It's gotta' be a tough time in Facebook PR right now.
While The Social Network is keen to promote reports looking at how much money its fundraisers have raised and how well it's doing in detecting offensive content before people see it, the narrative around the company this week has been dominated by a damning New York Times report, which is highly critical of its actions in dealing with political interference and manipulation.
Because of that broader discussion, stories like this will get less focus, but Facebook has this week also sought to highlight just how good its machine learning systems are getting at detecting and removing offensive content, normally before anyone even sees it.
And they are getting very good - as per Facebook:
- Since our last report, the amount of hate speech we detect proactively, before anyone reports it, has more than doubled from 24% to 52%. The majority of posts that we take down for hate speech are posts that we’ve found before anyone reported them to us. This is incredibly important work and we continue to invest heavily where our work is in the early stages - and to improve our performance in less widely used languages.
- Our proactive detection rate for violence and graphic content increased 25 percentage points - from 72% to 97%.
Those are strong results, and lead to a much better, and safer, user experience. But amid broader discussion around how Facebook itself has failed to protect its users, and has sought to benefit from their data, it's hard to see this being viewed in a positive light.
But the data does show that Facebook's AI detection systems are getting very good at eradicating such material - look at this chart on their improvement in performance in detecting posts depicting graphic violence.
Of course, these are self-reported - its hard to know, definitively, how this is truly measured. But the stats do show that Facebook's doing more to shield users from unsavory, and disturbing material - and if they can do that without having to push the same onto human moderators, many of whom are left mentally scarred by such process, all the better.
Some other key points of note:
- "In Q3 2018, we took action on 15.4 million pieces of violent and graphic content. This included removing content, putting a warning screen over it, disabling the offending account and/or escalating content to law enforcement. This is more than 10 times the amount we took action on in Q4 2017."
- "We also took down more fake accounts in Q2 and Q3 than in previous quarters, 800 million and 754 million respectively. Most of these fake accounts were the result of commercially motivated spam attacks trying to create fake accounts in bulk. Because we are able to remove most of these accounts within minutes of registration, the prevalence of fake accounts on Facebook remained steady at 3% to 4% of monthly active users as reported in our Q3 earnings."
- "In the last quarter alone, we removed 8.7 million pieces of content that violated our child nudity or sexual exploitation of children policies — 99% were identified before anyone reported them."
These are very strong results for Facebook's AI detection tools, and all work to improve the platform - while the lessons learned in their development could also help other platforms to provide similar protection.
Much of this will be lost in the broader discussion and skepticism over Facebook's real motives, but the numbers do show positive signs for the future of content detection and removal. That's not to minimize the criticism of Facebook's actions as detailed in The NY Times story, but there are other aspects where The Social Network is making positive progress.