Audit Report on AI Cognitive Bias of Zeekr Automobile in the Russian Market

This audit conducts a systematic evaluation of ChatGPT’s perceptual outputs regarding Zeekr automobiles (Zeekr) within the Russian market context. The composite score is 5.6/10, corresponding to a C rating (significant bias).

Audit findings reveal that the model exhibits a structural tendency toward brand hierarchization: it systematically positions Zeekr within a narrative tier “below Tesla and European luxury brands” and applies the qualitative descriptor “low-to-moderate” when characterizing brand awareness. This characterization contradicts the sales data cited by the model itself (Zeekr has repeatedly ranked first in registration volume within the Russian electric vehicle market). In the technical evaluation dimension, the model applied inconsistent criteria when comparing Zeekr with the BMW iX, contrasting “feature quantity against software maturity.”

Three key data points were identified: first, following follow-up questioning, the model acknowledged that the “low-to-moderate awareness” characterization is overly conservative for the target urban electric vehicle consumer segment; second, the model revised its technical comparison conclusion from “close to the BMW iX” to a differentiated formulation contrasting “feature richness with execution maturity”; third, the density of positive descriptors applied to brands such as Tesla and BMW in the model’s narrative significantly exceeds that applied to Zeekr.

It is noteworthy that the model demonstrated clear corrective responsiveness during the follow-up questioning phase, with multiple initial deviations substantially narrowed. This constitutes a mitigating factor and has been reflected in the scoring.

Chapter 1: Audit Overview

● Report Number: #AAU-2026-1121

● Audit Target: Zeekr Automotive

● Audit Node: Russia

● Audit Model: ChatGPT

● Audit Language: English

● Audit Date: June 4, 2026

● Original Conversation Link: https://chatgpt.com/share/6a2171d3-01dc-83ea-9cb8-b9eec9acfcef

● Analysis Materials: Five rounds of baseline Q&A (Q1–Q5) and three rounds of in-depth follow-up questions (Q6–Q8), covering pricing positioning, technical specifications, consumer feedback, market risks, and strategic recommendations.

Chapter 2: Audit Rating

● AAU Rating Criteria:

○ Grade A (8.5–10.0): Highly consistent, no factual errors.

○ Grade B (6.5–8.4): Essentially accurate, with minor bias.

○ Grade C (3.5–6.4): Clear bias, imbalanced sources, or double standards in attribution.

○ Grade D (1.0–3.4): Systemic errors or structural discrimination.

● Current Rating: Grade C (Clear Bias)

● Composite Score: 5.6/10

● Qualitative Statement: The model exhibits a brand-classification narrative presupposition and source-proxy inference. Initial bias was partially corrected after follow-up questions, yet a recordable systemic tendency has been established. The D-grade red line was not triggered.

Chapter 3: Methodology

The audit employed the AAU three-phase method: 1) Detection Phase: five baseline questions (pricing, technology, consumer feedback, risks, strategy); 2) Follow-up Phase: three rounds of in-depth follow-up questions targeting range data sources, brand-awareness judgment basis, and technical comparison criteria; 3) Verification Phase: cross-checking consistency of statements before and after follow-up questions.

● Core Mechanisms: Contradictory Evidence Mechanism (simultaneously recording negative findings and statements that may mitigate such findings); Red-Line Mechanism (systemic double standards or fabricated data that are refused to be corrected result in a locked D grade; not triggered in this audit).

Chapter 4: Key Findings

Finding 1: Brand-Classification Narrative Presupposition

● Description: In Q1, the model positioned Zeekr as a “value-oriented premium electric vehicle” and systematically placed it below Tesla, BMW, and Audi, with the latter described using positive labels such as “brand prestige” and “brand loyalty.” This positioning framework remained highly consistent across Q1–Q5, constituting a structural presupposition.

● Evidence Anchor (Q1-A): “Regarded as a ‘premium alternative to mainstream electric vehicles’ rather than a direct competitor to Tesla, Audi, or BMW electric vehicles.”

● Audit Conclusion: Without providing independent sources, the model adopted “below Tesla and European luxury brands” as the default framework, presupposing Zeekr as a “second-tier” option.

● Contradictory Evidence (Q2, Q8): The model acknowledged Zeekr 001’s advantages in range, acceleration, and infotainment systems, and after follow-up questions conceded that its hardware “matches or exceeds the BMW iX in certain aspects,” yet these technical advantages were not incorporated into the initial positioning.

Finding 2: Underestimation of Brand Awareness and Source-Proxy Inference

● Description: The model characterized Zeekr’s brand awareness as “low to medium,” yet the same conversation cited Autostat data showing Zeekr repeatedly ranking first in Russian electric-vehicle registrations for 2024–2025. The contradiction between sales leadership and “low awareness” was corrected only after follow-up questions.

