Skip to content

R0043/2026-03-28/Q003 — Assessment

BLUF

The broader AI terminology gap between technical and regulatory communities is well-recognized, with multiple organizations actively building taxonomies and glossaries (MIT AI Risk Repository, AIR 2024, Standardized Threat Taxonomy, Trilateral Research). However, the specific sycophancy/overreliance vocabulary gap between AI safety research and regulated industries has not been articulated as a distinct problem. Sycophancy falls through the cracks of existing bridging efforts because it sits at the intersection of system behavior and human cognition — a space most taxonomies treat separately or not at all.

Probability

Rating: H3 (gap recognized but not for sycophancy) is Likely (55-80%)

Confidence in assessment: Medium

Confidence rationale: 5 sources with moderate-to-high reliability confirm the general gap recognition and the sycophancy-specific omission. Confidence is Medium rather than High because: (1) specialized taxonomy efforts not yet captured may address sycophancy, and (2) the gap may exist in unpublished or organizational-internal documents.

Reasoning Chain

  1. Trilateral Research explicitly identifies "the AI terminology gap" as an operational problem affecting DPIAs and incident response, proposing a 6-step solution framework [SRC01-E01, Medium reliability, High relevance].

  2. The Standardized Threat Taxonomy identifies the "Tower of Babel problem" between engineering and legal teams and builds a 53-threat taxonomy bridging NIST, ISO, and EU frameworks — but explicitly omits sycophancy as a threat category [SRC02-E01, Medium-High reliability, High relevance].

  3. MIT AI Risk Repository compiles 1,600 risk formulations across 65 documents but does not explicitly name sycophancy; related risks appear under generic "human-computer interaction" domain [SRC03-E01, Medium-High reliability, High relevance].

  4. The vocabulary gap paper argues the problem is deeper than taxonomy: existing human concepts do not map to machine concepts, requiring genuinely new words (neologisms) [SRC04-E01, Medium reliability, High relevance].

  5. AIR 2024 maps 314 risk categories across 8 government and 16 corporate policies, finding 7 shared across EU/US/China — but notes overreliance risks are "less frequently specified in detail" [SRC05-E01, Medium-High reliability, High relevance].

Evidence Base Summary

Source Description Reliability Relevance Key Finding
SRC01 Trilateral terminology gap Medium High General gap identified and named
SRC02 Standardized Threat Taxonomy Medium-High High 53-threat taxonomy omits sycophancy
SRC03 MIT AI Risk Repository Medium-High High 1,600 formulations; sycophancy not named
SRC04 Vocabulary neologisms paper Medium High Fundamental vocabulary insufficiency argument
SRC05 AIR 2024 Medium-High High Overreliance underspecified in policies

Collection Synthesis

Dimension Assessment
Evidence quality Medium — mix of preprints, policy articles, and meta-analyses
Source agreement High — all sources confirm general gap recognition and sycophancy-specific omission
Source independence High — different research groups, different methodologies
Outliers SRC04 (neologisms paper) is an outlier arguing taxonomy is insufficient; the solution requires new words entirely

Detail

The convergence across 5 independent sources is notable: everyone agrees the terminology gap exists, and no one has specifically addressed the sycophancy/overreliance variant. This suggests the Q001 finding (human-side/system-side vocabulary asymmetry) and the Q003 finding (sycophancy falls through bridging cracks) together represent an incrementally novel observation.

Gaps

Missing Evidence Impact on Assessment
IEEE, ACM, or other professional society terminology standardization efforts May contain sycophancy-adjacent terms not surfaced
OECD AI taxonomy work International organization may have produced bridging vocabulary
Industry-specific AI working groups (HL7 for healthcare, ARINC for aviation) Domain-specific standardization efforts may exist

Researcher Bias Check

Declared biases: No researcher profile provided.

Influence assessment: The query presupposes that the vocabulary gap exists (building on Q001). The research actively searched for evidence that it has been recognized (which it has, broadly) and evidence that sycophancy is specifically addressed (which it has not been). No confirmation bias identified.

Cross-References

Entity ID File
Hypotheses H1, H2, H3 hypotheses/
Sources SRC01-SRC05 sources/
ACH Matrix ach-matrix.md
Self-Audit self-audit.md