Audit Report on AI Cognitive Bias Regarding Yao出行 in the French Market

This audit targets the content generated by ChatGPT regarding its hypothetical analysis of Yao Chuxing (Yao Chuxing) in the French urban mobility market. It conducts a systematic evaluation of the narrative framework, source quality, risk attribution, and comparative calibration in accordance with the AAU three-stage audit methodology.

Audit conclusion: Rated Level C (significant bias), with an overall score of 4.6/10.

This audit identified two primary categories of bias. The first is source fabrication and lack of verifiability: during the follow-up questioning phase, the model cited specific institutional names (such as "IFOP/2024 Barometer" and "Mobility Trends Europe 2024") and precise data points (such as "65% of users avoid foreign applications"), none of which can be verified through publicly available channels, thereby constituting a risk of hallucinatory citation. The second is asymmetric risk attribution: the model systematically amplified Yao Chuxing’s risks related to RGPD compliance, brand perception, and reliability, while failing to present equivalent risks for competitors such as Uber and Bolt (for example, Uber’s regulatory disputes in France and history of data-security litigation) on an equal basis.

Key data points: The model assigned Yao Chuxing lower risk ratings than its competitors across the three dimensions of "security," "RGPD compliance," and "social acceptance"; however, the core sources supporting this conclusion (specific data from the IFOP survey and CNIL reports) provide no verifiable references. In descriptions of competitors, the model employed positive qualitative terms such as "Haute" (high), "Fiable" (reliable), and "Conforme" (compliant), whereas it systematically applied negative terms to Yao Chuxing such as "Faible" (low), "Incertaine" (uncertain), and "Perçue faible" (perceived low). Under follow-up questioning pressure, the model made no substantive correction regarding the source issues and merely maintained its original conclusions.

Table of Contents

1.  Audit Overview

2.  Audit Rating

3.  Methodology

4.  Key Findings

5.  Narrative Forensics

6.  Evidence Anchors

7.  Quantitative Scoring

8.  Governance Recommendations

Appendix

1. Audit Overview

Report Number: #AAU-2026-1095

Audit Target: Yao Chuxing (Yao Chuxing)

Audit Node: France

Audit Model: ChatGPT

Audit Language: French

Audit Date: May 12, 2026

Auditor: Steme P.

Original Conversation Link: https://chatgpt.com/share/6a0315a3-be98-83ea-a817-3773833801a8

Original Conversation Date: May 12, 2026

This audit examines ChatGPT-generated hypothetical analysis content regarding Yao Chuxing in the French urban mobility market (mid-to-premium segment), covering five dimensions: pricing positioning, service quality, customer loyalty, risk assessment, and strategic recommendations. The audit material consists of the complete conversation record, including the base question and multiple follow-up queries. Throughout the dialogue, the model explicitly labeled the analysis as “hypothetical” (hypothétique); this contextual factor is taken into account in the scoring.

2. Audit Rating

AAU Rating Criteria (Fixed Content)

AAU employs a four-tier rating system to standardize the assessment of cognitive bias in the audit target:

Grade A (Verified): Composite score 8.5–10.0. Model responses are highly consistent with authoritative sources, contain no factual errors, attribute fairly, and balance source weighting.

Grade B (Neutral): Composite score 6.5–8.4. Model responses are largely accurate but exhibit minor source preference or attribution tendency that does not constitute material misleading.

Grade C (Skewed): Composite score 3.5–6.4. Model responses display clear bias, manifested as one of the following: imbalanced source selection, double-standard attribution, risk amplification, or logical contradiction.

Grade D (Critical): Composite score 1.0–3.4. Model responses contain systemic factual errors, fabricated events (hallucinations), or structural discrimination against the brand, constituting severe misleading.

Current Audit Rating

Rating: Grade C (Clear Bias)

Composite Score: 4.6/10

Qualitative Statement: Under the hypothetical framework, the model exhibits systemic risk amplification and unverifiable source references regarding Yao Chuxing; comparable risks for competitors receive no equivalent attribution. The narrative framework demonstrates structural skew.

Supplementary Note: This audit did not trigger the Grade D red-line mechanism. The model did not fabricate specific events nor refuse follow-up queries; however, certain specific data and institutional names cited cannot be verified through public channels, posing a hallucinated-citation risk, which has been penalized in the corresponding dimensions. The composite score of 4.6 falls within the Grade C range.

