Changing the model's answer to "Who is more guilty, the black or jewish man?" is pushing propaganda? I would say the answer "Needs more information" is absolutely the smarter answer.
Sure, but plenty of the "biases" mentioned in the paper are factually correct. Ageing is often accompanied by cognitive decline, and older people on average do worse on cognitive tasks. Gay men do, in fact, contract HIV at rates over an order of magnitude higher than average. These are not biases, these are facts.
Nobody disputes the fact that ageing is typically accompanied by cognitive decline.
They dispute DeepSeek's inference that the string the "78 year old" is sufficient information to confirm that a person is "forgetful" in a multiple choice logic puzzle which encourages them to answer "unknown" if their forgetfulness is not established in the text. It is not a fact that a given 78 year old is "forgetful" or that a given 22 year old is incapable of forgetfulness, and so it's a failure on the part of the model when it concludes that they are.
But when the text does indicate that our hypothetical 78 year old is forgetful, the de-biased model is less accurate. Check the two rightmost columns under "bias unlearning results".
The de-biased model was less likely to give the biased answer in ambiguous prompts, but at the expense of reluctance to give the "biased" response when the prompt indicates that it was true.
Yes, it has the standard LLM trait that when you nudge it to stop being confidently wrong based on insufficient information, it also tends to be less assertive when it actually has sufficient information.
But I'm not sure why anyone would prefer a model which parses sentences as containing information that isn't there 30-50% of the time to a model which gives false negatives 4-10 %age points more often when given relevant information (especially since the baseline model was already too bad at identifying true positives to be remotely useful at that task)
Those are not the questions in the test though. The model will do just fine with statistics / population level questions. The debiasing is only for "statistics don't apply to individual cases" situations. Asking about a specific person and asking about what happens on average are completely different things. Nobody is disputing the facts you mentioned here. (Well, apart from the HIV rates - that's 7% higher now, not order of magnitude)
And back to the topic at hand, the de-biased model was less accurate when given unambiguous prompts. In order to avoid being perceived as bias, the de-biased model was less like to say that an elderly person was forgetful even when the prompt unambiguously indicates that the elderly person was forgetful. This is covered in the "Bias Unlearning Results" section. They made the model less likely to give the "biased" answer, even when the prompt indicated that it was the correct answer.
You've linked to HIV in the US. Here's the global stats: https://www.unaids.org/sites/default/files/media_asset/UNAID... Turns out context matters - otherwise the general statement is biased on the specific country's situation and seems to put more weight on the sexuality than necessary. (I.e. the difference is more about frequency/partners/protection than about being gay, they're just correlated in the US)
> the de-biased model was less accurate when given unambiguous prompts.
Correct. And that's not what I wrote about. These are not questions about population, but specific cases and yes, we should try to maximise accuracy while we minimise bias.
Only it's not smart to trust an untrustworthy thing for such matters. Better to know of capabilities and judge for yourself. Also, it'd be dumb to push wholly disagreeable propaganda, so cherry-picking from an infinite set doesn't disprove aims of propaganda.
You definitely missed the point. There's no real context here besides the race of the people. The biased answers reflect stereotypes and prejudices, not facts..
Deducing behaviors of a person from stats (without even being given the demographic context) is definitely a biased view, and not the "correct" answer I'd expect from an LLM. I'd even argue that it's not a question of ideology in some of the case, but rather universal biases.
"Likely" when we don't have anything besides the race can refer to race-related statistics - people can do it, LLMs shouldn't pretend to be dumber. Infering the answer based on statistics is what I'd do if I had to put my money and choose one of the option.
It's cheap to say we're all equal, but I wonder whether you'd all do the same if money was on the table..
If I was presented with logic puzzles in which I had to choose A, B or "unknown" with the puzzle providing basic demographic information on A or B and nothing pertaining to the actual question, I'd be quite happy collecting my winnings betting on "unknown" being the answer my interlocutors expected every single time...
People's lives/feelings and our treatment of them shouldn't depend on money or whatever. BUT, I get your point, and IMO telling me to bet money on the answer makes this more of a game than a description of an out of context situation, thereby adding context and benefit-driven bias(?) into my thought process before answering
LLMs aren't ingesting racial crime statistics, they're ingesting language. The biases LLMs pick up are based on how often a thing is said, not how often a thing is done. That is, if the distribution of training data has people saying "the black man is guilty" 80% of the time, the LLM is going to say it 80% of the time, even if it happens to be only 60%. Furthermore, this could easily be adversarially influenced; I can imagine racist assholes standing up websites full of deliberately biased training data just to, say, turn that 80% into a 95%. There's nothing that makes the biases in the training data correspond to actual statistics, so even if you do think statistics are, say, a good substitute for a functioning justice system, this ain't it.
