Analysis finds algorithms designed by Facebook, Google, and others often struggle to detect doctored images and videos, a key tool for misinformation efforts
Wall Street Journal : Tweets: @dseetharaman , @saraheneedleman , @dseetharaman , @scottmaustin , and @dseetharaman Tweets: Deepa Seetharaman / @dseetharaman : NEW: Doctored photos & videos are a big, unresolved problem for tech companies who need to stamp out misinformation. Russian propagandists knew that. https://www.wsj.com/... Sarah E. Needleman / @saraheneedleman : “People will watch cat videos endlessly, but they won't take a minute to ascertain whether what they are being told is true or not.” https://www.wsj.com/... Deepa Seetharaman / @dseetharaman : For example, Black Matters US took an authentic 2013 image of a Nigerian boy and created a meme with a false stat on the life expectancy of black American men. Photographer wasn't aware until @WSJ called. https://www.wsj.com/... pic.twitter.com/Yif8sfT52W Scott Austin / @scottmaustin : Russian propagandists used a simple technique to get around Facebook's filters: Post doctored photos. Here's a photo taken at a pro-immigration rally, and an altered version that appeared on a Russian-linked Facebook page. https://www.wsj.com/... pic.twitter.com/qjjEKr6NVO Deepa Seetharaman / @dseetharaman : Facebook told us that they're going to include images and photos in fact-checkers queues in coming weeks & discussed the issue with them earlier this month. https://www.wsj.com/... pic.twitter.com/RkQXbWATrG
Context & Ripple Effects
This analysis lands three days after a [[a:926832|Mueller indictment showed how easily Facebook's Groups tools could be exploited by Russian actors]] — same adversary, different vector. Where that story was about distribution infrastructure, this one is about the payload itself: the doctored photos and videos flowing through it, which the detection systems at Facebook and Google were not built to catch.
It also marks a shift in what 'detection' means for these platforms. Facebook had already claimed progress spotting coordinated amplification through patterns of activity across fake accounts — but pattern-matching on accounts says nothing about whether an individual image is real. The later record bears out the gap: spreaders learned to evade content review by reposting images against different-colored backgrounds or cropping them, and by 2023 a study of 13M+ posts found nearly a quarter of sampled political images ahead of the 2020 elections contained misinformation.
First-order effects
- Facebook and Google face immediate pressure to move detection beyond account-level patterns into per-image and per-video authenticity checks — a capability the reporting shows they currently lack.
- Russian propagandists and similar operators retain a working playbook: doctor media, distribute it through organic-seeming channels like Groups, and stay ahead of classifiers tuned to older examples.
Second-order effects
- Adversaries industrialize evasion rather than volume alone — the documented tricks of recoloring backgrounds, swapping fonts, and cropping show the arms race moves to whatever perturbation defeats the classifier, forcing continuous retraining costs onto platforms.
- Verification burden shifts downstream to users and third-party fact-checkers, since the platforms' own tooling cannot arbitrate whether a given image is genuine.
Third-order effects
- If the pattern holds, platform integrity spending migrates from policing who posts toward authenticating what is posted — provenance signals attached at capture time becoming the durable answer classifiers can't provide.
- Doctored-media detection becomes a standing regulatory expectation for large platforms, as election-cycle failures like the midterm-era misinformation surge keep the issue on the public agenda.
The trend: Content moderation is shifting from identifying fake accounts and coordinated behavior to authenticating the media itself, with provenance standards emerging as the long-term complement to detection algorithms.