Report: misinfo spreaders have evaded Facebook's content review systems by reposting against different-colored backgrounds, changing fonts, or cropping images
@b_fung's latest. https://www.cnn.com/... Brian Fung / @b_fung : NEW: Facebook is seemingly unable to catch copies of misinformation it's already flagged as false, according to @Avaaz. The group found hundreds of posts that were simply recropped memes or reposts on different backgrounds to evade enforcement. https://www.cnn.com/... Andy Legon / @andrewlegon : Facebook has had 4 years to fix this - 2016 might have been an accident - This is just negligence! If the @Avaaz team can catch this, so can one of the most powerful companies ever... #disinformation #USElections2020 https://cnn.com/... Roger McNamee / @moonalice : Attention FB: When your best defense is that your size prevents you from fulfilling your responsibilities to society, you have lost the argument. #StopHateForProfit @FBoversight https://twitter.com/...
Context & Ripple Effects
This is the third time Avaaz has stress-tested Facebook's enforcement and found it leaking. An earlier Avaaz study flagged cracks and loopholes in the labeling system even while overstating the overall volume of false stories, and a later user survey found flagged posts on controversial topics like the 2020 election went labeled only a handful of times out of hundreds of sightings.
The new finding narrows the failure to something specific: Facebook's systems match known-false content but not near-copies, so a recropped meme or a repost on a different background resets the post to unreviewed. That gap was predictable — a 2018 analysis already showed Facebook's and Google's algorithms struggle to detect doctored images and videos.
First-order effects
- Spreaders of content Facebook has already labeled false regain distribution right before the US election, because each cosmetic edit — new background, changed font, crop — produces a fresh post that bypasses the existing flag.
- Facebook's fact-checking investment is partially nullified at the point of enforcement: the label exists, but the copy circulating in feeds carries no label.
Second-order effects
- Facebook is pushed toward perceptual or robust-content matching rather than exact-match detection, a costlier engineering problem its own 2018-era tooling was shown to be weak at.
- Rivals face the same evasion playbook — if trivial edits defeat Facebook's matching, the identical technique transfers to other platforms' review systems, forcing industry-wide fixes rather than a Facebook-only patch.
Third-order effects
- If trivially transformed reposts reliably defeat automated matching, enforcement shifts toward detecting the people and networks behind repeat spreading — consistent with Avaaz's later finding that a small cluster of mostly right-wing accounts drove a disproportionate share of voter-fraud posts — and strengthens the case for independent audits of platform enforcement.
- The pattern points to moderation becoming an arms race where external researchers like Avaaz function as de facto inspectors, raising pressure for regulators to mandate measurable enforcement standards rather than trust platform self-reporting.
The trend: Content moderation is settling into an arms race in which cheap cosmetic edits defeat automated matching, pushing platforms toward network-level enforcement and outside auditors as the real check on labeling.