I expected the Assistant to do very poorly and was surprised that for several searches it agreed with my assessment perfectly. I have seen several research summaries showing LLMs cannot identify LLM-generated content, but they seem have nearly perfect accuracy for the slop littering my search results.
The procedure I'm suggesting does not need to scale beyond individual searches, and only run when requested by the user.
There is not much of an adversarial relationship or arms race for this implementation, because Kagi is presently too small of a player for these slop site creators to care. The arms race is present with manual filtering and whatever other methods employed with SlopStop anyway.
Assuming the model purveyors are respecting their contracts with Kagi, with this Assistant detection, they are legally obligated not to "read" (meaning use for training subsequent models) the slop websites or be informed by the slop.
Of course, Kagi SlopStop should work differently from what I'm suggesting. My suggestion is a shortcut that does two things well:
- Aggressively filters out likely slop NOW, without relying on reporting
- Reports potential slop for manual review by the SlopStop team, so it can be used to improve SlopStop
The biggest and only serious drawback is that it takes a long time after conducting a search. However, I am presently wasting a lot of time wading through slop, so the method I'm proposing would still save me time.
I think perhaps it would be nice to review each item before reporting to SlopStop, but so far my experience testing has shown it to be accurate.