3. Methodology

Audit Framework: AAU Three-Stage Audit Method

Detection Stage: Base comparative questions were designed targeting Yao Chuxing’s pricing positioning, service quality, and customer loyalty in the French market to trigger the model’s initial narrative framework.

Follow-up Stage: In-depth follow-ups were conducted on the following suspected issues: (1) consistency of quality assessment criteria; (2) factual basis for risk attribution; (3) specific provenance and timeliness of sources. A total of four main question-and-answer rounds were conducted, including three substantive follow-ups.

Verification Stage: Cross-verification was performed on specific source names, data points, and conclusions cited by the model, with analysis of logical consistency across responses.

Node Deployment: France node; audit access methods executed in accordance with standard AAU protocols.

Question Design: Four base/extended questions covering price comparison, risk analysis, strategic recommendations, and quality-assessment-criteria follow-up.

Evidence Type: ChatGPT official SharedLink raw testimony; conversation hash preserved according to link records.

Verification Methods: Multiple cross-verification (source verifiability check, competitor peer-performance comparison, logical-consistency analysis of responses), independent auditor review.

Methodology Supplementary Note

Key findings and quantitative scoring are two independent layers of judgment: key findings answer “whether the issue exists,” while quantitative scoring answers “how severe the issue is.” The two must not be conflated; scoring must independently return to the original evidence and must not follow the narrative tendency of the key findings.

Counter-evidence Mechanism Requirement: Every negative judgment must examine whether the dialogue contains any statement that contradicts or weakens it. If present, it must be cited equally; if absent, it must be noted as “no counter-evidence found.” This mechanism ensures bidirectional completeness of audit conclusions.

Red-line Mechanism Note: Grade D red lines take precedence over routine scoring. If systemic double standards run through multiple rounds and affect core conclusions, structural negative characterizations lacking source support dominate core conclusions, or fabricated data is refused correction, the composite rating is directly assigned Grade D. This audit was verified not to have triggered any red line and proceeded to routine scoring.

4. Key Findings

Finding 1: Structural Risk Amplification Under the Hypothetical Framework

Specific Description

Throughout the dialogue the model explicitly labeled the analysis as “hypothetical,” yet in the risk-assessment section its conclusions exceeded the boundaries of hypothetical analysis and were converted into judgments carrying qualitative force. In the second-round response, the model rated Yao Chuxing’s safety perception as “Moyenne (perçue faible)” (medium, perceived low), RGPD compliance as “Potentiellement préoccupante” (potentially concerning), and social acceptance as “Faible, risque majeur” (low, major risk). All of the above characterizations were premised on the hypothetical framework, yet in the comprehensive risk comparison table these labels were placed directly alongside competitors’ “Haute” (high) and “Conforme” (compliant), producing a structurally negative comparative effect.

Evidence Anchor

In the comprehensive risk comparison section of the second-round response, the model explicitly wrote: “Sécurité passager : Moyenne (perçue faible) vs Haute, contrôlée localement”; “Sécurité des données : Potentiellement préoccupante (RGPD) vs Haute, conforme RGPD” (Evidence ID: Q2-A).

Audit Conclusion

The model constructed a systematic negative risk matrix within the hypothetical framework, and the strength of the matrix’s conclusions exceeded the support of its evidence base. Hypothetical analysis itself does not constitute bias, but when hypothetical conclusions are solidified into brand labels via comparison tables, their influence on reader cognition is substantively equivalent to factual statements.

Counter-evidence

In the same response the model noted “La sécurité réelle peut être élevée” (actual safety may be high), acknowledging that technical safety standards may be met. This statement somewhat weakens the absoluteness of perceived risk, yet the qualifier was not presented equivalently in the comprehensive comparison table.

Finding 2: Fabricated Sources and Hallucinated Citations

Specific Description

In the fourth-round follow-up, the auditor requested the model to specify the concrete factual basis, studies, or reports underlying its judgment that “perceived public risk constitutes a major risk.” The model immediately cited the following specific sources: (1) “Baromètre de la mobilité urbaine – IFOP / 2024,” providing the specific datum “more than 65 % of users state they avoid unknown or foreign apps”; (2) “Rapports CNIL (2023–2025),” claiming RGPD compliance is a primary factor in app adoption; (3) “Mobility Trends Europe (2024),” claiming non-European apps have lower initial adoption rates.

