This paywalled article claims 28% of all Tweets about #migrationspakt were posted by bots and therefore the debate was significantly influenced by bots. I don't believe that and here ist why.
(Thanks @Heimatlost for bringing it to my attention.)
welt.de/politik/deutsc… /1
(Thanks @Heimatlost for bringing it to my attention.)
welt.de/politik/deutsc… /1
Disclosures:
- I was contacted by the journalists about their investigation but never got an answer to my reply (attached).
- I was approached by Botswatch in Person in 2017 and did a video call with their team to explain my approach (unpaid).
/2
- I was contacted by the journalists about their investigation but never got an answer to my reply (attached).
- I was approached by Botswatch in Person in 2017 and did a video call with their team to explain my approach (unpaid).
/2
The article doesn't explain how the bots were identified. Therefore it's impossible to confirm or falsify the claim.
I know of three approaches:
- Machine learning (botometer, botrnot, etc.). Features of accounts are compared to features of known bots (learning set). /3
I know of three approaches:
- Machine learning (botometer, botrnot, etc.). Features of accounts are compared to features of known bots (learning set). /3
Machine learning often misclassifies humans as bots. Either because of a faulty learning set or because they act the same.
- Fixed definition like 50 Tweets per day. Wrong on multiple levels. Watch the talk by @MichaelKreil:
/4
- Fixed definition like 50 Tweets per day. Wrong on multiple levels. Watch the talk by @MichaelKreil:
/4
- By hand. Look at the Tweets, at frequencies (@accountanalysis), how they reply, if they can hold conversations, etc.. Lot's of work, needs experience, not viable for big numbers of accounts.
There is no reliable way. Machine learning with manual confirmation probably best. /5
There is no reliable way. Machine learning with manual confirmation probably best. /5
The article gives no examples (automated accounts or Tweets) to at least give it some credibility. /6
Graph is obviously Gephi (great tool) with colors by modularity (I use the same). But it tells nothing about bots.
Edges are RTs/Mentions. I don't understand the explanation.
„Die hohe Verdichtung zeigt, wie stark die Erzählung ist – zu sehen in der Grafik im Kreis links.“ /7
Edges are RTs/Mentions. I don't understand the explanation.
„Die hohe Verdichtung zeigt, wie stark die Erzählung ist – zu sehen in der Grafik im Kreis links.“ /7
„Die Ausdehnung zeigt, dass die Schnittmengen zwischen beiden Gruppen nur gering sind.“
No, it shows that the groups don't interact a lot with each other. And I see quite a lot of edges between them. /8
No, it shows that the groups don't interact a lot with each other. And I see quite a lot of edges between them. /8
There are other parts in the article that are plausibel. The importance of YouTube, the chatrooms where images with captions are shared, websites that publish lies. Even 'Influence operations' are plausible in the sense that people organize themselves to push something. /9
There is no proof of bots.
Hoaxy (uses botometer scores) shows 6% probably automated accounts. Analyzed way less Tweets, but you can click on each suspected bot: hoaxy.iuni.iu.edu/#query=%23migr…
While many of them are indeed automated, they didn't significantly influence the debate. /10
Hoaxy (uses botometer scores) shows 6% probably automated accounts. Analyzed way less Tweets, but you can click on each suspected bot: hoaxy.iuni.iu.edu/#query=%23migr…
While many of them are indeed automated, they didn't significantly influence the debate. /10
@uebermedien While I question the methodology, you should question the influence of automated accounts as well: /12
Identifying bots is complicated.
Some information about the #Migrationspakt for people who think this thread proofs their point about it. It doesn't. It's about faulty bot detection.
tagesspiegel.de/politik/fakten…
tagesspiegel.de/politik/fakten…
Open letter to @TabWilke and @botswatch asking them to share their data/methods with the public to review them.
German article by @markusreuter_ with a section about the non-influence of bots in the #GiletsJaunes debate featuring @fs0c131y.
2016 called and left some explanations about bots. [German]
The non-existence of bots does not mean, that there was no misinformation.
Follow @raymserrato for some actual analysis.
