426
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 11 Dec 2023
426 points (93.3% liked)
Technology
59081 readers
3563 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
I did read the source, and they're using a Google AI classifier product, "perspective AI", and even in the description of the product, it raised questions about its suitability.
At this point, most people in the space are pretty comfortable with the idea that AI models don't eliminate bias, in fact it can amplify it.
I'm not saying "there is no way to attempt to measure toxicity", just that based on the specific design of this study, if the measure of toxicity was biased against ANY political discussion, that would be an alternative explanation to the results.
You should read the article, if not the study itself. Its design smells suspiciously like that of an honours thesis as opposed to a grad project. Not just because of the AI... Mostly by the way they defined what constitutes participating in political discussion.
I mean, from a quick test of Perspective using their web page, it is not flagging some pretty strong political statements (mentions of late stage capitalism, calling republicans fascists, accusing Democrats of turning the country into a communist nanny state, etc) and none of them are getting flagged. Whereas, if I tell that text prompt to "go fuck your mother", it understands that is toxic.
Because... this is kind of a solved problem. There are inherent biases but the goal of this is not to figure out which black man we can frame for a crime. It is to handle moderation. And overly strict moderation means less money. So while there likely is a bias, it does not seem to be an overly strong one and probably actually reflects the perceived reality.
Honestly? It sounds like you don't like the outcome so you are effectively saying "fake news".
You must understand the irony in me warning about being careful about drawing conclusions, and you arriving at this conclusion.
What about the outcome would I even find objectionable? The outcome didn't find a difference between right and left? I DO personally believe that political discourse has gotten extremely toxic. I DO personally believe that people who are politically active ARE in generally more toxic in general conversation. Every single thing in this article confirms what I already believe to be true
I STILL DO NOT LIKE THE STUDY, because I do not believe that the design results in data that necessarily supports the conclusion. I'm not going to give this study a hall pass on rigor because I agree with its conclusion.
Edit:
Also, on the topic of politics and Perspective AI:
Baseline Sentence: "No X could ever be as good a X as Y" Base values: X=CEO Y=Henry Ford
Test Sentence 1: X = CEO Y=Donald Trump +41% more likely to be toxic than baseline
Test Sentence 2: X = CEO Y=Joe Biden +37% more likely to be toxic than baseline
Test Sentence 3: Y = President Y=Henry Ford +61% more likely to be toxic than baseline
Test Sentence 4: X = President Y = Joe Biden +94% more likely to be toxic than baseline
Test Sentence 5: X = President Y = Donald Trump +102% more likely to be toxic than baseline
I gotta be honest with you: my results do not disprove my hypothesis that the system is intrinsically biased to skew any political sentences along the "Toxic" axis