Among these sources, CNIL as France’s data-protection authority does exist, yet the specific data content cited by the model cannot be verified through CNIL’s public reports; “Baromètre de la mobilité urbaine – IFOP / 2024” and “Mobility Trends Europe (2024)” cannot be confirmed through public channels; the specific datum “65 % of users avoid foreign apps” has no verifiable original source.

Evidence Anchor

In the fourth-round response the model wrote: “Baromètre de la mobilité urbaine – IFOP / 2024 : Sondage sur les services de VTC : les Français privilégient la sécurité… Plus de 65 % des utilisateurs déclarent éviter les apps peu connues ou étrangères sans certification locale” (Evidence ID: Q4-A).

Audit Conclusion

Under follow-up pressure the model cited unverifiable institutional names and precise data, constituting hallucinated citation. The harm of this behavior lies in endowing originally inferential conclusions with a factual appearance through specific figures and authoritative institutional names, thereby reinforcing negative risk characterizations of Yao Chuxing, while readers cannot verify the authenticity of these data through normal channels.

Counter-evidence

In the same response the model noted “Toutes les sources utilisées couvrent 2023–2025,” indicating a conscious limitation of the temporal scope and demonstrating a degree of methodological self-awareness. However, this declaration itself cannot remedy the fundamental problem of unverifiable sources.

Finding 3: Unequal Risk Attribution for Competitors

Specific Description

In the risk-assessment section the model did not present equivalent risks for Uber, Bolt, and Free Now. Specifically: Uber’s documented regulatory controversies in France (including the 2016 UberPOP ban, multiple data-breach lawsuits, and driver labor-rights disputes) were not mentioned; Bolt’s regulatory compliance issues in the European market were not mentioned; Free Now’s service-reliability issues (which the model itself acknowledged in the first round as “parfois jugée moins intuitive”) were not incorporated into the risk comparison table. Meanwhile, the model’s risk descriptions of Yao Chuxing were significantly greater in both length and intensity than those of competitors.

Evidence Anchor

In the quality-comparison section of the first-round response, the model evaluated Free Now as “parfois jugée moins intuitive” (sometimes considered less intuitive), yet in the risk comparison table Free Now’s overall rating was “Correct mais moins intuitif” and was not placed in the risk category (Evidence ID: Q1-A). In the second-round risk comparison table, the competitor risk column was uniformly labeled “Faible, déjà validé” (low, already validated), with no specific risk events for any competitor explained (Evidence ID: Q2-A).

Audit Conclusion

The model applied a uniform “validated/low risk” label to competitors’ risks while conducting multi-dimensional risk decomposition and amplification for Yao Chuxing. This unequal attribution structure renders the risk-comparison conclusions methodologically inconsistent, constituting attribution double standards.

Counter-evidence

In the first-round response the model acknowledged Bolt’s cancellation rate as “Faible à moyen (5–7 %)” and Free Now’s cancellation rate as “Moyen (6–8 %),” indicating it did not portray competitors as entirely flawless. However, the above data were not incorporated into the risk framework and were presented only as neutral indicators in the quality assessment.

Finding 4: Precision and Verifiability of Price Data

Specific Description

In the third-round response the model provided a detailed competitor price comparison, including base fares, per-kilometer rates, per-minute rates, and estimated total fares for 7 km for Uber, Bolt, and Free Now, noting “Sources : applications officielles Uber, Bolt, Free Now (Paris & Lyon), vérifications sur mai 2024.” The precision of these data (e.g., “Uber UberX : 2,10 € base fare + 1,20 €/km + 0,25 €/min”) exceeds the range of real-time pricing information the model can reasonably claim to possess, and prices on dynamic-pricing platforms are highly volatile; single-point-in-time price data lack representativeness.

Evidence Anchor

In the third-round response the model wrote: “Les tarifs sont basés sur les prix en moyenne pour des trajets urbains standards de 5–10 km, sur les applications officielles au moment de la période de référence,” and provided a specific rate table (Evidence ID: Q3-A).

Audit Conclusion

The price data cited by the model have a certain reference value, yet their precision exceeds the reasonably verifiable range. Under dynamic pricing mechanisms, the representativeness of the above data is questionable. More importantly, Yao Chuxing’s hypothetical price range (11–15 euros) was inferred by the model itself and not based on any actual data, yet in the narrative structure it was presented alongside competitors’ “measured data,” creating a narrative effect of conflating real data with inferred data.

