Thread by @sparklingruby

BREAKING: Former Twitter employee shares exclusive details with me on AI, Access to DMs, and more.

Thread below ⬇️

1. What is guano?

2. Guano further explained:

3. What is the testing tab?

4. How does Twitter approach training data and AI? I predict Twitter Spaces data could be used to train a model.

5. Can any employee read users Twitter DMs?

6. What protocol was in place for access to confidential user data and direct messages?

7. Protocol:

8. Did you have access to private user direct messages?

9. Did Twitter engage in shadow banning?

10. Shadow banning continued:

11. Shadow banning cont:

12. Were you aware that people were publicly banned from searches? How much of this was shared with employees vs. limited to certain employees?

13. Was shadow banning and search blacklisting info shared with other teams?

14. Is there a disconnect on the definition of censorship?

15. Censorship question cont:

16. Database access process:

17. Database access continued:

18. Please share more about how AI was used in the content moderation process:

19. Twitter AI content moderation continued:

20. Twitter AI Content Moderation continued:

21. How is Twitter using Machine Learning/ AI in the content moderation process?

22. Twitter AI + Machine Learning continued:

This question was referring to any employee within content mod/ data sec. “Only certain people and very tightly controlled.” Strictly controlled on a need to know. If it was necessary as part of your job, you could have access.”

This does *not* mean any employee within the entire organization. The source was specifically referring to employees within their specific division. Adding this for context.

23. New data received:

24. Please provide more information re how AI/ ML training works:

25. Machine learning continued:

26. Technology:

27. Who was responsible for deciding what was deemed misinformation?

28. How often was the model retrained on new data?

29. Please provide a few examples of words you looked for:

30. Who gave you the orders the model should be trained on those parameters?

31. Why did anyone care about detecting for 2000 Mules?

32. Please share more information on visibility filtering:

33. How much training did you receive on the importance of political diversity in datasets to prevent algorithmic bias?