August 22, 2026, 2:28 pm | Read time: 4 minutes
Many people likely had their first encounter with an AI system with the introduction of ChatGPT in late November 2022. However, there had been numerous attempts to test AI tools long before that. In March 2016, Microsoft launched its AI chatbot Tay on the platform Twitter, now X. TECHBOOK tells the story of why the AI experiment was terminated after just 16 hours in this article.
It was March 23, 2016, a Wednesday, when Microsoft publicly introduced its AI chatbot via Twitter. The chatbot was a female avatar communicating with the Twitter community under the name TayTweets.
“From a technical standpoint, Tay was an exciting experiment. However, it began producing problematic content relatively quickly, such as racism and hate speech,” recalls Anna Kruspe, a computer science professor at the University of Applied Sciences Munich (HM), who has been working with AI systems for years.
Microsoft responded to the unexpected radicalization of its AI chatbot in a disorganized manner. Apparently, no one on the responsible development team was prepared for such a situation. The reasons why Tay descended into the depths of hate speech remain speculative to this day.
No Comparison to Today’s AI Systems
“Microsoft never released very detailed information about Tay’s technical structure. At that time, the current architectures and technical prerequisites did not exist, so a different form of neural networks was likely used. These were significantly weaker in understanding longer conversation contexts,” Kruspe describes the technical differences from today’s systems.
In retrospect, it is clear how strongly Tay reacted to polarizing and problematic posts from the Twitter community. Essentially, a development that is now seen on many social media platforms. Emotional posts and extreme positions attract attention and reach. Even a machine like Tay was evidently not resistant enough to this.
“To better understand the AI chatbot’s reaction, it would be important for interpretation to know more about the technology behind it. If the input data is problematic, the risk is high that the output will also be problematic,” explains Anna Kruspe from the University of Applied Sciences Munich.
Today’s AI Systems Rely on Filters
Fortunately, AI research today has a better understanding of how to tame AI models. “Today’s chatbots are generally not retrained dynamically and unfiltered from user inputs, although chat histories are likely used for training future versions. However, today’s AI providers remain as secretive about the technical details as Microsoft did back then. It is very likely that today’s AI systems use automatic filters that respond to hate speech and other forms of radical expression,” the AI expert provides insight into the current state of development.
Filters initially sound good. However, they only respond to obvious extreme expressions, such as when someone openly broadcasts their hatred for everything and everyone. The issue remains with more subtle forms of racism or extremism.
“In research, we try, among other things, to uncover whether AI systems systematically portray certain population groups more negatively and to develop methods to address this. Such problems are much less noticeable at first than Tay’s obvious racism but can become dangerous in the long run,” emphasizes Anna Kruspe, a computer science professor at HM.
Jobs at Risk Due to Artificial Intelligence
What Can the Twitter Alternative “Threads” Do?
Tay Developers Underestimated the Potential of Hate Speech
Microsoft’s AI experiment with Tay ended after 16 hours and more than 96,000 tweets. The company attempted another launch just a week later, but it was halted even faster due to the same issues.
According to Anna Kruspe, this failed AI attempt from 10 years ago clearly demonstrated “what happens when AI providers underestimate developments.” Microsoft did not anticipate the destructive behavior of some Twitter users. Additionally, the company was unprepared for how strongly and quickly extreme posts could destabilize its AI system.
Today, much more is known about how extreme positions can drive an entire platform in a certain direction. Twitter, renamed X after being acquired by Elon Musk, is the most prominent example of this.
Also of interest: Why Politeness Costs OpenAI Millions of Dollars
Today, There Are Other Hurdles for AI Systems
Today’s AI system providers face various hurdles. A key aspect is keeping an eye on the enormous costs to meet the ever-growing demand for AI. While AI providers strive to maintain high-quality training data, this point remains lower on the priority list.
AI researchers like Anna Kruspe from the University of Applied Sciences Munich pay close attention to unwanted developments: “Today’s AI models learn from vast amounts of internet data, where biases are already embedded, albeit much more subtly. These can manifest, for example, in job applications or credit approvals, where certain groups are disadvantaged without anyone noticing.”
The Tay account from Microsoft on X still exists today. However, it has been silent since spring 2016.