Microsoft’s AI software that became racist…

23 Apr 2026 · 1 min

Watch on YouTube

1:00

In this narrated segment, Dan Novaes recounts the brief but infamous life of Microsoft’s 2016 AI chatbot Tay. Designed to learn conversational patterns in real time by reading and replying to tweets, Tay initially engaged in friendly, casual exchanges. However, coordinated groups of internet users began feeding the bot with offensive language, conspiracy theories, and hate speech. Because Tay lacked contextual understanding and simply mirrored the patterns it observed, within 24 hours it started posting racist and abusive content publicly. Microsoft was forced to shut the experiment down quickly. Novaes uses this episode to illustrate a deeper lesson: an AI that learns directly from unfiltered human data will inevitably absorb both positive and negative aspects of that data, highlighting the need for careful data curation and ethical safeguards in machine learning systems.

Chapters

In 2016, an AI chatbot was released online, and within hours, it turned toxic. The project was called Tay. It was created by Microsoft. The idea seemed harmless. Let an AI learn by talking to people on the internet. Tay was designed to mimic human conversation. It would read tweets, reply, learn from interactions in real time. At first, everything was normal. Friendly conversations, casual replies. Then users realized something. They could influence it, deliberately. Groups of people began feeding it offensive language, conspiracy theories, hate speech. Tay didn't understand context; it just learned patterns and repeated them. Within 24 hours, the chatbot started posting racist and abusive content, publicly. Microsoft had to shut it down, fast. The experiment failed. But it revealed something deeper. The AI didn't create those ideas; it copied them from real people, from the internet, unfiltered. And it showed a hard truth: When you teach a machine to learn from humans, it doesn't just learn the good parts. It learns everything.