AXIØM.talk update 26.3–Trust & Safety

I’ve launched the latest update to my Personal, Digital Friend app which runs on-device with iPhone® models that support Apple Intelligence® and my main focus was improving the Trust & Safety of the experience.

AI Safety

We’re already familiar with the concept of safety and moderation from our experiences with social media. Over the past few years, governments around the world have rolled out new regulations imposing stricter content and access standards for minors on both platforms like Facebook/Instagram, and on app platforms like Apple’s App Store® and Google Play.

Now, AI Safety is quickly becoming part of the popular zeitgeist, as debates about misinformation, algorithmic bias, and automated decision-making enter mainstream awareness. Just as society adapted to the idea of platform accountability in the social media era, we are beginning to internalize that AI systems need proactive measures to ensure they operate responsibly and align with common-sense values.

AXIØM.talk update 26.3 includes 3 new Trust & Safety features

1. Trust & Safety Agent–runs every time on the on-device LLM (the same that powers Apple Intelligence®) responds to the user

   1. It checks both the immediate response, as well as the tone of the overall message history, to ensure it is not steering or unintentionally enforcing sycophancy or unsafe behaviour

   2. It uses a custom set of System Instructions, inspired by Anthropic’s Claude Constitution,

2. the app changes how it responds based on the user’s age, as provided to the app via Declared Age Range API, this is a privacy-preserving system that ensures age-appropriate language and tone are applied without the app learning any exact birthday of the User

3. to protect a User’s Perspective, warning the user when their duration of use becomes too long while balance user agency to use their devices and the app they paid for responsibily, I have designed a new Warning that gradually “warms up” as the user’s session goes up to the recommended time to take a break, and cannot be dismissed, rather it gradually “cools down” by the same amount of time it took to appear

The Rejected Trust & Safety Feature

I had intended to include one more feature for younger users, the ability for the app to:

1. read if the User had Parental Controls dictating a maximum Film Classification level

2. add this guidance to the system prompt for even more specificity of the tone and appropriateness of responses

Unfortunately this feature did not survive App Review, specifically Guideline 2.5.1 of taking a public API for a specific purpose and using for another purpose. I absolutely respect App Review’s position on this, as there are bad actors trying to find loopholes on Apple Platforms for ad-tracking and bypassing privacy projections all the time.

That being said, I am debating creating a FB Request that this type of data could really help developers in crafting safe AI experiences for younger users.

---

I hope you all give these new features a whirl as AXIØM.talk is currently -50% (USD $12.95 → $6.49) off for the next few days!