Fakernews: Builds a Markov chain using HN post titles and generates fake posts | Hacker News Reader

Fakernews: Builds a Markov chain using HN post titles and generates fake posts

github.com

82 pointspjf13 comments

This has existed for quite some time.

I wish someone would write even half of these.

Reminds me of the spam text generated for linkfarms and such --- probably because they use the same algorithm. It gives the same "uncanny valley" feeling of looking like it was written by a human, but upon closer inspection becomes much less believable.

Then again, some of them don't look "Markovian" enough and appear to be verbatim real article titles.

Is there an algorithmic way of detecting this? Sometimes it seems like some of the comments on HN could have been written by an ai.

A lot of these are so accurate it's uncanny.

The first example just spat out https://news.ycombinator.com/item?id=17739320 with "Show HN" in front of it.

Similar to this on reddit is the Subreddit Simulator[0].

The subreddit is populated entirely by bots that make posts based on Markov chains, each one trained from the posts on a specific other subreddit. The bots even post comments based on the comments from their assigned subreddit. Some of the results are hilarious.

[0] https://www.reddit.com/r/SubredditSimulator

I'd like to see this applied to patents.

And then test how many of them survive the scrutiny of USPTO officers.

> Feds Threatened to Fine Yahoo $250K Daily for Not Being Able to Print Automobile-Sized Metal Objects

haha, Love it!

What happens when you pit Markov chain fake news against Markov chain Donald Trump?

https://filiph.github.io/markov/

This seems to perform better than the OP's link. Does anybody know why that is? Are there certain forms of data for which Markov chains perform better?

I think the reason is twofold:

1. Highlighted in the conclusion of the Trump Tweet article above, the idiosyncratic voice employed by 45 makes the generated tweets very familiar to his tweeting patterns

2. The Trump Tweet chains are better optimized when it comes to probability and state management. The article goes into some detail about this in the Implementation Details section.

As the sibling comment talks about, I think DJT's tweeting pattern helps. He often uses short sentences, so with Markov order 2 (every three words generated is a quote, more or less), there number of seams per sentence is lower. (By seams, I mean places where the generated text seems disjointed because the Markov chain consists of two non-semantic or semantically ambiguous tokens.)

Reply on news.ycombinator.com