... Pardon me, but are you under the impression that most neural networks are purely unsupervised learning with no opportunities for operators to bias their results?
If so, this is is a misapprehension. Even unsupervised learning systems are subject to data based biasing. The selection of input, the decisions on how to partition the dataset, the decisions made on how to judge and measure overfitting, and potentially hundreds of other small hyperparameter decisions made by operators can substantially change the output of the system. In the case of the top post, it's very clear that "does this increase time-on-site" was an operator-chosen scoring function for the results of the system in question.
Frighteningly, these decisions are often unexamined and unrecorded. This has lead to a series of impressive-sounding systems that can neither be reproduced nor truly audited. [0]
And that's just the approximated function. The tensors it outputs are then further processed in some way, as they're often of no value in isolation. The decisions on how to use those outputs also have a profound impact on how we view the results of learning systems.
If anything, we're not cautious enough as a society about using this technology. We see all sorts of crazy in the news that's wrapped up in the over-hype. We see law enforcement failing to use even basic tools [1] because they're so excited.
You're right of course, though that's still quite a bit of distance from the deliberate political interference explained in the OP. I find business goals like increasing engagement to be morally neutral vs political goals like shaping people's views.
If so, this is is a misapprehension. Even unsupervised learning systems are subject to data based biasing. The selection of input, the decisions on how to partition the dataset, the decisions made on how to judge and measure overfitting, and potentially hundreds of other small hyperparameter decisions made by operators can substantially change the output of the system. In the case of the top post, it's very clear that "does this increase time-on-site" was an operator-chosen scoring function for the results of the system in question.
Frighteningly, these decisions are often unexamined and unrecorded. This has lead to a series of impressive-sounding systems that can neither be reproduced nor truly audited. [0]
And that's just the approximated function. The tensors it outputs are then further processed in some way, as they're often of no value in isolation. The decisions on how to use those outputs also have a profound impact on how we view the results of learning systems.
If anything, we're not cautious enough as a society about using this technology. We see all sorts of crazy in the news that's wrapped up in the over-hype. We see law enforcement failing to use even basic tools [1] because they're so excited.
[0]: http://science.sciencemag.org/content/359/6377/725
[1]: https://gizmodo.com/defense-of-amazons-face-recognition-tool...