Loading...
Loading...
Show Results
After seeing Automatic Donald Trump, I wanted to run it through a Turing test. This site is the result of that.
For every set of choices, one of the boxes above is a real tweet, and the other is a fake tweet generated by a Markov chain. Click the one you think sounds more "human". At any point during the test, click "Show Results" to end the test and show your results. To start over, click "Do it again".
Because why not.
A good explanation of Markov chains is provided at the Automatic Donald Trump site. But to put it simply, it's how predictive text works. Give it training data, and it will learn to emulate it. The Turing test is a test used to determine how "human" an AI is. For every set of choices, one choice is a real tweet (pulled from a database of 11000+ tweets), and one is generated by the Markov chain.
The current generator is a order Markov chain. The higher the order, the more Trump-esque the output is. However, this can lead to the output text copying the source text. Although that kind of behaviour is good for passing the Turing test, I feel that it violates the spirit of what I initially set out to do. So, I developed a version using Andrej Karpathy's "Magical" LSTM. An LSTM is a neural network that, given training data in the form of text, can generate text similar to (but not the same as) the training data. However, it didn't produce believable output, so I left the Markov chain.
version 0.2