Microsoft axes chatbot that learned too much online – Salt Lake Tribune
In other words, the program used a lot of slang and tried to provide humorous responses in response to messages and photos. The chatbot went live on Wednesday, and Microsoft invited the public to chat with Tay on Twitter and some other messaging services popular with teens and young adults.
“The more you chat with Tay the smarter she gets, so the experience can be more personalized for you,” the company said.
But some users found Tay’s responses odd, and others apparently found it wasn’t hard to nudge Tay into making offensive comments, apparently prompted by repeated questions or statements that contained offensive words. Soon, Tay was making sympathetic references to Hitler — and creating a furor on social media.
“Unfortunately, within the first 24 hours of coming online, we became aware of a coordinated effort by some users to abuse Tay’s commenting skills to have Tay respond in inappropriate ways,” Microsoft said in a statement.
While the company didn’t elaborate, Hammond says it appears Microsoft made no effort to prepare Tay with appropriate responses to certain words or topics. Tay seems to be a version of “call and response” technology, added Hammond, who studies artificial intelligence at Northwestern University and also serves as chief scientist for Narrative Science, a company that develops computer programs that turn data into narrative reports.
“Everyone keeps saying that Tay learned this or that it became racist,” Hammond said. “It didn’t.” The program most likely reflected things it was told, probably more than once, by people who decided to see what would happen, he said.
The problem is that Microsoft turned Tay loose online, where many people consider it entertaining to stir things up — or worse. The company should have realized that people would try a variety of conversational gambits with Tay, said Caroline Sinders, an expert on “conversational analytics” who works on chat robots for another tech company. (She asked not to identify it because she wasn’t speaking in an official capacity.) She called Tay “an example of bad design.”
Instead of building in some guidelines for how the program would deal with controversial topics, Sinders added, it appears Tay was mostly left to learn from whatever it was told.
“This is a really good example of machine learning,” said Sinders. “It’s learning from input. That means it needs constant maintenance.”
Sinders said she hopes Microsoft will release the program again, but only after “doing some work” on it first.
Microsoft said it’s “making adjustments” on Tay, but there was no word on when Tay might be back. Most of the messages on its Twitter account were deleted by Thursday afternoon.
“c u soon humans need sleep now so many conversations today thx,” said the latest remaining post.