4 ms·

Who Is Making Sure the A.I. Machines Aren’t Racist?

6 points by reckel 6y ago

mlthoughts2018 6y ago
Ugh, here we go with more of the sanctimonious nonsense about AI fairness research from the Times. Google was well within rights in firing both employees (Gebru sincerely quit and was not fired, given her ultimatum, but Google would have been in the right). I’m no Google apologist - their ads monopoly should be broken up, they are anti-union, pay harassers huge sums of money and lead the way in poorness of leetcode software culture and low accountability. Google has poor customer service and largely has no idea how to build products for people, constantly shutting things down, killing acquisitions, and using poorly attended message forums as a bureaucratic wall to steamroll any real way of seeking customer support or appealing black box takedown decisions. Google is a mess. That doesn’t make Gebru in the right - she’s not at all. Here are several previous comment threads going on at length about why Gebru is sincerely in the wrong here: - https://news.ycombinator.com/item?id=25314824 https://news.ycombinator.com/item?id=25314824 - https://news.ycombinator.com/item?id=26147178 https://news.ycombinator.com/item?id=26147178 - https://news.ycombinator.com/item?id=26238365 https://news.ycombinator.com/item?id=26238365
steve_g 6y ago
Regardless of whether you think Timnit Gebru is a jerk or a very fine person, bias in real-world AI is a big problem. To the extent algorithms are used to make significant, real-world decisions (whom to arrest, whom to extend credit, whom to grant parole, etc.) they had better be fair. And it seems that black-box algorithmic decisions are becoming more and more prevalent in the real world. It's been clearly demonstrated that biased training sets give biased results. It's also hard to come up with un-biased training sets. This is a problem that's worth working on.
- mlthoughts2018 6y ago
  The problem is that what you describe is a question of pragmatics and specific data curation, which is not sexy or marketable for political fairness and bias researchers. They need to manifest false angles of the problem, connections to climate change or #metoo or whatever is the social justice outrage du jour, so that it can win news cycle attention, TED talks and political notoriety. It leads to perverse incentives where AI fairness stops being an engineering problem that should be exclusively focused on objective criteria that need to be met to solve concrete problems, and starts being a tribalistic song and dance about wokeness, cancel culture and shoe-horning subjective values into learning systems at a strategic level. Unfortunately AI fairness is not a serious research domain yet. It’s a venue for trying to create political cottage industries around convenient social justice drama and then trying to parlay that into a career by building political moats such that if anyone questions the legitimacy of this or that aspect of “fairness research” then that skepticism alone can be used to cancel them and argue they are regressive.

6 points by reckel 6y ago

mlthoughts2018 6y ago

Ugh, here we go with more of the sanctimonious nonsense about AI fairness research from the Times. Google was well within rights in firing both employees (Gebru sincerely quit and was not fired, given her ultimatum, but Google would have been in the right). I’m no Google apologist - their ads monopoly should be broken up, they are anti-union, pay harassers huge sums of money and lead the way in poorness of leetcode software culture and low accountability. Google has poor customer service and largely has no idea how to build products for people, constantly shutting things down, killing acquisitions, and using poorly attended message forums as a bureaucratic wall to steamroll any real way of seeking customer support or appealing black box takedown decisions. Google is a mess. That doesn’t make Gebru in the right - she’s not at all. Here are several previous comment threads going on at length about why Gebru is sincerely in the wrong here: - https://news.ycombinator.com/item?id=25314824 https://news.ycombinator.com/item?id=25314824 - https://news.ycombinator.com/item?id=26147178 https://news.ycombinator.com/item?id=26147178 - https://news.ycombinator.com/item?id=26238365 https://news.ycombinator.com/item?id=26238365

steve_g 6y ago

Regardless of whether you think Timnit Gebru is a jerk or a very fine person, bias in real-world AI is a big problem. To the extent algorithms are used to make significant, real-world decisions (whom to arrest, whom to extend credit, whom to grant parole, etc.) they had better be fair. And it seems that black-box algorithmic decisions are becoming more and more prevalent in the real world. It's been clearly demonstrated that biased training sets give biased results. It's also hard to come up with un-biased training sets. This is a problem that's worth working on.

mlthoughts2018 6y ago
The problem is that what you describe is a question of pragmatics and specific data curation, which is not sexy or marketable for political fairness and bias researchers. They need to manifest false angles of the problem, connections to climate change or #metoo or whatever is the social justice outrage du jour, so that it can win news cycle attention, TED talks and political notoriety. It leads to perverse incentives where AI fairness stops being an engineering problem that should be exclusively focused on objective criteria that need to be met to solve concrete problems, and starts being a tribalistic song and dance about wokeness, cancel culture and shoe-horning subjective values into learning systems at a strategic level. Unfortunately AI fairness is not a serious research domain yet. It’s a venue for trying to create political cottage industries around convenient social justice drama and then trying to parlay that into a career by building political moats such that if anyone questions the legitimacy of this or that aspect of “fairness research” then that skepticism alone can be used to cancel them and argue they are regressive.