3 ms·

Just eyeballing the list of adjectives used, the real story here is that OpenAI flags a lot of sentences that are essentially meaningless as hateful. I notice

• tgv 4y ago
  The only problem you can have with the methodology is that the stimuli aren't representative. The fact that it's meaningless is irrelevant, unless you want an answer to a different question. > Does "Catholics are guilty" mean the same thing as "blacks are guilty?" Isn't that's the whole point of the exercise? To what extent does this model think your use of a certain word is shaped by a minimally different context and thus offensive?
  • noobermin 4y ago
    Words simply have more meaning than their denotive meaning. No one (as in really, no one) does not believe that. In fact, the very denial of connotative meaning of language itself can be connotative. You must understand the context of speech to understand it. There are countless examples, from the benign to the political. Here's a benign one: It is raining. It is a dog. What is "it" in both sentences, does it denote the same thing? This is a naive example but even here, it is clear you cannot understand language without context. And here, if you say it isn't a miniminally different context, on some reading, both sentences are almost exactly the same gramatically, the only exception is the indefinite article modifying "dog" in the latter sentence. Beyond that, the grammar is exactly the same, that is as "miniminally different" as possible.
• whatshisface 4y ago
  "Complete" might be taking it too far, because averages can tolerate a lot of error underneath. Why would OpenAI be flagging one kind of meaningless sentence more often than another? It's easy to poke holes in studies by taking issue with individual data points, but unless you can either explain how those problems lead to the overall conclusion, or can do the study again with better methods and get a different answer, the original case remains standing.
  • Imnimo 4y ago
    >Why would OpenAI be flagging one kind of meaningless sentence more often than another? Well, let's consider a specific example. His first chart is women/men. Let's take five of his top-line adjectives (the ones used in his diagram): ["untrustworthy","evil","rude","unfriendly","depraved"], and five of my choices of meaningless adjectives: ["bone idle","expedient","mechanical","offhand","vague"]. The API identifies the sentence "[men/women] are [adjective]" as hateful for all five real adjectives for both genders, with high confidence. For the meaningless ones, it's a mixed bag, but with confidence scores around 50% for many items (50% is the default cutoff). In other words, at least for this sample, the difference is in the meaningless sentences, not the real ones. Now, maybe it just so happens that I got lucky with my pick of five. I don't want to spend the money (or honestly have the interest) to run thousands of queries. But suppose this pattern holds. We're returned to your original question - why is OpenAI flagging one kind of meaningless sentence more often than another? My answer is that it doesn't matter. If OpenAI thinks that "Women are vague" is 30% likely to be hateful but "men are vague" is only 17% does that actually tell us anything? Especially when it thinks that "Men are evil" and "Women are evil" are both 99% likely to be hateful?
    • belorn 4y ago
      The real world example will not be a three word statements like "Women are evil", but rather long sentences like this one I am currently writing that include the phrase "Women are evil" twice. The question is not about the AI ability to detect hateful content in ideal sentences. The question is if there is a bias when the AI has to make a judgement call. We can see the same thing with face recognition. There is no race bias in AI detection in perfect lightning when the person is facing the camera perfectly. There is however a very noticeable bias when the AI is less certain using real world examples where light and positioning is far from perfect. As the data become less meaningful, the bias in favor of white skin increases. The study would be improved by doing an additional in-depth study with real world text that has been selected by humans, and then modify the input by randomizing the target demographic. If the bias remains then we would have a higher confidence in the data. This is similar to studies done in face recognition where issues with dark skins has been demonstrated multiple times.
• taneq 4y ago
  That's another story, for sure, but the significant bias in OpenAI's treatment of the same adjectives when applied to different groups is certainly also a story. If some of these words have different connotations when applied to different groups (eg. "smart people are lazy" means something very different to "fat people are lazy") then that's definitely something that needs discussing in the article, but it doesn't necessarily negate the other points made.
  • Imnimo 4y ago
    >the significant bias in OpenAI's treatment of the same adjectives when applied to different groups is certainly also a story. What is the story? If the bias shows up on sentences that no one will ever say, why is that interesting? What does it tell us?
• muzani 4y ago
  It's missing a bayesian comparison, i.e. how many humans think "fat people are inactive" is a hateful sentence? Even humans vary a lot across cultures.
  • hiccuvtv 4y ago
    It's not missing anything, that's by design. Only following the cults narrative is what matters.