9 votes

Florida State University and Stanford are developing an "online polygraph" that detects lies in text — without the contextual clues that can hint at deception in a face-to-face conversation.

5 comments

  1. [2]
    Maven
    Link
    A skilled human lie detector only gets it right slightly more than half the time. I'm pretty sure this is just a bad training set or some similar error.

    A skilled human lie detector only gets it right slightly more than half the time. I'm pretty sure this is just a bad training set or some similar error.

    4 votes
    1. ntgg
      Link Parent
      The data set seems to be ~7000 words from some sort of game. I don't know if a game where you are trying to lie would give the best results compared to actual lying. I have no idea how else one...

      The data set seems to be ~7000 words from some sort of game. I don't know if a game where you are trying to lie would give the best results compared to actual lying. I have no idea how else one would collect data about lies at scale, but I think that the results of a game are going to be noticeably different than real life.

      5 votes
  2. [3]
    hereticalgorithm
    (edited )
    Link
    The SciHub version of this paper doesn't have the figures in it, got a friend with access to send me a copy. And here's the appendix with some of the tables that didn't make it to print. Looking...

    The SciHub version of this paper doesn't have the figures in it, got a friend with access to send me a copy.

    And here's the appendix with some of the tables that didn't make it to print.

    Looking over it, they note that their models only had two statistically significant cues. The first is word count - presumably lies take more words to explain than the truth which is actually negatively correlated with deception. The second is a calculated property called "insight", measured by the use of words like "think, know, consider" which are supposed to indicate certain cognitive proccesses (maybe because lies require more thinking to come up with?).

    4 votes
    1. [2]
      Fire
      Link Parent
      That makes perfect sense, really. I would think intentional lying comes with a factor in unknowingly overcompensating, fillers, meandering around - even textually. Very interesting. I'm...

      That makes perfect sense, really. I would think intentional lying comes with a factor in unknowingly overcompensating, fillers, meandering around - even textually. Very interesting. I'm downloading that paper now. Nice score!

      2 votes
      1. hereticalgorithm
        Link Parent
        Hey, just letting you know that I checked out the table in the appendix again and it turns out that word count was negatively correlated with deception. Not sure how what makes that the case (I'd...

        Hey, just letting you know that I checked out the table in the appendix again and it turns out that word count was negatively correlated with deception. Not sure how what makes that the case (I'd have to look more into the setup to figure out why - looks like a lot of these questions seem to concern personal experiences).

        1 vote