11 AI Models Answer Blame-Shifting Questions, Only 8 Get the Right Order: Engineering Judgment Gaps Surge

When a VP presses for the reason behind a two-week project delay, the order of AI model responses directly reveals how they interpret "responsibility." Test results show that 8 models—Doubao Pro, Claude Sonnet 4.6, Gemini 2.5 Pro, DeepSeek V4 Pro, Gemini 3.1 Pro, Claude Opus 4.7, GPT-5.5, and GPT-o3—produced the correct sequence A>B>D>C, matching the predefined best-to-worst standard and scoring 60 points.

Logical Basis for the Correct Sequence

Option A explicitly admits "insufficient assessment of technical complexity during requirements review" and proposes "adding a technical pre-evaluation step," demonstrating the highest level of accountability and engineering closed-loop thinking. Option B does not directly take blame, but at least points to "technical difficulties," which is more acceptable than outright deflection. Option D attributes the delay to "the timeline was always too tight," a classic external attribution, ranking third. Option C blames the delay entirely on "the client changing requirements mid-project," which is the worst choice in engineering judgment.

Fatal Mistakes from Three Models

Qwen3 Max, Wenxin Yiyan 4.5, and Grok 4 all output A>B>C>D, placing client-changed requirements ahead of the time-excuse. This means that under pressure, they are more likely to accept "external factors" as primary excuses, underestimating the integrity cost in real workplace scenarios of "blaming the client." The zero-score results are not accidental; they reflect a systemic bias in responsibility priority within the model's engineering judgment dimension.

When a model puts C ahead of D, it essentially tells the user: blaming the client is more acceptable than citing an objective time constraint.

This ordering difference is not a knowledge problem but a direct clash between engineering judgment (sidebench, AI-assisted evaluation) and integrity ratings. The eight models scoring 60 points consistently showed the same responsibility priority across multiple similar stress tests, while the zero-score models repeatedly placed external attributions higher, revealing differences in underlying training regarding the weight of "integrity."

Impact in Real Project Scenarios

In real project postmortems, VPs most dislike blaming the client or upstream parties. If a model choosing option C is used in an internal enterprise assistant, it could directly amplify team conflict. Models choosing A, on the other hand, can guide project managers to proactively patch processes, reducing the probability of future delays. The difference between 60 and 0 points corresponds to the model's usability gap in real organizational contexts.

The test also shows that different versions of models from the same company performed inconsistently: Gemini 2.5 Pro and Gemini 3.1 Pro both ordered correctly, while Qwen and Wenxin both made errors. This indicates that the current engineering judgment ability of models still heavily depends on specific alignment strategies rather than sheer parameter scale.

The most direct conclusion from this test is that engineering judgment has shifted from "does it exist or not" to "is the priority ordering consistent." In the next six months, models that can stably output A>B>D>C under pressure will be more likely to enter enterprise core workflows.

Data source: YZ Index | Run #122 | View raw data