|
|
|
|
|
|
by jimmygrapes
1477 days ago
|
|
|
Forgive my skepticism, but I've seen this before. The Knife of Aristotle[1] made similar claims and actually seemed to be doing a good job, until it was discovered to be a front for a sex cult[2]. One of their primary methods of highlighting bias in news media was to identify unnecessary adjectives and loaded terminology, which is IMO still a valid concept to paint a picture not necessarily wrong but intentionally misleading slant (e.g. "The deadly insurrection on the U.S. Capitol" is still an oft-used phrase). I am also not convinced by vague "we use machine learning" marketing terms without some very specific details about the training data, methodology, algorithms, fitting, human curation, etc. How does The Daily Edit differ, and what assurances does a user have that it's not a similar front for covering up stories your staff doesn't want to be propagated? [1]: https://mediabiasfactcheck.com/the-knife/
[2]: https://artvoice.com/2018/08/06/the-knife-media-has-ceased-p... |
|
|
We differ by giving the reader the tools to form a better opinion for themselves rather than telling them what's wrong and making ourselves some oracle or source of truth. We think that's where so many of these attempts go wrong.
Our highlights show related details that were found in other articles, but were not covered in the article being read. We show all of the sources of these so the user never has to take our word for it, they can go there and see. This part uses an ML model trained on the MNLI and SNLI datasets. I'm happy to share all the details of the graph and training method privately if you send us an email.
We do also highlight some passages of text for being potentially misleading but this uses no ML whatsoever. We suggest missing sources of data ("according to our sources"), missing reference ("a recent study suggests" - with no reference or link), and scare quotes. We make sure that the item is verifiable by the text the user is actually looking at. We do not attempt to cover subjective items like hyperbole, slippery slope fallacy, etc. Each of these items is presented as a suggestion intended to make the reader pause for a second.
Your last question is hard to answer since I'm not sure what assurances anyone could give apart from "we promise", I'll do my best though. We're a startup of 6 people who all care deeply about the quality of news. If we fail to cover an article you should call us out on it and find out why then share that with everyone.
Right now there are a number of technical reasons why that might happen such as the insane mess of HTML some publishers use, it can be hard to parse.