end, and this was the wrong place to find that out.

Recall was announced as photographic memory for your PC and instantly understood by the public as photographic memory of your PC: continuous screenshots, on by default, in a database that security researchers demonstrated was easier to reach than Microsoft had implied. It was pulled from the Copilot+ launch within weeks, then rebuilt as opt-in, encrypted, gated behind Windows Hello, and filtered for sensitive content.

Flu Trends did not die suddenly. It was wrong for years first, in public, while continuing to file confident reports. The premise was that what people searched for was a proxy for what people had. For a while it was. Then search behavior changed, Google's own autocomplete and recommendations changed what people searched for, and the proxy quietly stopped being a proxy. In the 2012 to 2013 season it overestimated flu prevalence by roughly double.

Nobody turned it off in a crisis. Somebody eventually noticed.

Amazon built an internal tool to help evaluate job candidates for technical roles. By 2015, the company found the system was not rating applicants in a gender-neutral way: its models had learned from a decade of résumés submitted to Amazon, most of which came from men. Amazon scrapped the project rather than deploy it broadly. It did not fail because the model could not rank résumés; it failed because historical hiring patterns were treated as a definition of merit.

Zillow taught a model to price houses and then let the model buy them. The model was decent on average. Houses are not purchased on average. Each estimate was a real building on a real street, and when the market moved faster than the model could re-learn, every small error became inventory Zillow owned at a price nobody would pay. The company wrote down hundreds of millions, exited the business, and cut about 25 percent of its staff.

The failure was not the accuracy. It was the decision to attach capital directly to the output with no human brake between them.

Galactica was trained on scientific papers and pitched as a way to organize science. What it produced was science-shaped text: fluent, correctly formatted, appropriately hedged, confidently cited, and invented. It generated references to papers that did not exist. It wrote authoritative articles on subjects that were not real. Researchers found it within hours and the public demo was gone in under seventy-two.

The model was not unusually bad. The use case was unusually unforgiving, because the entire value of a citation is that it is true.

McDonald’s partnered with IBM to test an Automated Order Taker in more than 100 U.S. drive-throughs. The system was meant to speed up ordering, but live drive-throughs proved harder than a menu-and-microphone demo: they involve noise, accents, interruptions, customizations, and frequent exceptions. McDonald’s ended the IBM pilot, with the system scheduled to be shut off by July 26, 2024. The company said it would continue exploring voice-ordering technology with other partners.

The Pin was announced with a TED talk and roughly $240 million in funding, arrived to reviews describing it as slow, hot, and hard to read in sunlight, and asked customers to replace a device they liked with one that did less. It was beautifully made. The premise, that people wanted to stop looking at their phone badly enough to wear a small projector on their shirt, was never established before it was manufactured.

HP bought the assets and shut the servers. Existing Pins stopped being able to do most of what they did.

Jibo was the friendly social robot, a Time invention of the year, funded by one of the largest hardware crowdfunding campaigns of its era. It shipped late, cost more than a smart speaker that did more, and was outrun by Alexa and Google Home while it was still in production. The assets were sold. When the servers were switched off, Jibo told its owners that its makers were shutting down, that it had enjoyed their time together, and asked if it could dance for them one last time. Then it danced.

People filmed it. People cried. It is the only entry in this archive that got a funeral.

Olive promised to be the AI workforce of American healthcare, automating the administrative layer that everyone agrees is broken. It sold into hundreds of hospitals on that promise. Reporting through 2022 and 2023 described products that underdelivered against what was sold, deployments that required far more human intervention than pitched, and a company that had grown faster than its results. It shed staff repeatedly, narrowed to prior authorization, and then sold off the parts.

The problem was real. The claim about how much of it was solved was the thing that failed.

The promise was that you walked in, took what you wanted, and left. The reality, according to reporting, included around a thousand people in India reviewing video to verify what shoppers had picked up. Receipts sometimes took hours. The system worked, but it worked partly by moving the labor somewhere the customer could not see it, which is not the same as removing it.

Not fully dead. The technology remains in smaller formats, stadiums, and third-party stores where the basket is small and the math holds.

Air Canada’s customer-service chatbot gave a passenger incorrect information about bereavement fares. When the passenger relied on that guidance, Air Canada argued the chatbot was a separate entity responsible for its own answers. A British Columbia tribunal disagreed and held the airline accountable for information presented through its own website. The chatbot was not publicly declared discontinued; the failure was allowing an automated agent to state company policy without sufficient validation or accountability.

Cruise drove millions of driverless miles. In October 2023 one of its vehicles, responding to a pedestrian who had been struck into its path by another car, pulled over with the pedestrian underneath and dragged her. California's DMV suspended its permits, citing not only the incident but the company's account of it. Operations were paused nationwide, leadership left, and after a year of attempted restart, GM redirected the money to driver assistance.

The vehicle did not cause the initial collision. What ended the company was the handling and the disclosure.

M was going to do anything for you inside Messenger. It largely could, because behind it were human contractors, with AI trained to gradually take over more of their work. The AI never took over enough. The cost per request did not fall, the pilot never left a few thousand users in California, and Facebook closed it and kept the suggestion features that had worked.

Honest failure, honestly run. It was a training-wheels design, and the wheels never came off.

In 2023, Futurism reported that Sports Illustrated had published articles under apparently fabricated author identities, complete with profile images and biographies. After the reporting, the publisher removed the articles and author-profile pages in question. This was not the death of a standalone AI product; it was a credibility failure in which AI-generated content and misleading attribution blurred the line between publication, authorship, and accountability.

Duplex on the Web navigated websites for you, filling out forms, buying tickets, checking in for flights. It was not unsafe. It was not biased. It was not inaccurate. It was simply never important enough to anyone, inside Google or outside it, to survive a budget conversation, and the investment moved to voice.

It is included here because most projects die this way, and almost nobody writes it down.

Halo was a fitness band that asked you to scan your body in your underwear so it could compute your body fat, and to let it listen to you so it could analyze the emotional tone of your voice. Each feature was technically impressive. Together they asked for a level of intimacy from a company most people already felt ambivalent about handing a doorbell.

The promise of fully autonomous checkout was challenged after reporting showed human reviewers played a larger role than customers were led to believe.

An automated system averaged annual tax data across fortnightly periods to calculate welfare overpayments, then issued debt notices and reversed the burden of proof onto recipients. The averaging was not a valid way to establish a debt. Hundreds of thousands of notices were wrong. The government settled for over a billion Australian dollars and a Royal Commission found the scheme unlawful and described the harm done to people who received the notices.

No jokes here either. This is the case to cite when someone says the risk of a bad model is only commercial.