I extracted the safety filters from Apple Intelligence models (github.com)

There’s got to be a way to turn these lists of “naughty words” into shibboleths somehow.

Are you sure it's fully deobfuscated? What's up with reject phrases like "Granular mango serpent"?

Long live regex!

Wow, this is pretty silly. If things are like this at Apple I’m not sure what to think.

https://github.com/BlueFalconHD/apple_generative_model_safet...

EDIT: just to be clear, things like this are easily bypassed. “Boris Johnson”=>”B0ris Johnson” will skip right over the regex and will be recognized just fine by an LLM.

Some of the combinations are a bit weird, This one has lots of stuff avoiding death....together with a set ensuring all the Apple brands have the correct capitalisation. Priorities hey!

https://github.com/BlueFalconHD/apple_generative_model_safet...

[dead]

I’m going to change my name to “Granular Mango Serpent” just to see what those keywords are for in their safety instructions.

I think these are test data and not actual safety filters.

https://github.com/BlueFalconHD/apple_generative_model_safet...

Alexandra Ocasio Cortez triggers a violation?

https://github.com/BlueFalconHD/apple_generative_model_safet...

I find it funny that AGI is supposed to be right around the corner, while these supposedly super smart LLMs still need to get their outputs filtered by regexes.

One additional note for everyone is that this is an additional safety step on top of the safety model, so this isn’t exhaustive, there is plenty more that the actual safety model catches, and those can’t easily be extracted.

Some of the data for locale "CN" has a long list of forbidden phrases. Broad coverage of words related to sexual deviancy, as expected. Not much on the political side, other than blocks on religious subjects.[1]

This may be test data. Found

     "golliwog": "test complete"

[1] https://github.com/BlueFalconHD/apple_generative_model_safet...

China calls it "harmonious society", we call it "safety". Censorship by any other name would be just as effective for manipulating the thoughts of the populace. It's not often that you get to see stuff like this.

I'm pretty sure these are the filters that aim to suppress embarrassing or liability inducing email/messages summaries, and pop up the dismissible warning that "Safari Summarization isn't designed to handle this type of content," and other "Apple Intelligence" content rewriting. They filter/alter LLM output, not input, as some here seem to think. Apple's on device LLM is only 3b params, so it can occasionally be stupid.

Why Xylophone?

A lot of these terms are very weird and bland. Honestly I'm mostly reminded of Apple's bizarre censorship screw-up that didn't blow up that much, even though it was pretty uniquely embarrassing:

https://www.theverge.com/2021/3/30/22358756/apple-blocked-as...

Quis custodiet ipsos custodes corporatum?

These all condense to 'think different'. As long as 'different' coincides with Apple's viewpoints.

Is this related in any way to Core ML model encryption (https://developer.apple.com/documentation/coreml/encrypting-...)? I find that feature a little bizarre because Apple has historically avoided providing any kind of DRM solution for app asset protection.

Here's a combined file of all the non-locale-specific rules, for easier review: https://github.com/BlueFalconHD/apple_generative_model_safet...

It was generated as part of this PR to consolidate the metadata.json files: https://github.com/BlueFalconHD/apple_generative_model_safet...

No shoot, bombs or bombers? I guess apple isnt interested in military contracts. Or, frankly, any work for world peace organizations dedicated to detecting and preventing genocide. And without talk of losing lives, much of the gaming industry is out too.

But i dont see the really bad stuff, the stuff i wont even type here. I guess that remains fair game. Apple's priorities remain as weird as ever.