balajis
@balajis.eth
I want a simple open source JS text widget that automatically detects and rejects AI input. - It constantly keeps up to date with the latest models - As the user types, it displays the probability that the input is AI - If the user pastes in text, the AI probability increases - It flat-out rejects input if AI probability is >50%, by graying out submit - It doesn't show anything unless the AI probability is high Basically, a good slice of people now paste in AI slop because they're lazy, can't write, or don't realize how detectable it is. I don't want to manually detect it with my eyes, I want AI to do that and reject it prior to submission. Put the widget up at noaijs.com or a similar domain. Allow developers to subscribe for a hosted version, or even perhaps buy a memecoin (ha!) to support.
13 replies
12 recasts
97 reactions
patxol
@patxol.eth
If the code is open sourced, then it’s useless.
1 reply
0 recast
0 reaction
Daniel Lombraña
@teleyinex.eth
Why?
1 reply
0 recast
0 reaction
patxol
@patxol.eth
Because you can’t have an efficient spam detection system that reveals its detection method. It would be instantly obsolete
2 replies
0 recast
1 reaction
Andrei O.
@andrei0x309
On the other hand if is not open-sourced it could be completely wrong. It could use any unorthodox methods like random, it may not be deterministic and so on and so forth, so not being open source could also be interpreted as useless.
1 reply
0 recast
1 reaction
patxol
@patxol.eth
If not open sourced it can be bad, I agree. Same if being open sourced. But that’s not being closed source that would make it useless.
1 reply
0 recast
1 reaction
Andrei O.
@andrei0x309
There are cases when reveling the data points will not make the algorithm completely useless. Twitter a few years back reaveled all data points. There are certain things that you can't easily gain like for example is your account from 2008? When it comes to AI because it's really impossible to not have false positives, without seeing the data points is also useless. As there's no way to prove your algorithm is even working. Anyway many researchers say it's impossible to detect LLM output reliable, if the model does not embed signature in the output. I also tend to agee with these researchers. This whole notion that you can do that reliable is contested and the author ignores that and writes as if the matter is settled.
0 reply
0 recast
0 reaction
