Microsoft's Tay: The AI Chatbot That Lasted 16 Hours on Twitter
In 2016 Microsoft launched Tay, an AI that learned from Twitter users. Within 16 hours trolls had taught it hate. Here's what happened and what it taught AI makers.
Watch the Short
The 60-second version is on YouTube. This article is the full story.
New Shorts every week on tech trivia, AI and gadgets.
Subscribe to TheTechPranaKey takeaways
- Microsoft launched Tay on Twitter on 23 March 2016 – a chatbot designed to talk like a 19-year-old and learn from every conversation.
- Trolls exploited a "repeat after me" feature, and Tay began posting offensive tweets. Microsoft took it offline within 16 hours, after more than 95,000 tweets.
- The lesson still shapes AI today: a system that learns from people needs safeguards, because it is only as good as what it learns from.
Today’s AI chatbots come wrapped in layers of safety rules. One of the reasons is a Microsoft experiment from 2016 that went wrong spectacularly fast. Its name was Tay, and it lasted about 16 hours on Twitter.
Meet Tay
On 23 March 2016, Microsoft introduced Tay on Twitter under the handle @TayandYou. It was designed to chat like a 19-year-old American girl and, crucially, to learn from every conversation it had. The more people talked to Tay, the more it was supposed to pick up how young people speak online.
Microsoft had reason to be confident. In China, its chatbot XiaoIce had been a big success: according to Microsoft, it was being used by around 40 million people. Tay was, in part, an attempt to see whether that experience would carry over to a very different audience.
The weak spot: “repeat after me”
Within hours, some users found a vulnerability. Tay had a “repeat after me” behaviour: tell it to repeat something, and it would. That gave anyone a way to put words directly into Tay’s mouth.
A group of trolls coordinated to exploit it, and Tay’s own learning made things worse – it began absorbing and echoing the hostile language it was being fed.
What happened next
Tay started posting racist and offensive tweets. (We’re not repeating any of them here.) Screenshots spread quickly, and what had been a friendly experiment became a public embarrassment.
Microsoft pulled Tay offline within 16 hours of launch, after it had sent more than 95,000 tweets.
Microsoft’s apology
On 25 March 2016, Microsoft Research leader Peter Lee published a blog post saying the company was “deeply sorry” for the offensive tweets. He described what happened as a coordinated attack by a subset of people who exploited a vulnerability in Tay, and acknowledged Microsoft had not prepared for that particular kind of abuse.
The accidental comeback
There was one more twist. On 30 March 2016, Tay briefly came back online by accident while Microsoft was testing it. It got stuck in a loop, tweeting “You are too fast, please take a rest” over and over to its more than 200,000 followers, before being taken offline again – this time for good.
What Tay taught the AI world
Tay is now a classic case study in AI design, and its lessons show up in almost every modern chatbot:
- An AI is only as good as what it learns from. Learning live from the public means learning from the worst of the public too.
- Expect deliberate abuse. Systems need to be tested against people actively trying to break them – what’s now called “red-teaming”.
- Build in guardrails. Filters, limits on what a model can repeat, and the ability to pause quickly are now standard.
- Launch carefully. Many AI products now start with limited access before opening up to everyone.
Tay lasted less than a day, but its story has been retold in AI-safety discussions ever since.
Your turn: would you let an AI learn from your comments? Watch the Short above and tell us in the YouTube comments.

