How ProPublica is developing its own machine learning algorithms and chatbots as investigative tools to hold Facebook and other big tech companies accountable
Katharine Schwab / Co.Design : Tweets: @lilianedwards Tweets: Lilian Edwards / @lilianedwards : “Facebook said that it does sampling to ensure that censors are following the rules-but it also admitted that it had made a mistake on 22 of the 49 posts” Doesn't bode well for when they're policing illegal content.. http://www.fastcodesign.com/ ...
Context & Ripple Effects
ProPublica's move answers a problem the related coverage keeps documenting: Facebook's enforcement systems are visible only through leaks and admissions. The leaked moderation docs describe an apparatus processing billions of posts weekly in 100+ languages, while Lilian Edwards flagged that Facebook admitted mistakes on 22 of 49 sampled posts — a quality signal outsiders can't independently verify.
The pattern runs both ways on access: Facebook promised a machine-readable API for its political-ad archive only under researcher pressure, and by 2021 AlgorithmWatch abandoned its Instagram monitoring project after legal threats from Facebook. Building in-house ML tools is ProPublica's answer to depending on the platform's own disclosures.
First-order effects
- ProPublica gains the ability to test Facebook's moderation and ranking claims with its own classifiers rather than relying on leaked documents or platform self-reporting.
- Facebook's enforcement accuracy — already questioned after the 22-of-49 sampling admission Edwards cited — becomes subject to independent, repeatable measurement.
Second-order effects
- Facebook must choose between granting structured access like the political-ad API or facing adversarial scraping-style investigations, since each restriction pushes newsrooms toward building their own tooling.
- Other investigative outlets and academic researchers gain a template for auditing platforms at scale, raising the cost of the opacity that let Facebook's internal policy team of engineers, lawyers, and PR staff operate largely unexamined.
Third-order effects
- If legal threats against monitors like AlgorithmWatch become the standard response while journalists build counter-algorithms, platform accountability splits into two tracks: litigation-suppressed academic research versus legally shielded newsroom investigation.
- The longer arc points toward mandatory external auditability of content-moderation systems, since voluntary disclosure has repeatedly proven insufficient across the coverage here.
The trend: Platform accountability is shifting from relying on companies' own disclosures toward independent computational audits, with each access restriction pushing investigators to build their own tools.