Thread by @sundersays

Twitter has produced a paper on how algorithms amplify of political content
- In 6 of 7 countries (except Germany), algorithm boosts mainstream right party content over mainstream left content.
- Reasons unknown.
- this effect is strongest in the UK
blog.twitter.com/en_us/topics/c…

Right/left amplification by country

By country

Caveat: boost for party content does not carry through for individual Mps from that party

Why? Twitter says Don't Know, more research needed.

Suggests could be a media effect. (But nb party v MP)

Or party tactics/strategy. (Have worked out how to game it. Or differences in images, language that help right).

Nb, Cons & Labour party HQ content does this. But Cons/Lab MP content does not differ markedly.

I did not spot/follow how far algorithmic boosts depend primarily on user behaviour, or are driven by factors independent of it.

Mainstream left has more users than mainstream right.(Maybe left users may do a lot of hate-tweet sharing, which algorithms may then amplify further?)

A guess (esp as I'm not clear re algorithm/user dynamics). But I wonder if
- some right party HQs produce stickier/spreadable content than left (eg simpler language? Or do more A/B testing?)
- but CDU do not have an edge on SPD on this
- Cons do over Lab (HQ), but not on MP v MP

Hypotheses
- platform biased (by design?? Accident?)
- right parties researched & gamed system
- media bias assists right (outside Ger)
- right social media comms have something left comms lack (but not at MP level)
- more online support can backfire (if share opponents' content)

Or are there other hypotheses re this data. I feel party vs MPs, and "but more in UK, Spain, but not in Germany" must be clues, which hypotheses need to have a potential explanation of.

Is platform culpable?

Very?
- if deliberate bias (unlikely)
- if gamed by illegit means (eg things like bots)

Somewhat
- If doesnt know why
- if algo incentivises corrosive practice

No
- if diff arises from eg party testing/skill (unless an active responsibility to 'balance')

Finding out more about how the *aggregate* lack of bias between MP tweets by party relates to the characteristics of MPs (across parties) who are high boost, or medium & low boost could provide important clues about the drivers of the party content algo bias or impact effect.

Caveat. 5 month study (April to Aug 2020)
- first quarters of lockdown
(Was party content different to normal? Eg opposition less partisan?)
- the key research Q is whether partisan effect is steady, narrower or bigger in heightened times (esp campaigns year & short campaign)

These impacts would matter more (whether attributed to platform bias, or to party efficacy) if they replicate across platforms, and if they are strong in campaign periods. Those are the key "known unknowns" arising from this preliminary research.

The Canada Liberal/Conservative gap is wider than the UK Labour/Conservative gap, with Spain also above average.

https://twitter.com/JohnMacGXxxx/status/1451450637380694028?s=20

Thread from the research team

https://twitter.com/ruchowdh/status/1451250772356698122?s=20

Full thread

https://twitter.com/ruchowdh/status/1451248122605146112?s=20

Guardian: a straight, nuanced report on the research findings
theguardian.com/technology/202…