Bypass Censorshop of DEI Terms

fragdaspdf.de

2 points·mnewme··4 comments

It's not unusual for text classifiers and other tools to know about alternate ways to write things with alternate unicode characters. Commercial LLMs see through them just fine, other systems explicitly squash " " to "unicode italics".

I cut and pasted the output of that thing into Copilot and Copilot had no problem understanding what was being talked about.

One summer I was working on an attempted foundation model based on character-CNNs and one topic for discussion was if we should build feature vectors for characters that have a lot of knowledge built in [1] or whether we should expect the model to learn all that at the byte level.

Anyhow, AI sees right through it today and you're not hiding anything by making everything you write look like a ransom letter.

[1] that italic i would have the i bit set and the italic bit set, the Chinese character 狐 has the Chinese bit and the simplified bit and the the h bit and the u bit and the rising tone bit set with something to disambiguate it from the other characters that sound the same

yes of course, but for mass checking, I highly doubt they will use LLMs though

You are right, this is just a quick thingy to bypass shallow checks

Another possibility is to use plain words and avoid coded words. Often this gets better results.

For instance, there are some men who will, when asked "Are you gay?" say "no" but if you ask "Do you have sex with men?" they will answer "yes". [1]

[1] https://en.wikipedia.org/wiki/Men_who_have_sex_with_men

This automatic word censorship reminds me eerily of the short story "The City of Silence" by Ma Boyong.

part 1 https://web.archive.org/web/20120118104342/http://worldsf.wo...

part 2 https://web.archive.org/web/20120201032550/http://worldsf.wo...

Reply on news.ycombinator.com