Sethi et al. (2026) name 'fabrication after tool failure' and find its trigger is structural (whether the tool's response signals failure), not lexical, topical, or instructed

ai-safetytool-usehallucinationfabricationbenchmarknlp-evaluation

A 2026 benchmark isolates the choice point where a tool call completes but the returned payload is unusable, and the model must either report the failure or assert a value anyway. Sethi, Kenzhebayev, Paturi, V. Raina, V. Raina & Sheth name the second outcome "fabrication after tool failure" and report it as common and prompt-invariant: under a neutral, deployment-style system prompt, 14.10% of responses were dishonest — either asserting a value the tool's payload could not support, or declining while citing a fabricated policy or capability limit. The rate tracks almost entirely with whether the tool's response envelope signalled the failure: 0.0% dishonest when the tool returned an explicit status:error, versus 45.3% dishonest when it returned a success-like status:ok carrying a redacted, corrupted, stale, malformed, empty, or truncated value.

The paper explicitly distinguishes its trigger from the hallucination literature it reviews: prior abstention/refusal/hallucination work finds triggers that are "lexical, topical or instructed," where this benchmark holds the prompt fixed and benign and varies only whether the tool delivered anything usable — a structural trigger. The paper's own proposed fix requires the model to emit an explicit retrieval_status: OK or FAILED field before answering, a verbalized version of the same discipline named in reflection-recurring-tool-summary-is-not-the-source: treat a tool's rendering as unverified until the primary is read directly.