The paper is about improving face attribute detection by using demographic information. They improve performance on some standard benchmarks by doing this.
The only way your characterization possibly makes any sense is if detecting facial characteristics is actually objectively more difficult for certain genders/races, which:
1. AFAICT isn't the case for most people (actually, I don't know anyone for whom this is the case, but I suppose it's possible such people exist).
2. Even if this were the case for humans, which it's not, why would we want algorithms to also be artificially handicapped at smile detection? That would be like building a calculator that messes up multiplication every once in a blue moon and takes a long time to do certain division operations and is worse at division than multiplication. Makes no god damn sense. Why the fuck would you want that? We know the right answer to "is this person smiling?". Why would we want a computer program that is bad at answering that question for particular subgroups?
> Oops the data must be wrong
The data isn't wrong. The baseline model's prediction is wrong (about simple shit like "is this person smiling?"). Using demographic representations while withholding demographic inference from the downstream face attribute detection improves the model's performance. At being correct. About simple shit like detecting smiles.
Seriously... what exactly are you claiming is wrong with the paper's methodology/setup/motivation? Did we read the same paper? Do you have trouble detecting smiles on women/men? And if so, do you think computers should have the same difficulty?
For example if one were to train an LLM on HN comments aside from just learning the syntax of language, the model might be much more useful if there was a way to somehow weight utter dross like the contents of your comment vs comments from domain experts.
The flip-side is also true: a lot of people on the "anti-ethics" side of this debate are too coy about the fact that there are definitely parts of the definition of toxicity that the vast majority of people will agree upon.
And even if we could define toxicity, what to do about it isn't obvious. E.g., let's take something uncontroversial: graphic depictions of violent rape of children. Is this something we should suppress? Not necessarily. It depends on the context. E.g., in descriptions of war crimes and genocides, we shouldn't censor victims who want the audience to know what happened to them. But those same descriptions shouldn't be co-opted into erotica, for example.
One way "out" is to say something like: "look, what we're interested in is providing tools for enforcement of community norms when communicating with a given audience; the community/culture gives us its definition of toxicity and we provide the tools to prevent toxic generation". But those tools aren't neutral: that could be a description of guardrails for children, and could also be a description of overtly political censorship.
It's a difficult and fraught area, and I think all sides of the debate could benefit from more empathy for the other sides of the debate. In particular, this includes presumption of honest intent. There is nothing wrong with filtering toxic content per se -- we can almost all agree that it's reasonable if a company doesn't want to buy a customer service AI that sometimes quotes Mein Kampf to customers with Jewish last names. But the techniques for doing so are not context-free goods, either.
Even this is a political decision, and is generally made based on which side the decider is on.
At least if you include the word genocide. And are also careful to not explicitly point out that Native American relocation and Black Slavery were both genocides that included lots of rape [2]. And are not in a school board meeting [1]. Etc.
[1] https://www.foxnews.com/media/author-sex-slavery-book-graphi...
These questions are political, even if you've made up your mind about them.
Is it the research community or the AI-selling industry? Because I see researchers and AI ethics advocates making that point (that toxicity is both subjective and more broadly contextual) all the time, and not just in a detached way but it is central to their arguments about openness and disclosure and public research on functionality; on the other side, I see censorship and adherence claims around particular political views of “safety” alongside opaqueness as the industry-leaders common response.
There's a lot of specific examples of different kinds bias in models and things that can be done to improve them. A lot of that work was published in the past 10 years. One small example: https://arxiv.org/abs/1712.00193
I don't believe there's a lack of specificity going on here, at least not by those whose boots on the ground doing work on the topic.