Inspiration
Like many of us, we're tired of low quality mass produced "slop" on the internet. We made this extension to better identify this content, while preserving privacy. Existing extensions for this goal send your data to servers; here, all your browsing data is processed locally, so you never have to worry about some random remote server having all your browsing history.
What it does
Our algorithm performs several different levels of checks.
- Is a site already in a list of low-quality websites? [DEFINE THIS]
- Does an article's content closely match content from a known low quality website?
- Does this page's text content register as "likely AI" using statistical analysis?
- If the site contains comments/reviews; e.g. YouTube comments or Amazon reviews, are they "suspiciously" correlated? That is, [GIVE AN EXAMPLE OF THIS]
How we built it
This tool is packaged as a Chrome extension, written in Typescript and built with Vite.
We use a variety of algorithmic and statistical techniques to flag suspected "slop."
To start, we store a list of low quality websites. This is pulled from github1 and github2. We check if a domain is in this list efficiently via storing all these sites in a bloom filter.
For detecting pages which are similar to known "slop" articles, we utilize locality sensitive hashing to generate a "fingerprint" for each page. That fingerprint is compared to
Challenges we ran into
[Getting it to integrate naturally into web pages, without being too visually cluttered/obstrusive]
Accomplishments that we're proud of
It works! This extension clearly labels content deemed likely to be slop, and it also doesn't noticeably slow down the browser. We also have observed very few false positives so far, which is a great sign!
What we learned
Perhaps unsurprisingly, processing all user data locally is more limited. We can't send off data to servers with powerful GPUs; these constraints forced us to implement more efficient techniques to detect slop. This limitation made the project more interesting to work on.
What's next for SlopSniffer
Image detection is next on our list. [ELABORATE]
We would also love to tune our algorithms more. In this project, there's a balance between performance, false positives, and false negatives. Testing these algorithms on more data would give us a better sense of how we should tweak it.
