R0048/2026-04-01¶
Investigated what corporate and government AI training programs teach employees about AI limitations, sycophancy, and hallucination. Found that training is widespread but shallow on specific failure modes, sycophancy is completely absent from all training materials, and hallucination is addressed but characterized as random error without connection to sycophancy.
Queries¶
Q001 — Corporate AI Training Content — Medium confidence
Query: What do standard corporate AI training courses teach employees about AI limitations?
Answer: Training is widespread but covers limitations at varying depths. The DOL AI Literacy Framework names "hallucinations and accuracy limits" explicitly. The NHS framework names automation bias. Most programs address limitations generally without specific failure-mode education.
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Comprehensive failure-mode coverage | Inconclusive | — |
| H2: Widespread but shallow | Supported | — |
| H3: Perfunctory/checkbox-driven | Partially supported | — |
Confidence: Medium · Sources: 7 · Searches: 4
Q002 — Sycophancy Warnings — Medium-High confidence
Query: Do any training materials warn about sycophancy or the AI tendency to tell users what they want to hear?
Answer: No. The term "sycophancy" does not appear in any training material. The NHS framework addresses automation bias (human-side) but not sycophancy (AI-side). The gap between AI safety research awareness and training content is complete.
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Training warns about sycophancy | Eliminated | — |
| H2: Adjacent concepts only (automation bias) | Supported | — |
| H3: No related concepts at all | Partially supported | — |
Confidence: Medium-High · Sources: 6 · Searches: 3
Q003 — Hallucination Training — Medium-High confidence
Query: How do training materials characterize hallucination? Is the hallucination-sycophancy connection made?
Answer: Hallucination is the most widely addressed failure mode. The DOL names it explicitly. However, training characterizes it as random fabrication, not as a spectrum including user-expectation-confirming errors. No training connects hallucination to sycophancy despite research establishing shared neural mechanisms.
| Hypothesis | Status | Probability |
|---|---|---|
| H1: Comprehensive (spectrum + sycophancy link) | Eliminated | — |
| H2: Random-error framing only | Supported | — |
| H3: Hallucination not addressed | Eliminated | — |
Confidence: Medium-High · Sources: 5 · Searches: 2
Collection Analysis¶
Cross-Cutting Patterns¶
| Pattern | Queries Affected | Significance |
|---|---|---|
| Research-training disconnect | Q001, Q002, Q003 | AI safety research documents specific failure modes extensively; none of this knowledge appears in training |
| Gradient of specificity by domain stakes | Q001, Q002 | Healthcare (NHS) > Government (DOL) > Corporate in specificity of failure-mode coverage |
| Human-side vs AI-side framing gap | Q002, Q003 | Training addresses human behavior (verify outputs, don't overtrust) but not AI behavior (AI actively generates agreeable outputs) |
| Hallucination as named but mischaracterized | Q001, Q003 | Hallucination successfully entered training vocabulary but is understood as random error, not as a spectrum including sycophancy |
Collection Statistics¶
| Metric | Value |
|---|---|
| Queries investigated | 3 |
| Answer with Medium-High confidence | 2 (Q002, Q003) |
| Answer with Medium confidence | 1 (Q001) |
Source Independence Assessment¶
Sources across the three queries represent genuinely independent perspectives: government frameworks (DOL, GSA, NHS), regulatory (EU AI Act), corporate training (NAVEX, Deloitte, Microsoft), academic research (Stanford/Science, Tsinghua), policy analysis (Georgetown, Brookings), and professional organizations (IAPP, IPR). No single upstream source drives the findings. The convergence across these independent sources on the same pattern — training exists but does not address sycophancy or the hallucination-sycophancy connection — significantly strengthens confidence in the findings.
Collection Gaps¶
| Gap | Impact | Mitigation |
|---|---|---|
| Internal training materials (paywalled/internal) | Cannot assess actual depth of content delivery | Focused on publicly visible descriptions; noted as limitation |
| Post-training assessment data | Cannot determine whether employees learn what training covers | Out of scope for current research; flagged as revisit trigger |
| Non-English markets | EU/Asian training may differ | Searched EU AI Act; non-English content out of scope |
| Classified/government security training | Military AI safety training may address sycophancy | Not publicly accessible; acknowledged as blind spot |
Collection Self-Audit¶
| Domain | Rating | Notes |
|---|---|---|
| Eligibility criteria | Low risk | Consistent criteria across all three queries |
| Search comprehensiveness | Low risk | 9 searches, 82 results dispositioned, multi-vocabulary strategy |
| Evaluation consistency | Low risk | Same GRADE/bias framework applied to all 18 sources |
| Synthesis fairness | Low risk | Researcher bias compensation applied throughout; counterevidence highlighted |
Resources¶
Summary¶
| Metric | Value |
|---|---|
| Queries investigated | 3 |
| Files produced | ~110 |
| Sources scored | 18 |
| Evidence extracts | 18 |
| Results dispositioned | 21 selected + 61 rejected = 82 total |
Tool Breakdown¶
| Tool | Uses | Purpose |
|---|---|---|
| WebSearch | 17 | Search queries across training, policy, and research |
| WebFetch | 18 | Page content retrieval (10 succeeded, 8 failed/403) |
| Write | ~110 | File creation |
| Read | 12 | File reading (specifications, placeholder files) |
| Edit | 1 | Instance index update |
| Bash | 7 | Directory creation, placeholder generation |
Token Distribution¶
| Category | Tokens |
|---|---|
| Input (context) | ~200,000 |
| Output (generation) | ~80,000 |
| Total | ~280,000 |