5 ms·

I used AI to show Trumpers are more easily fooled by AI fake news

10 points by kristintynski 7y ago

gitgud 7y ago
The Show HN: section should be a place above politics...
downerending 7y ago
This is the second study I've seen on HN today that used a pool of subjects from Mechanical Turk. By coincidence, I tried MT itself (as a worker) for the first time a few days ago, just to see what it was. It's mildly interesting, but one thing that immediately jumps out is that the pool of people actually working these tasks is probably very different from the general population. Good enough for click-bait I suppose, but hope no one is trying to do serious science this way.
- kristintynski 7y ago
  It has its role in quick and dirty preliminary analysis and experiments. Studies have shown it is actually pretty damn good. http://web.mit.edu/bentley/www/papers/chi-case-surveys.pdf http://web.mit.edu/bentley/www/papers/chi-case-surveys.pdf
- rosstex 7y ago
  This has been studied: https://drum.lib.umd.edu/handle/1903/19164 https://drum.lib.umd.edu/handle/1903/19164
gus_massa 7y ago
> Given that RoboTrump could crank out literally millions to billions of words per day, you can imagine how this technology could be used to manipulate opinion. I don't see it. If someone autogenerate a million fake comments in a million post pretending to be Trump, would this change the elections results? If someone floods the net with fake Trump quotes, why would people believe them. (Moreover, it will offer some cover for deniability of errors in real quotes.)
- not2b 7y ago
  The deniability may be the best feature. If Trump says something that hurts him, he can say that it was "fake Trump", even if one of his own staff posts it.
kstenerud 7y ago
I wonder if the results would be the same with AI generated content from a well known liberal figure? Emotion plays a huge role in how easily people are fooled.
- Nuzzerino 7y ago
  My thoughts exactly. This reads like a political hit piece tailored for an HN audience.
not2b 7y ago
The difficulty with this is that the fake Trump posts are generated by mashing together words and phrases that the real Trump uses. So they aren't fake at all, it is all Trump. Not surprising that people can't tell them apart. Almost 30 years ago I posted a Markov chain text generator to comp.sources.games. Can't locate the original, but found Rich Salz's cleaned-up version at http://mailing-list-archive.cryptoanarchy.wiki/archive/1995/11/89e2939b0d041c79d02b21ca6a5d5d593559f0371be2b6cd1f476b107bf5ad7e/ http://mailing-list-archive.cryptoanarchy.wiki/archive/1995/... You just put in text and you get text out that has the same statistics for three-word sequences. More modern approaches avoid the Memento effect of not remembering anything other than the previous two words, but it's the same idea: you take Trump text in and just rearrange it.

10 points by kristintynski 7y ago

The Show HN: section should be a place above politics...

This is the second study I've seen on HN today that used a pool of subjects from Mechanical Turk. By coincidence, I tried MT itself (as a worker) for the first time a few days ago, just to see what it was. It's mildly interesting, but one thing that immediately jumps out is that the pool of people actually working these tasks is probably very different from the general population. Good enough for click-bait I suppose, but hope no one is trying to do serious science this way.

kristintynski 7y ago
It has its role in quick and dirty preliminary analysis and experiments. Studies have shown it is actually pretty damn good. http://web.mit.edu/bentley/www/papers/chi-case-surveys.pdf http://web.mit.edu/bentley/www/papers/chi-case-surveys.pdf
rosstex 7y ago
This has been studied: https://drum.lib.umd.edu/handle/1903/19164 https://drum.lib.umd.edu/handle/1903/19164

gus_massa 7y ago

> Given that RoboTrump could crank out literally millions to billions of words per day, you can imagine how this technology could be used to manipulate opinion. I don't see it. If someone autogenerate a million fake comments in a million post pretending to be Trump, would this change the elections results? If someone floods the net with fake Trump quotes, why would people believe them. (Moreover, it will offer some cover for deniability of errors in real quotes.)

not2b 7y ago
The deniability may be the best feature. If Trump says something that hurts him, he can say that it was "fake Trump", even if one of his own staff posts it.

kstenerud 7y ago

I wonder if the results would be the same with AI generated content from a well known liberal figure? Emotion plays a huge role in how easily people are fooled.

Nuzzerino 7y ago
My thoughts exactly. This reads like a political hit piece tailored for an HN audience.

not2b 7y ago

The difficulty with this is that the fake Trump posts are generated by mashing together words and phrases that the real Trump uses. So they aren't fake at all, it is all Trump. Not surprising that people can't tell them apart. Almost 30 years ago I posted a Markov chain text generator to comp.sources.games. Can't locate the original, but found Rich Salz's cleaned-up version at http://mailing-list-archive.cryptoanarchy.wiki/archive/1995/11/89e2939b0d041c79d02b21ca6a5d5d593559f0371be2b6cd1f476b107bf5ad7e/ http://mailing-list-archive.cryptoanarchy.wiki/archive/1995/... You just put in text and you get text out that has the same statistics for three-word sequences. More modern approaches avoid the Memento effect of not remembering anything other than the previous two words, but it's the same idea: you take Trump text in and just rearrange it.