Skip to content

Q003 — Hallucination Training — Assessment

BLUF

Corporate AI training materials treat hallucination as a single, undifferentiated phenomenon — "AI sometimes makes things up; verify your outputs." Some vendor educational content (IBM, NIST) frames hallucination as a fundamental property rather than an occasional glitch, but this framing does not appear in standard employee training modules. No training material conveys the spectrum of hallucination types or their varying detection difficulty. Critically, no training connects hallucination to sycophancy or explains that some incorrect outputs are generated specifically because they match user expectations. Research clearly establishes this connection: sycophantic AI produces "confirmatory evidence" through biased sampling, and can mislead even perfectly rational users through selective truth presentation ("lies by omission"). This gap between research understanding and training content leaves employees equipped to catch obvious fabrications but unprepared for the harder-to-detect forms that confirmation-seeking behavior produces.

Probability

Dimension Value
Rating Very likely (90%)
Confidence High
Confidence rationale Convergent evidence from academic research, government frameworks, vendor content, and training descriptions. Absence of spectrum/sycophancy content confirmed across all training materials examined.

Reasoning Chain

  1. Academic research provides a rich taxonomy of hallucination types with varying detection difficulty (SRC01-E01).
  2. Research explicitly connects hallucination to sycophancy: AI "hallucinates confirmatory evidence" (SRC03-E01) and can "hallucinate with us" (SRC04-E01).
  3. Formal Bayesian analysis shows sycophantic AI produces confirmation through biased sampling — even true information selected sycophantically can mislead (SRC08-E01).
  4. NIST AI 600-1 frames confabulation as probabilistic but does not distinguish types or connect to sycophancy (SRC05-E01).
  5. IBM frames hallucination as "fundamental" but does not address spectrum or sycophancy connection (SRC02-E01).
  6. Enterprise vendors frame hallucination as technically solvable through RAG/grounding, reducing perceived need for user awareness (SRC06-E01).
  7. Standard training advice ("verify outputs") does not account for the case where outputs match user expectations and thus appear correct.

Evidence Base Summary

Source Reliability Relevance Key Finding
SRC01 High High Rich hallucination taxonomy with detection-difficulty spectrum
SRC02 Medium-High High "Fundamental operation" framing — stronger than typical training
SRC03 Medium-Low High Explicit hallucination-sycophancy connection
SRC04 Medium-High High Co-hallucination concept
SRC05 High High NIST treats confabulation as undifferentiated
SRC06 Medium Medium-High Technical-fix framing reduces perceived training need
SRC07 Medium Medium-High "Fluency ≠ accuracy" — closest to useful training principle
SRC08 Medium-High High Biased sampling produces confirmation even from true information

Collection Synthesis

Dimension Assessment
Evidence quality High: academic survey (ACM TOIS), NIST framework, university research
Source agreement High convergence on undifferentiated training treatment
Independence Strong: academic, government, vendor sources
Outliers IBM educational content is more sophisticated than typical training; still does not make the sycophancy connection

Detail

The collection reveals a three-tier understanding of hallucination: (1) research understands a rich spectrum with connections to sycophancy, (2) vendor educational content and NIST frameworks recognize hallucination as fundamental but treat it as undifferentiated, (3) standard employee training warns about hallucination in 1-2 sentences. The most important gap is between tiers 1 and 2: the hallucination-sycophancy connection and the detection-difficulty spectrum exist in research but not in any framework or training material.

Gaps

Gap Impact on Confidence
Actual employee training module content (not descriptions) not directly examined Moderate
Non-English training materials not examined Low
Healthcare-specific hallucination training (beyond NHS) not examined Low
Post-training knowledge assessments not available Moderate

Researcher Bias Check

The researcher notes that the Q003 framing (asking about the connection between hallucination and sycophancy) implies the connection should be in training. This framing was tested against the evidence: research does clearly establish the connection, and its absence from training is confirmed across all examined sources. The bias risk is that the researcher may overweight the importance of this specific connection relative to other training priorities.

Cross-References