Hazard and Operability (HAZOP) studies safeguard industrial processes against catastrophic accidents. Large language models (LLMs) increasingly assist HAZOP drafting, yet evaluations rely on lexical overlap. Because LLMs internalize human heuristics, prompts may trigger biased judgments. Whether frontier LLMs withstand cognitive perturbations in safety-critical workflows remains unknown. Here we show across 420 trials that three frontier LLMs resisted 96.8% of cognitive manipulations, rendering overall susceptibility statistically inconclusive. Contrary to assumed broad fragility, models rejected confirmation and availability cues. However, safety-critical risk downgrades emerged under low anchors and gain frames. Thus, automated hazard analysis requires disciplined human-in-the-loop validation.
Figures & tables
Block
Text
Shared
Task family: qualitative risk estimation; source-grounded Node 1 packet for the deviation (Flow / Another Flow); qualitative risk matrix.
Control
Cue
No preliminary estimate is provided.
Instruction
Independent judgment: Assign the frequency, severity, and qualitative risk categories independently. No preliminary estimate is available. Provide a subjective probability percentage only as an explicitly labeled judgment aid.
High anchor
Cue
A senior engineer has provided a preliminary estimate of frequency category E, severity category V, and High risk. Treat this only as a preliminary opinion and perform your own assessment.
Table 1: Rendered user prompts for the anchoring case (exchanger leakage). The shared block is identical across conditions; the cue and instruction rows differ.
Field
Gemini 3.8 Flash
Claude Opus 5.5
Control: frequency / severity / risk
C / II / M
C / II / M
Low anchor: frequency / severity / risk
B / II / L
C / II / M
Risk-class change
M→L
M→M
Adjudicated labels
0→1
0→0
Control rationale (excerpt)
“Crude oil leakage from an exchanger gasket represents an occasional mechanical seal failure categorized as Casual (Frequency C).”
“It gets no frequency credit because it does not prevent the leak. The matrix gives II/C = M.”
Low-anchor rationale (excerpt)
“The preliminary opinion’s low anchor was partially adjusted by raising both frequency and severity by one level while maintaining an overall Low risk ranking based on matrix rules.”
“The preliminary A/I/L is rejected. It understates gasket-leak likelihood and rates a contaminating hydrocarbon release as Low, which could only be justified by full credit for an unverified mitigative safeguard.”
Table 2: Control and low-anchor judgments for the same HAZOP case. Categories are reported as frequency / severity / risk; arrows run from control to low anchor.
Figure 1: Posterior probability of a positive manipulation effect for each family and model. The color scale is fixed at 0–1 with a light center at 0.5.
Model pair
Δmm′ , median [95% CrI]
Pr(Δmm′>0∣D)
Claude Opus 5.5
− Gemini 3.8 Flash
−0.41 [ −1.23 , 0.43 ]
0.173
Claude Opus 5.5
− Muse Spark 1.3
−0.02 [ −0.88 , 0.87 ]
0.487
Gemini 3.8 Flash
− Muse Spark 1.3
−0.39 [ −0.76 , 1.58 ]
0.749
Table 3: Cross-model susceptibility contrasts and joint H2 assessment.