Facebook's Q3 transparency report: 0.1%-0.11% of what users see violates hate speech rules; 95% of hate speech takedowns are proactive, up from 24% in 2017
The VergeAdi Robertson
Context & Ripple Effects
Facebook's Q3 report is the latest step in a metrics pivot it has been telegraphing all year: Q1's action on 9.6M pieces of hate speech and Q2's doubled 22.5M removals were volume stories, while this report leads with prevalence — just 0.1%-0.11% of what users see breaks the rules — and with 95% of takedowns initiated by Facebook itself rather than user reports, up from 24% in 2017.
The proactive-detection climb has been steady — 68% to 80% software-identified through 2019, 88.8% in Q1 — so the Q3 number reads as the maturation of an AI-first enforcement stack. The stakes of the prevalence framing show up a year later, when Facebook disputes a WSJ report by citing a 0.05% prevalence figure, and when the Transparency Center institutionalizes the 'removed before reported' stat.
First-order effects
Facebook's own enforcement now barely depends on users: at 95% proactive detection, human reports are a rounding error, so the platform's AI classifiers — not its community — set the effective enforcement boundary.
The headline metric changes from how much Facebook removes to how little users see, which makes quarter-over-quarter removal counts like Q2's 22.5M a secondary stat in Facebook's own reporting.
Second-order effects
Critics and journalists lose the easy takedown-volume attack surface and must instead contest the self-reported prevalence number itself — the dynamic that surfaces in Facebook's later dispute with the WSJ over its 0.05% figure.
Rivals face pressure to publish comparable prevalence-and-proactivity stats, since a platform reporting 0.1% prevalence with 95% proactive detection sets the disclosure bar other social networks get measured against.
Third-order effects
If the pattern holds, platform accountability shifts toward company-defined prevalence metrics audited only by the companies that compute them — Facebook's Transparency Center and its disputed WSJ rebuttal both point to self-reported numbers becoming the industry's standard of proof, with external verification as the unresolved question.
The trend: Content-moderation reporting is moving from takedown volumes to AI-driven proactive detection and self-reported prevalence, turning Facebook's own measurements into the central battleground over platform accountability.
Facebook's Q3 content moderation report showed major improvements over last quarter, as some moderators returned to the office. These new stats only help build moderators' case that they're core to Facebook and ought to be more than contractors https://www.protocol.com/...
The Indian government sent the highest ever user data requests to Facebook, second only to the US, in the first six months of 2020, the social media giant said in its biannual transparency report released on November 19. 1/7
@Reuters Update: Facebook released their biannual transparency report on Friday. It showed they censored 834 posts in Vietnam upon gov't request in the first half of 2020. 14 profiles were removed “in error”, the report says. https://twitter.com/...
Lack of transparency reduces the value of FB transparency reports. Statements like “automation removed 95% of the hate speech that was removed” elide FB's willingness to permit significant amounts of hate speech. https://twitter.com/...
While other social media companies have implemented aggressive new polices to combat the spread of disinformation on their platforms, Facebook has failed to do the same. Today @RepKatiePorter and I wrote to @Facebook CEO Mark Zuckerberg demanding a more rigorous response. https:/…
Last year @PeterDutton_MP said right-wing extremists were “on the dark web.” In reality, they are also hiding in plain sight, on platforms like Facebook, spreading hate speech & propaganda. It's a growing threat the Morrison Govt must take seriously. https://www.reuters.com/...
Facebook says Artificial Intelligence detects 94.7% of hate speech removed from platform. Facebook needs to shut it's operation today and open at a later time. Militia groups and Hate groups congregate on Facebook to set meetings! https://www.cnbc.com/...
These numbers on now many pieces of misinformation Facebook labelled seem impressive, but they miss is the gargantuan effort from fact-checkers working in overdrive all year. It's not just the company's achievement. https://www.engadget.com/...
What's fascinating about this report is that their data points are p meaningless. Facebook's definition of hate speech is extremely narrow and their enforcement of policies are inconsistent. The issue isn't just speed it's that they don't consider a lot of hate to be hate speech.…
In 2020Q3, hate speech prevalence was 0.10-0.11%. Our work aims to minimize the prevalence of hate speech, while also reducing mistakes we make when removing content. More details on our work on hate speech here: https://about.fb.com/... (6/7)
CTO Mike Schroepfer commends the development of FB's AI in catching 95% of hate speech before it's reported. In '17, he said it was around 25%. Remember that the co's moderators wrote a letter yesterday saying FB's AI isn't as good as ppl think it is https://twitter.com/...
Credit where it's due: people have been calling for Facebook to release prevalence stats (the ‘but how much did you MISS?’ question) for a long time. Yes, there's still many more qs (auditing; breakdown by country/affected communities/seriousness) but this is a big step forward h…