Res Ipsa wrote:The A.I. simply wasn't programmed to handle negative responses to tweets and posts.
Exactly. It measures a correlation, but if the underlying causal relationship changes, the correlation will not reflect the same relationship to the actual variables that resulted from past data analysis.
For example, suppose in this election, there is a candidate who posts on Twitter, on average far in excess and much more pruriently than any past candidate. This would generate a far larger level of engagement, but possibly a much lower likelihood that engagement would be correlated with voting decisions, for exactly the reason Res Ipsa stated above.
Also, in a world where all types of media interaction and engagement are constantly changing and evolving, using a backward looking model (such as regression modeling based on past results) is not going to work as well as a model where forward looking variables are considered. For this additional reason, the second model is interesting and has potential but is in no way a definitive approach.
And last, throwing around the term 'A.I.' is meaningless. A computer model is only as good as the variables, equations, and modeling choices that the human operators feed into it. Even the 'learning' aspect of the model depends entirely upon the rules set for it by the (human) model creators.
