Thread by @gchaslot

YouTube's deep learning AI boosted massively two news channels in the US last month:

➡️Newsmax TV went from 3M to 110M views/month (x33!)
➡️New Tang Dynasty TV (NTD): 6M to 80M views/m (x13)

They both made wild claims of voter fraud. Why does the AI like them?

Thread ⬇️

I already made a thread about Newsmax TV last week (see below), so let's focus on NTD

2/

https://twitter.com/gchaslot/status/1334615032617897990

New Tang Dynasty TV was funded by the Falun Gong religious movement, that got banned in China, and thousands of followers have died under Chinese custody.

This channel existed for years, but its views just exploded last month. Why?

3/
en.wikipedia.org/wiki/Falun_Gong

Today, TDN was one of the most recommended news channel from my dataset from 800+ channels.

Only Fox News and NBC scored higher among news channels
algotransparency.org/index.html?dat…

4/

Here are the type of videos that got massively recommended today:


5/

Here are more examples of videos of NTD that got millions of views thanks to massive recommendations:

6/

So why does the algorithm like to recommend unfunded claims of voter fraud? Let's take one example:

1⃣ "1.2M fraudulent votes?" is a catchy title. So people click. People wait for evidence, and often have to watch the full video.

7/

2⃣ Even if they're not convinced by "the evidence" they often watch a lot of the video. So the algorithm takes it as a signal that it *should* recommend the video to more people.

8/

It's the exact same thing that happened with flat earth from 2006 to 2019, until I confronted Google about it.

Flat earth is now in harmful content and not recommended, unlike baseless voter fraud accusations.

9/

https://twitter.com/gchaslot/status/1064527592428986368

Google needs to invest seriously AI that take into account truth and ethics.

If we don't do anything, we'll live in a post truth world where people are divided into factions that don't talk to each other.

It's not a bug, it's a feature that increases ad revenue.

10/