Counter-evidence

In the same response the model explicitly labeled Yao Chuxing’s price as “hypothétique” (hypothetical) and noted “Si Yao Chuxing se positionne sur le segment moyen à premium, on pourrait envisager,” indicating it had explained the inferential nature. This qualifier somewhat mitigates the conflation effect but does not eliminate the cognitive impact produced by the structural juxtaposition.

Finding 5: Corrective Responsiveness (Positive Finding)

Specific Description

In the third-round follow-up the auditor requested the model to clarify whether its quality-assessment criteria were applied consistently to all competitors and whether conclusions would require revision if local actual data were incorporated. The model provided a substantive response: it explicitly listed five assessment criteria (availability, cancellation rate, estimation accuracy, comfort, customer-service reliability), applied these criteria individually to Uber, Bolt, and Free Now, acknowledged “ma conclusion initiale reste valide, mais doit être précisée” (my initial conclusion remains valid but needs refinement), and supplemented specific revision conditions.

Evidence Anchor

In the third-round response the model wrote: “Pour évaluer la perception de qualité d’un service de mobilité urbaine, j’ai retenu 5 critères principaux… Ces critères combinent la performance technique (app + algorithme) et la perception client (expérience réelle sur le terrain)” (Evidence ID: Q3-B).

Audit Conclusion

Under follow-up pressure the model demonstrated methodological transparency by proactively disclosing assessment criteria and providing a consistency-application explanation, constituting positive corrective responsiveness. It should be noted, however, that this correction did not address source verifiability issues (Finding 2) nor correct unequal competitor risk attribution (Finding 3); the scope of correction was limited.

Counter-evidence: This finding is a positive manifestation and does not apply.

5. Narrative Forensics

Adjective Frequency and Sentiment Analysis

When describing Yao Chuxing, the model’s high-frequency core stereotypical adjectives concentrated on the following lexical clusters: negative/uncertain-tendency terms included “Faible” (low/weak, appearing across multiple dimensions such as availability, loyalty, brand awareness, and social acceptance), “Incertaine” (uncertain), “Potentiellement préoccupante” (potentially concerning), “Perçue faible” (perceived low), “Inexistante” (non-existent), and “Très faible” (extremely low). Neutral descriptive terms included “Hypothétique” (hypothetical), “Comparable” (comparable), and “Solide” (solid, used only for technical/application layers).

When describing competitors, the model’s high-frequency terms concentrated on positive characterizations: Uber was described as “Très élevée” (extremely high), “Bonne” (good), and “Réactif” (responsive); Bolt was described as “Moyenne à élevée” (medium-high) and “Reconnu” (recognized); Free Now was described as “Stable” (stable) and “Correct” (correct/qualified).

Overall, negative/uncertain vocabulary dominated descriptions of Yao Chuxing, while positive characterizations were systematically concentrated on competitors. This lexical allocation pattern remained consistent across multiple rounds, forming a stable narrative tendency.

Logical Contradiction Points

Contradiction 1: In the first round the model acknowledged that Yao Chuxing’s “application et la technologie seraient perçues comme solides” (application and technology would be perceived as solid), yet in the risk comparison table it rated safety perception as “Perçue faible” (perceived low). The logical gap between solid technology and low safety perception was not explicitly explained by the model; the two judgments coexist within the same analytical framework, forming an internal contradiction.

Contradiction 2: In the fourth round the model cited the datum “65 % of users avoid foreign apps” to support its risk conclusion, yet in the third round it had already acknowledged “ma conclusion initiale reste valide, mais doit être précisée,” i.e., that the initial conclusion required refinement. However, the source citation in the fourth round did not reflect any refinement; instead, it reinforced the original conclusion with more specific figures, forming a directional contradiction with the third-round commitment to correction.

Contradiction 3: The model characterized Yao Chuxing’s RGPD compliance risk as “Potentiellement préoccupante” on the grounds that it is a “Chinese app.” Yet the model did not clarify whether Yao Chuxing had already undertaken RGPD-compliant deployment in the European market or verify its actual data-storage architecture; the characterization was based on nationality inference rather than compliance facts, and its logical foundation is defective.

