- More than 24 hours ago, Mark Zuckerberg released a statement on the quarterly results on hot topics like business revenue and hate speech. He boldly claims that 90% of hate speech cases have been able to be automatically detected.
- The Simpson’s paradox in probability and statistics and states that when grouping, combining, aggregating or averaging sets of information, some information may be lost. In other words, the devil is in the details.
Early in any software engineers career, you learn about the infamous edge cases: extreme or rare inputs and behaviors your algorithm is never free of.
For example, if you have a simple function that averages a set of values as an input:
function average(values):
sum = 0
for x in values:
sum += x
return sum / values.size
For non-coding users, what this code does is that for a set of values i.e. {1, 2, 3} it sums each element to a cumulative variable, and finally divides the sum by the number of elements in the set. In summary, the average we know.
However, this code might not be used by everyone in the same way:
- Some might try to give non-numerical values in the sets
- Some might mistakenly pass null i.e. non-existent values to the set
- The input could be another structure that is not a set i.e. the dictionary
- The numbers in the set might be so large that the sum can’t be contained in a computer’s memory
You can notice that these behaviors are not part of the majority, or expected behavior. This is what we call edge cases.
Now back to Zuckerberg’s 90% statistic. This number can be interpreted as follows: if 100 users make a comment, 90 of the comments made by the user assuming they only make one would be detected as hate speech. There are two problems when applying this to our reality.
What is Facebook’s definition of hate speech? In a small research I carried out in my personal account, from 14 hate speech comments against BIPOC and the LGBTTQ+ community only 4 were flagged as hate speech, which had a simple formula <subject> is <adjective>. Unfortunately, I wasn’t able to perform a bigger and more statistically significant test. However, it could be questioned if Facebook’s definition of hate speech includes comments against LGBTTQ+ and BIPOC. Now let’s make a huge assumption that this 4 out of 14 comment detection is true at a greater scale. Where do the 90% of comments detected come from then?
Who comments on Facebook? The better question to answer is who has access to Facebook. The answer is anyone who can afford internet access. Unfortunately, this doesn’t favor low-income households and BIPOC. In America, “the high average cost of internet service is unaffordable for many U.S. consumers, contributing to a longstanding digital divide that disproportionately affects low-income households and BIPOC communities.” (New America, 2020) I can neither deny nor confirm the following, but we must ask ourselves if the automated hate speech prevention tool by Facebook favors those with the most priviledge. After all, Facebok’s CEO has stated that “men are trash” are hate speech, and this is something that has always made most of us in the LGBTTQ+ and BIPOC community to roll our eyes.
Now back to the Simpson’s paradox. By summarizing the effectiveness of their hate speech prevention tools and guidelines behind this 90% statistic, Facebook’s CEO losses information. It loses information on the possibility that some of the sets of comments that weren’t unable to be detected were hurtful to BIPOC and LGBTTQ+ communities. Like the extreme values in our average function, we’re the edge cases. And although addressing edge cases is considered a gold standard in software engineering and Facebook’s coding practices, Mark Zuckerberg doesn’t address these caveats in his statement, nor does he point out to tangible efforts other than third party consulting on how they’re trying to overall “make conversations healthier”. BIPOC and LGBTTQ+ are an unaddressed edge case.
Even worse, there might be a possibility that from the 100 cases of the 90 out of 100 ratio, there might exist more that haven’t even been identified by any proxy, person or expert as a hate speech comment. COVID has taught us that authorities might choose to consider obscuring sampling, and that bias has to be addressed in measuring the effectiveness of our solutions to critical situations.
In summary:
- Facebook continues to obscure and ignore the power and priviledge disparity suffered historically by minorities.
- Facebook needs to foster sensitivity metrics like precision and recall.
- Facebook needs to acknowledge there might be some cases it might be missing under their radear.
And that hasn’t changed with the statement he just released after the Anti Trust Hearings the major tech companies, like the one Facebook is part of, held some days ago. The statement generates the same questions without any authentic answers.
References: