Why AI sometimes makes mistakes

The systematic root causes of AI failure: hallucinations, training distribution shift, adversarial prompt injection, ambiguity, and lossy compression of knowledge in parameter weights.

Real-World Mental Model

“An eager student guessing on a test question they never studied: they write very confident-sounding paragraphs, but the underlying facts are completely fabricated.”

INTERACTIVE ENGINE: FAILURE MODES

Live Reactive

Hallucination & Confabulation

HIGH // Plausible falsehoods delivered with high confidence.

Root Engineering Cause:

Model is an autoregressive token predictor optimizing grammatical fluency, not an ontological database of truth.

Production Mitigation:

Implement Grounded RAG with strict citation verification and zero-temperature decoding.

CORE PRINCIPLES

What You Must Understand

Hallucination: Model generates syntactically fluent text with false factual claims.

Distribution Shift: Real-world inputs differ from training data distributions.

Context Length Degradation: "Lost in the Middle" attention degradation in long prompts.

Reference Implementation

// Guardrail against hallucination
if (!groundedInContext(response, retrievedDocuments)) {
  fallbackToHumanReview();
}

Production Failure Mode

Assuming newer or larger models have zero hallucinations without implementing automated verification and RAG citation checks.

View Full 26-Topic Curriculum Table

Loading AI Engineering Foundations...