Update your Cookie Settings to use this feature.
Click 'Allow All' or just activate the 'Targeting Cookies'
By continuing you accept Avaaz's Privacy Policy which explains how your data can be used and how it is secured.
Got it
We use cookies to analyse how visitors use this website and to help us provide you the best possible experience. View our Cookie Policy .
Avaaz Report

BRIEF: How Facebook’s AI is failing American voters ahead of Election Day

Facebook is allowing ‘repeat misinformers’ to flood the platform with clones of misinformation, racking up 142M views and counting

October 9, 2020

Back to Disinfo Hub
Report Image

Variations of misinformation already marked false by Facebook - such as content alleging Presidential Candidate Joe Biden is a paedophile, that President Trump’s family stole money from a kids charity, or that multiple stamps are needed for mail-in ballots - are slipping through Facebook's detection system and being viewed millions of times ahead of the elections, according to Avaaz.  

While Facebook says it will apply ‘strong warning labels’ to fact-checked content and reduce its distribution1, Avaaz’s new investigation found that nearly half of the fact-checked misinformation content (42%) we analysed is managing to circumvent Facebook's own policies and remain on the platform without a label, earning millions of interactions.

Flaws in Facebook's fact-checking system mean that even simple tweaks to misinformation content are enough to help it escape being labeled by the platform. For example, if the original meme is marked as misinformation, it's often sufficient to choose a different background image, slightly change the post, crop it or write it out, to circumvent Facebook's detection systems.

As a result, many Facebook pages seem to be slipping under Facebook’s radar for being a “repeat offender” and avoid being down ranked in users' News Feeds. This leaves them free to go viral ahead of the US Elections in November.

We identified 119 ‘repeat misinformers’ -- pages that we found have shared misinformation content at least three times, which was fact-checked by one of Facebook’s fact-checking partners in the US over the last year. Together, these pages generated an estimated 5.2 billion views in a year2.

All this is happening despite Facebook’s recent claim that its AI is already “able to recognise near-duplicate matches” and “apply warning labels”, regarding Covid-19 misinformation and exploitative content, noting that “for each piece of misinformation [a] fact-checker identifies, there may be thousands or millions of copies.”3 Our findings suggest that Facebook’s detection systems are failing to keep up and need to be made much more sophisticated to prevent disinformation going viral.

We were able to identify 738 posts without fact-checking labels that contained variations of misinformation content that was debunked by Facebook’s fact-checking partners in the US. These 738 posts - which presumably should have a fact-checking label and be down ranked in the algorithm - have racked up 5.6m interactions and an estimated 142m views.

Facebook must act urgently to label and down rank both this misinformation and the actors that systematically spread it ahead of the US 2020 elections. That’s why Avaaz is calling on Facebook to immediately label all variations of the same misinformation, downrank pages and groups that systematically spread it, and correct the record to all users who’ve seen misinformation4.

Update as of October 7, 2020: Avaaz provided Facebook with our evidence on Thursday, Oct 1st . About six days later (Wednesday, Oct 7th) only 4% of the 738 false posts flagged to Facebook without a label now do have a label, and 3% appear to have been removed. The other 93% are still on the platform without a label. The lack of urgency and narrow scope of action by Facebook to act on the evidence provided is especially troubling, imagining such false claims going viral just weeks ahead of the elections and Facebook’s detection system doesn’t catch them5.

Email media@avaaz.org for the entire report.


  1. https://www.facebook.com/journalismproject/programs/third-party-fact-checking
  2. Our methodology to calculate ‘estimated views’ can be found in the methodology section below.
  3. https://ai.facebook.com/blog/using-ai-to-detect-covid-19-misinformation-and-exploitative-content/
  4. https://secure.avaaz.org/campaign/en/correct_the_record_study/
  5. The screenshots and measures represent the state of the posts as we found them during our research, and before sending our findings to Facebook.