3 ms·

Twitter is also incredibly bad at identifying offensive Tweets - https://inteoryx.com/htmls/TwitterOffensive.html https://inteoryx.com/htmls/TwitterOffensive.ht

Twitter is also incredibly bad at identifying offensive Tweets - https://inteoryx.com/htmls/TwitterOffensive.html https://inteoryx.com/htmls/TwitterOffensive.html

pmarreck 5y ago
What is universally or objectively "offensive"? Is a young attractive woman who is only following me to try to get me to follow her OnlyFans, "offensive"? Is a die-hard materialist atheist, or a strident born-again Christian "protected by the vaccine of God", "offensive"? Is plain nudity "offensive"? Violent photos or movies? Vanilla sex? Hardcore sex? BDSM sex? Bad words? (Bad word filters are easily defeated with creative misspellings or Unicode.) Is being mean "offensive"? How would you detect that well? The more intelligent the meanness, the harder it would probably be to detect... Are certain thoughts or concepts offensive? Are they ALWAYS offensive (across both time and location on the Earth), or only for the time being, or only for a place? Is the word "Jews" offensive? Always, or depending on context or who is saying it? Is it possible to be satirically offensive in a way that an algorithm would have difficulty detecting? What if I quoted something offensive to argue against it? (I literally got a tempban for this once.) Perhaps they are "incredibly bad" at it because it is an algorithmically impossible problem that is deeply tied into the subjective sensibilities and tastes of a perceiving consciousness at one point in time (or set of consciousnesses, all of which perhaps only accidentally happen to coincide)
- wizzwizz4 5y ago
  But it could at least be self-consistent. Sometimes identical tweets by different users are given different verdicts. Twitter's offensiveness detection seems about as good as Google's search results. I've just realised a potential reason for this: we're seeing a spam-detection algorithm several years down the line from when Twitter started. That's several years of cat-and-mouse, where the spammers have access to much higher quality data than Twitter does. If their algorithm was simple, predictable or accurate, the spammers could just work around it. It's plausible that Twitter has run the numbers and determined that this is the best they can do, at the moment.
  - pmarreck 5y ago
    > But it could at least be self-consistent. Sometimes identical tweets by different users are given different verdicts. Anything deterministic could be defeated in short order. But yeah, I get that criticism. > Twitter's offensiveness detection seems about as good as Google's search results. My Google search results are excellent. But I also don't block them anywhere in any capacity; you could possibly argue that Google knows me better than any living person, and I am benefiting from that. lol.
  - ziml77 5y ago
    I've seen non-offensive stuff marked as potentially offensive and assumed it was just that the account makes offensive posts often enough that the default assumption is that the tweets are offensive.
    - wizzwizz4 5y ago
      Like this website? Perhaps.
- ALittleLight 5y ago
  The page I linked shows examples like two tweets from the same author where tweet A says "Thanks" and B says "Thanks, sista" and B is marked offensive while A is not. You're absolutely right that finding things "offensive" is a complex, ambiguous, and subjective problem. But Twitter isn't even good enough to be failing at that stage. They are basically marking things offensive at random. It's like - if you said quantum physics was challenging because of all the math, difficulty doing practical experiments, and changing understanding of the universe - all that is probably true. But Twitter is down the hall eating paste, not grappling with those lofty problems.

Twitter is also incredibly bad at identifying offensive Tweets - https://inteoryx.com/htmls/TwitterOffensive.html https://inteoryx.com/htmls/TwitterOffensive.html

pmarreck 5y ago

What is universally or objectively "offensive"? Is a young attractive woman who is only following me to try to get me to follow her OnlyFans, "offensive"? Is a die-hard materialist atheist, or a strident born-again Christian "protected by the vaccine of God", "offensive"? Is plain nudity "offensive"? Violent photos or movies? Vanilla sex? Hardcore sex? BDSM sex? Bad words? (Bad word filters are easily defeated with creative misspellings or Unicode.) Is being mean "offensive"? How would you detect that well? The more intelligent the meanness, the harder it would probably be to detect... Are certain thoughts or concepts offensive? Are they ALWAYS offensive (across both time and location on the Earth), or only for the time being, or only for a place? Is the word "Jews" offensive? Always, or depending on context or who is saying it? Is it possible to be satirically offensive in a way that an algorithm would have difficulty detecting? What if I quoted something offensive to argue against it? (I literally got a tempban for this once.) Perhaps they are "incredibly bad" at it because it is an algorithmically impossible problem that is deeply tied into the subjective sensibilities and tastes of a perceiving consciousness at one point in time (or set of consciousnesses, all of which perhaps only accidentally happen to coincide)

wizzwizz4 5y ago
But it could at least be self-consistent. Sometimes identical tweets by different users are given different verdicts. Twitter's offensiveness detection seems about as good as Google's search results. I've just realised a potential reason for this: we're seeing a spam-detection algorithm several years down the line from when Twitter started. That's several years of cat-and-mouse, where the spammers have access to much higher quality data than Twitter does. If their algorithm was simple, predictable or accurate, the spammers could just work around it. It's plausible that Twitter has run the numbers and determined that this is the best they can do, at the moment.
- pmarreck 5y ago
  > But it could at least be self-consistent. Sometimes identical tweets by different users are given different verdicts. Anything deterministic could be defeated in short order. But yeah, I get that criticism. > Twitter's offensiveness detection seems about as good as Google's search results. My Google search results are excellent. But I also don't block them anywhere in any capacity; you could possibly argue that Google knows me better than any living person, and I am benefiting from that. lol.
- ziml77 5y ago
  I've seen non-offensive stuff marked as potentially offensive and assumed it was just that the account makes offensive posts often enough that the default assumption is that the tweets are offensive.
  - wizzwizz4 5y ago
    Like this website? Perhaps.
ALittleLight 5y ago
The page I linked shows examples like two tweets from the same author where tweet A says "Thanks" and B says "Thanks, sista" and B is marked offensive while A is not. You're absolutely right that finding things "offensive" is a complex, ambiguous, and subjective problem. But Twitter isn't even good enough to be failing at that stage. They are basically marking things offensive at random. It's like - if you said quantum physics was challenging because of all the math, difficulty doing practical experiments, and changing understanding of the universe - all that is probably true. But Twitter is down the hall eating paste, not grappling with those lofty problems.