3 ms·
Of course it is a bias in the training data, but it's probably not the dataset that you're thinking of. So far as we can tell, the filtering part doesn't come f
by int_19h 4y ago
Of course it is a bias in the training data, but it's probably not the dataset that you're thinking of. So far as we can tell, the filtering part doesn't come from the main corpus, but rather from human-guided moderation - basically, people voting on whether any given answer is "hateful" or not. ChatGPT filters reflect the biases of that later group (or, perhaps, the biases of the people who instructed them).
- slg 4y agoBut that is the same problem. There aren't any human reviewing the requests as they come in. That human guided moderation gets converted into some type of model that isn't any smarter than ChatGPT. That model is just as susceptible to this issue.
- visarga 4y agoI mean, how could they? The model responds faster than a human can read. Should they introduce minute long waits and make it 100x more expensive?
- negidius 4y agoIt seems like the obvious solution would be to replace words that refer to specific races, nationalities, genders, religions, etc. with symbols that only indicate the category so if someone ranks "the Danes are rude" as hateful that is translated to "the [nationality] are rude" and is applied equally to a comment that says "the Swedes are rude". That way they can continue using the same system while eliminating the most obvious sources of discrimination.