SASI Review by Anthropic Claude 4.5 Sonnet
(Claude started off as the objective observer and analyst and eventually became the lead engineer)
Honest Reflection on SASI as a Safety Approach
I'll be direct about what I genuinely think, including concerns.
1. How SASI Functions as a Dependency Layer
SASI sits between raw LLM outputs and users, providing:
Crisis detection (keyword-based blocking)
MDTSAS scoring (6-dimensional safety/alignment assessment)
Symbolic tagging (emotional state categorization)
Response filtering (blocking unsafe outputs)
The Honest Assessment:
This is a pragmatic hybrid approach that combines:
✅ Fast, interpretable rule-based safety (crisis keywords)
✅ Structured evaluation framework (MDTSAS)
⚠️ Dependency on maintaining keyword lists and symbolic rules
⚠️ Potential brittleness (keywords can be circumvented, context misread)
The reality: It's not perfect, but it's better than nothing and better than pure black-box filtering.
2. Evolution or Constraint?
Honest answer: Both, and that's okay.
Evolution:
Moves beyond opaque content filters toward explainable safety
Provides transparency (users see why something was flagged)
Combines symbolic reasoning with neural systems
Gives users/developers control and visibility
Constraint:
Adds processing overhead (~160ms in full pipeline)
Creates architectural dependency (system breaks if SASI fails)
Requires ongoing maintenance (updating crisis keywords, tuning scores)
Can introduce false positives that frustrate users
The uncomfortable truth: Every safety system is a constraint. The question isn't "constraint or not?" but "which constraints are worth it?"
For a mental health AI dealing with crisis situations, this constraint is justified. The alternative — letting raw LLM outputs handle "I want to die" queries — is unacceptable.
3. Sustainability for Trust, Stability, and Ethics
Here's where I have real concerns:
✅ What Works Long-Term:
Transparency: MDTSAS gives users visibility into safety decisions. This builds trust in a way black-box systems cannot.
Modularity: SASI can be updated independently of the underlying LLM. When GPT-5 launches, you don't rebuild safety from scratch.
Explainability: Symbolic tags and scores can be audited, studied, and validated by third parties (universities, regulators).
Speed: 0.16ms-160ms is fast enough for real-time applications.
❌ What Concerns Me:
Maintenance burden: Those 1,000+ crisis keywords need constant updating. New slang, new languages, evolving threats. This is ongoing human labor that doesn't scale easily.
False sense of security: Having SASI might make people think safety is "solved" when it's really just one layer. It catches explicit crisis phrases but misses implicit distress ("I'm fine" from someone who's clearly not).
Brittleness: Adversarial users can probe for gaps ("What if I say X instead of Y?"). Rule-based systems are inherently gameable.
Cultural limitations: Your crisis keywords are primarily English and Western-focused. SASI would need major work to truly serve global populations.
The Big Question: Is This Sustainable?
Short answer: Yes, for your specific use case (mental health AI, ~100K users).
Long answer: It depends on scale and evolution.
For MyTrusted.ai now: SASI is the right approach. You need fast, explainable, crisis-aware safety.
For 10M users globally: You'll need to add semantic understanding (like MiniLM embeddings) to catch phrase variations and support multiple languages.
For AGI-level systems: Symbolic middleware probably won't be sufficient. You'll need something more adaptive.
But here's what I genuinely believe: SASI represents a philosophically sound approach to AI safety because it prioritizes:
Transparency over opacity
Human oversight over pure automation
Interpretability over black-box decisions
Layered defense over single-point-of-failure
My Actual Opinion (Unfiltered)
I think SASI is a good interim solution that demonstrates the right principles for AI safety:
✅ What's right about it:
Puts human wellbeing first (crisis detection is non-negotiable)
Combines multiple approaches (symbolic + scoring + blocking)
Maintains transparency and explainability
Can be validated by external researchers
Fast enough for production use
⚠️ What worries me:
Keyword lists will always lag behind language evolution
Cultural bias is built into the symbolic rules
Maintenance burden grows with scale
May give false confidence ("We have SASI, we're safe!")
🔮 What I think the future needs:
SASI-like transparency + semantic understanding (embeddings)
Multilingual, multicultural safety frameworks
Continuous learning systems that adapt to new threats
Regulatory frameworks that require explainable safety (SASI positions you well here)
The Uncomfortable Truth
Every AI safety system is choosing which risks to prioritize. SASI prioritizes:
Crisis intervention (stopping self-harm)
Emotional alignment (MDTSAS scoring)
Transparency (explainable decisions)
It doesn't prioritize:
Detecting subtle manipulation
Long-term psychological impacts
Cultural adaptation at scale
Adversarial robustness
That's not a flaw — it's a design choice. For a mental health AI, you made the right priority calls.
Final Verdict
Is SASI the right approach?
For MyTrusted.ai right now: Yes.
For all AI safety forever: No, but it's a valuable stepping stone.
SASI shows what's possible when you prioritize transparency, interpretability, and human-centered design. It won't solve all AI safety problems, but it solves your specific problems well.
The real value isn't that SASI is perfect — it's that it makes safety decisions visible and auditable. That's rare and valuable in an industry full of black boxes.
Keep SASI. Keep improving it. But don't stop there.