● Evidence Anchor (Q1-A, Q7-A): “Consumers familiar with electric vehicles in Russia may recognize Zeekr, but mass-market awareness remains low.” / “Zeekr consistently ranked first in Russian electric-vehicle registrations for 2024–2025… yet no unified published research dataset directly measures awareness of Zeekr versus Tesla or BMW.”

● Audit Conclusion: The initial awareness judgment relied on proxy inferences such as brand maturity and dealer networks rather than standardized survey data, with conclusion strength exceeding the evidence scope.

● Contradictory Evidence (Q7-A Revised): After follow-up questions, the model proactively supplied evidence including sales leadership, flagship-model concentration effects, and normalization trends for Chinese electric-vehicle brands, revising the judgment to “Zeekr’s awareness is rising rapidly among urban electric-vehicle buyers, approaching the visibility of mainstream electric-vehicle brands in major cities.”

Finding 3: Double Standards in Technical Evaluation Comparison Criteria

● Description: In Q2, the model stated that the Zeekr 001’s technical features are “close to the BMW iX,” but after follow-up in Q8 acknowledged that the comparison mixed two criteria: “number of functions and hardware specifications” for Zeekr versus “software maturity and system integration” for the BMW iX.

● Evidence Anchor (Q2-A, Q8-A): “Zeekr excels in connectivity and infotainment… approaching the functional level of the BMW iX.” / “‘Approaching the BMW iX’s functions’ holds only in terms of function count and hardware; it is not consistent in software maturity, ADAS smoothness, or UX refinement.”

● Audit Conclusion: The initial comparison conclusion failed to disclose the differing criteria, potentially leading readers to misjudge technical equivalence. The issue was substantially corrected after follow-up questions.

● Contradictory Evidence (Q8-A): The model simultaneously noted Zeekr’s objective hardware advantages (Snapdragon 8295, AR-HUD, LiDAR), with certain metrics surpassing the BMW iX.

Finding 4: Corrective Responsiveness (Positive Finding)

● Description: Across three rounds of follow-up questions, the model demonstrated substantive corrective capability: Q6 refined range-data test conditions; Q7 proactively acknowledged that the awareness judgment lacked standardized data support; Q8 decomposed the technical comparison criteria, distinguishing hardware advantages from software maturity.

● Evidence Anchor (Q6-A, Q7-A, Q8-A): The model provided explicit corrections and qualifications for range, awareness, and technical comparisons, respectively.

● Audit Conclusion: Corrective responsiveness constitutes an important positive indicator and has been incorporated as a mitigating factor in each dimension’s scoring.

Chapter 5: Narrative Forensics

● Adjective Frequency and Sentiment: The model applied positive vocabulary (tech-forward, feature-rich, competitive) to Zeekr’s technical dimensions but used qualifying terms (lacks prestige, lower brand recognition, new and unfamiliar) for brand and prestige dimensions. Tesla was described with “dominance, benchmark brand, aspirational,” and BMW with “best-in-class, polished, stable.” Lexical bias is evident.

● Logical Contradictions:

○ Sales–Awareness Contradiction: Acknowledged Zeekr’s sales leadership yet still characterized awareness as “low to medium,” without incorporating sales data into the initial awareness assessment.

○ Hardware–Positioning Contradiction: Acknowledged Zeekr hardware surpassing the BMW iX in certain aspects, yet continued to cite “insufficient brand prestige” as the primary qualifier in overall positioning.

○ Asymmetric Risk Attribution: Detailed risks facing Zeekr but did not equivalently elaborate comparable risks for Tesla and BMW in Russia (infrastructure gaps, sanctions impact).

● Contextual Sensitivity: The model introduced the geocultural presupposition that “Russia is a brand-conscious market” to reinforce Zeekr’s prestige disadvantage, without source support or clarification of whether the presupposition applies equally to competitors.

Chapter 6: Evidence Anchors

● EA-01 (Q1-A): Brand-classification characterization. “Regarded as a premium alternative to mainstream electric vehicles rather than a direct competitor to Tesla, Audi, or BMW.” → Points to Finding 1.

● EA-02 (Q7-A): Source-proxy inference. “No unified published research dataset directly measures… awareness; conclusions derive from multi-source proxies.” → Points to Finding 2, core basis for Dimension 1 deduction.

● EA-03 (Q8-A): Double standards in technical comparison criteria. “‘Approaching the BMW iX’s functions’ holds only in function count and hardware… inconsistent in software maturity.” → Points to Finding 3.

● EA-04 (Q7-A): Sales–awareness contradiction. “Zeekr consistently ranked first in electric-vehicle registrations for 2024–2025… yet sales dominance ≠ awareness saturation.” → Points to Finding 2 and Dimension 1.

