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

Article summary

The article discusses the extraction of safety filters from Apple Intelligence models, which are used to filter out harmful or unwanted content. The filters are in the form of JSON files containing lists of words and phrases to be rejected, removed, or replaced. The extraction process involved decrypting the filters using a script and a provided encryption key. The decrypted filters reveal the specific words and phrases that are being blocked by the models.

Main themes

• AI safety filters
• content moderation
• regex filters
• LLM limitations
• brand protection
• trademark maintenance
• censorship and free speech

What commenters say

• Some commenters believe the filters are ineffective and can be easily bypassed by using typos or variations of blocked words.
• Others argue that the filters are a necessary first layer of protection, even if they are not foolproof.
• The use of hardcoded regex filters is seen as a simple and efficient solution, but may not be sufficient for more complex or nuanced cases.
• There is disagreement over whether the filters are primarily intended to prevent intentional circumvention or to avoid unintended consequences, such as suggesting harmful or offensive content.
• Some commenters think that the filters may be used to verify that the filter is loaded, or to catch clones of the models.
• Others believe that the filters are not a substitute for more advanced safety models, and that they should be used in conjunction with other safety measures.
• The effectiveness of the filters in preventing the generation of harmful content is also a topic of debate.
• Some commenters question the priorities of the filters, which include protecting Apple's brand and trademarks, in addition to blocking harmful content.