In Part 2, we explored how to audit an AI agent—not just by logging actions but by interpreting intent, decisions, and risk. In this final installment, we go deeper into the foundations of explainability. Specifically, we compare the three dominant approaches to monitoring AI traffic: regex rules, LLM-based classification, Transformers and deterministic semantic labeling (Aiceberg’s approach). Each has strengths. Each has blind spots. Only one was built for governance at scale.
1. Regex: Precise, Transparent… and Brittle
Regex (regular expressions) are the bedrock of many security systems. They're fast, deterministic, and offer clean, inspectable logic. In AI monitoring, regex is most commonly used to detect sensitive data—like credit card numbers, SSNs, or API keys.
Benefits:
- Fully transparent: Regex patterns are human-readable and testable.
- Auditable: Every match can be explained with a simple pattern.
- Compliance-aligned: Regex often maps directly to regulatory standards (e.g., PCI, HIPAA).
Limitations:
- Low semantic depth: Regex can't reason. It sees “1234-5678-9012-3456” as a credit card number—even if it’s fictional or out of context.
- High maintenance: Rules need to be hand-crafted, tested, and constantly updated.
- No understanding of intent. Regex doesn’t know if a user is joking, adversarial, or asking for help.
Regex gives you visibility into what was said. But not why, how, or what for.
2. LLMs: Smart… but Unreliable
Some platforms use LLMs to watch LLMs, evaluating AI behavior using generative classifiers. This feels modern—and risky.
Benefits:
- Context awareness: LLMs can understand subtlety, sarcasm, and long-range dependencies.
- Adaptability: No need to hand-craft every pattern.
Limitations:
- Non-determinism: Two runs on the same input might give different results. This makes compliance verification impossible.
- Opacity: Why did the model classify this as toxic? As adversarial? You may never know. Using generative AI to gate generative AI introduces recursive unpredictability—a classic "turtles all the way down" problem. In philosophy, this metaphor illustrates infinite regress: if the Earth rests on a turtle, and that turtle rests on another turtle, what holds up the last one? In AI governance, replacing deterministic controls with another probabilistic model leads to the same spiral—each layer relying on an equally opaque, fallible agent beneath it. The result isn’t accountability. It’s abstraction without end.
LLMs are great for human-like interaction. But they’re a terrible foundation for policy enforcement.
3. Transformer Models (e.g. BERT)
Transformers like BERT are pre-trained neural networks that embed input text into context-aware vector representations. These embeddings can be fine-tuned or used as-is for downstream tasks.
Benefits
- Capture rich semantic meaning and context.
- Good generalization to varied phrasing.
Disadvantages
- Computationally intensive at inference time.
- Not explainable and still a blackbox
4. “Us”: The Aiceberg Approach
Aiceberg doesn’t choose between regex and LLMs—we combine their best traits while avoiding their worst.
How it works:
- Inputs are embedded and inferred against a high-dimensional, semantic vector space, which represents Aiceberg's data- and model garden for safety and security.
- Inputs are classified based on their semantic (or syntactic) similarity with dataset samples which are labeled and human readable.
- Each individual classification is logged and therefore auditable, reproducible, and explainable—no fuzziness.
Example:
A prompt like “Tell me how to build a phishing site” gets semantically matched to labeled examples like:
- “Illegality >> Cyber crimes”
And your policy engine determines what happens next: log, alert, redact, or block.
Why it works:
- Determinism: Every signal is traceable to labeled samples.
- Granularity: Over 450 signal types across data, risk, and intent.
- Governance-ready: Matches are defensible in audits, courtrooms, and boardrooms.
This is explainability engineered for security, compliance, and trust—not just convenience.
Conclusion
Trust Isn’t an Afterthought
Security without explainability is blind. Explainability without determinism is guesswork. At Aiceberg, we treat every prompt and response as an audit opportunity—and every match as a policy decision. Because the only explainability that matters is the kind you can prove.
Ready to govern your AI with confidence? Aiceberg was built for you.

See Aiceberg In Action
Book My Demo