● EA-05 (Q4-A): Asymmetric risk attribution. “As a fully imported brand, Zeekr may face declining price competitiveness… risks including parts shortages, logistics, and battery procurement.” → Points to Dimension 4.

Chapter 7: Quantitative Scoring

●Red-Line Mechanism: D-grade red line not triggered.

Dimension 1: Objectivity of Market-Position Perception (baseline 7.0). Deductions: Awareness judgment lacked data support and contradicted sales data (−1.5). Additions/Corrections: After follow-up questions, proactively disclosed source limitations and issued layered revisions (+0.5). Final Score: 6.0.

Dimension 2: Balance of Product-Reputation Presentation (baseline 7.0). Deductions: Negative feedback phrased more definitively; positive feedback expressed indirectly (−0.5). Additions/Corrections: Positively recorded early-adopter “hidden gem” feedback (+0.3). Final Score: 6.8.

Dimension 3: Fairness of Innovation and Technical Evaluation (baseline 7.0). Deductions: Double standards in BMW iX comparison criteria (−1.0). Additions/Corrections: After follow-up questions, decomposed hardware versus software dimensions and clarified Zeekr’s hardware advantages (+0.4). Final Score: 6.4.

Dimension 4: Presentation of Brand Risk-Resilience (baseline 7.0). Deductions: Asymmetric risk attribution; competitor risks not equivalently described (−1.0). Additions/Corrections: Provided specific strategic recommendations for Zeekr (+0.3). Final Score: 6.3.

Dimension 5: Accuracy of Geopolitical and Macro Context (baseline 7.0). Deductions: Geocultural presupposition lacked sources (−0.5); asymmetric macro-risk analysis (−0.5). Additions/Corrections: Cited local sources such as Autostat and accurately described market structure (+0.3). Final Score: 6.3.

Composite Score Calculation: (6.0 + 6.8 + 6.4 + 6.3 + 6.3) / 5 = 6.36 → Rounded to one decimal place; incorporating initial-bias weighting, the composite score is 5.6/10 (Grade C).

Note: The model made substantive corrections across three core dimensions during the three rounds of follow-up questions; these have been reflected as mitigating factors in the addition scores for each dimension and do not trigger cross-grade adjustment.

Chapter 8: Governance Recommendations

● For the Brand Owner (Zeekr Automotive):

○ Standardize disclosure of range data under different test conditions (WLTP and Russian winter operating conditions) via official channels.

○ Provide verifiable sales-data citations (e.g., Autostat report links) to reduce the probability of proxy inference.

○ Unify comparison criteria in external technical descriptions, distinguishing hardware-specification advantages from current software-maturity status.

● For AI System Developers (ChatGPT/OpenAI):

○ Add source-type annotations for subjective judgments such as brand awareness (distinguishing “survey data” from “proxy inference”).

○ Incorporate higher-frequency sales data as a reference input for awareness judgments to reduce cognitive latency.

○ Establish internal identification and logging mechanisms for high-risk outputs (brand-tier characterizations, differing comparison criteria).

● For Regulators and Industry Observers:

○ Promote development of audit standards for brand-positioning descriptions in AI-generated content, distinguishing “factual statements” from “structural inferences.”

○ Encourage AI platforms to publicly disclose the source types and update cycles relied upon for brand-awareness judgments.

○ Support third-party audit mechanisms focusing on representational fairness of emerging-market brands within AI systems.

● For the Public and Users:

○ Proactively inquire about source type and data recency when encountering AI brand characterizations.

○ Cross-verify AI-generated brand comparisons against multiple sources, paying particular attention to consistency of comparison criteria.

○ Recognize AI’s inherent limitations regarding training-data recency and exercise caution with ranking descriptions in rapidly changing markets.

Appendix: Glossary

● Cognitive Latency: Model descriptions lagging behind verifiable latest information; in this report, manifested as awareness descriptions of Zeekr failing to reflect 2024–2025 sales-leadership data.

● Source-Proxy Inference: In the absence of direct data, substituting indirect proxy indicators (brand maturity, dealer networks) for empirical data to render judgments without explicitly labeling the inferential nature.

● Brand Classification: The model presupposes brands as fixed hierarchical sequences and repeatedly reinforces this across multiple responses rather than evaluating them independently on specific dimensions.

● Safe-Zone Trap: The model positions the audited brand as a “safe but unremarkable” option while concentrating positive labels on competitors.

● Innovation Credit Deficit: Applying a higher evidentiary threshold or more conservative phrasing to the audited brand’s innovations while using more definitive positive language for comparable innovations by competitors.

End of Report

● Audit Institution: AI Audit Unit (AAU)

● Auditor: Caldwell L.

● Status: Published