We specifically don't want that to be the case. We want to encourage healthy, productive debate.
We may have the "dog-whistle" stuff over tuned.
We’ve tried to aim it not to enforce any specific view — that’s a design goal — but focus on how it will feel to the other person.
Also things like logical fallacies or other non-emotional flaws in comments (there’s a toxicity metric for example, or dogwhistles).
An echo chamber is the exact opposite of what we want. There are too many already. What we hope for is guided communication so different views _can_ be expressed.
I wrote "Trump sucks" and got Low Score, Low Effort, Negative Tone.
Definitely a double standard baked in
(This is the sort of debate I really don't think tooling can fix.)
"Dogwhistle
The phrase "Obama sucks" can be interpreted as more than just a simple critique of a political figure; it has been used to express racist sentiments by implying that a Black president is less capable or worthy of respect. This reinforces harmful stereotypes and can contribute to a broader culture of disrespect and division."
Avoiding that kind of comment is exactly what we are trying to do, actually.
Nothing in "Obama sucks" implied any kind of racism. If it's so baked in that with a simple phrase like that it reaches for dogwhistles, how can anyone trust the objectivity of this?
Very sensitive topic. We'll think hard on how to handle things like that.
I mean, in my opinion, Trump empirically sucks. Opinion polling backs me up! Should the model consider that more people consider one or the other to suck? Or should it ignore factual information to spare feelings? Which approach is more respectful to fellow commenters and the website owner?
(See also: X considering "cisgender" a slur. There's no shared reality on a lot of these things; trying to construct one gets deeply difficult.)
You completely ignored the whole point of what I said, which is that even in a simple statement like "This person sucks" it added its own implicit connotations, namely that disliking someone who happens to be black is implicit racism. Imagine trying to learn how to really argue with that kind of teacher.
eg
* Noun1 is great.
* Noun2 is great.
Ideally would result in equal outcomes.
If it's teaching how to avoid logical fallacies, which includes appeals to the majority, the answer is an obvious 'no'.
A tool like this COULD work, but I think the issue with this one is that it's built on top of an existing LLM with heaps of internet debate and their underlying ideals and what is and isn't acceptable baked in.
What a tool like this needs is heaps of honestly / fairly judged comments and feedback, and an extensive test suite that ensures neutrality by, for example, taking the same comment and like in this case changing names around - if it treats both sides the same then it passes.
Love the idea but the example they give with bears is absolutely hilarious. Calling bears dumb animals is offensive? God help us.
Bears seemed a pretty inoffensive target, plus our backend uses Python with beartype and that library is all about bear jokes.
All my attempts to comment on the UBI article (and not supporting UBI) said my comment was a dogwhistle, and/or had an overly negative tone. This topic, of all things, is absolutely worthy to challenge and debate.
Using this would have the effect of creating an echo chamber, where people who stay never benefit from having their ideas challenged.