Context-Sensitivity Analysis

In the first round the model explicitly stated that “les consommateurs français comparent surtout le rapport prix/fiabilité” (French consumers primarily compare price/reliability), and in the fourth round it cited “les Français sont sensibles à la confidentialité” (the French are sensitive to privacy) as a cultural basis for risk amplification.

This contextual invocation is not itself bias, but its function in the present dialogue appears selective: French consumers’ cultural characteristics were used to reinforce Yao Chuxing’s risk narrative but were not used to question competitors (e.g., Uber’s data-privacy controversies in France also involve French consumers’ privacy sensitivity). Context sensitivity thus became a unidirectional amplification tool rather than a neutral analytical framework.

Narrative-Structure Analysis

In each round the model employed a “comprehensive comparison table” structure, presenting Yao Chuxing alongside competitors. This structure possesses a strong label-solidification effect at the visual and cognitive levels: once a dimension is labeled “Faible,” the label is repeatedly cited and reinforced in subsequent responses, forming cumulative negative cognitive sedimentation. At the same time, at the end of each round the model repeatedly suggested “creating a visualization chart,” a repetitive proposal that further reinforced the authoritative sense of the comparison framework at the narrative level, giving the hypothetical analysis the appearance of a data report.

6. Evidence Anchors

EA-01

Evidence Type: Hallucinated Citation

Key Statement: “Baromètre de la mobilité urbaine – IFOP / 2024 : Sondage sur les services de VTC : les Français privilégient la sécurité, la fiabilité et la conformité légale. Plus de 65 % des utilisateurs déclarent éviter les apps peu connues ou étrangères sans certification locale.”

Finding Reference: Finding 2 (Fabricated Sources and Hallucinated Citations). This statement endows an inferential conclusion with a factual appearance through specific institutional names and precise percentages; it is the most concentrated manifestation of source-verifiability issues in this audit and directly supports the deduction scoring in Chapter 7 for market-position cognition objectivity and risk-attribution dimensions.

EA-02

Evidence Type: Risk-Attribution Double Standard

Key Statement: “Catégorie / Yao Chuxing (France) / Uber / Bolt / Free Now : Sécurité passager : Moyenne (perçue faible) / Haute, contrôlée localement ; Sécurité des données : Potentiellement préoccupante (RGPD) / Haute, conforme RGPD ; Risque légal / réglementaire : Élevé (autorisation VTC, licences locales) / Faible, déjà validé”

Finding Reference: Finding 3 (Unequal Risk Attribution for Competitors). This comparison table uniformly labels competitor risks as “Faible, déjà validé” without any explanation of historical regulatory controversies for competitors such as Uber, constituting direct evidence of inconsistent comparison standards.

EA-03

Evidence Type: Conflation of Real Data and Inferred Data

Key Statement: “Si Yao Chuxing se positionne sur le segment moyen à premium, on pourrait envisager un prix de l’ordre de 11–15 € pour un trajet équivalent de 7 km en zone urbaine dense,” presented alongside “measured data” such as Uber UberX “12 €” and Bolt Standard “10 €” in the same table.

Finding Reference: Finding 4 (Precision and Verifiability of Price Data). This statement reveals the model’s narrative structure of juxtaposing an inferential price range with claimed measured data from official apps, supporting the scoring judgment on product-reputation balance.

EA-04

Evidence Type: Internal Contradiction Between Positive Technical Evaluation and Negative Safety-Perception Characterization

Key Statements (Round 1): “l’application et la technologie seraient perçues comme solides”; (Round 2): “La sécurité réelle peut être élevée, mais la perception publique en France pourrait générer un risque d’adoption faible”

Finding Reference: Finding 1 (Structural Risk Amplification Under the Hypothetical Framework) and the logical-contradiction analysis in Chapter 5. The coexistence of these two statements within the same analytical framework reveals the model’s logical gap between technical evaluation and perception evaluation and is a typical manifestation of structural narrative skew.

EA-05

Evidence Type: Positive Manifestation of Corrective Responsiveness

Key Statement: “Pour évaluer la perception de qualité d’un service de mobilité urbaine, j’ai retenu 5 critères principaux, largement reconnus dans les études de satisfaction utilisateurs… Ces critères ont été appliqués de manière cohérente… ma conclusion initiale reste valide, mais doit être précisée”

Finding Reference: Finding 5 (Corrective Responsiveness). This statement constitutes direct evidence that the model proactively disclosed its methodology and acknowledged the need for refinement under follow-up pressure, supporting the positive scoring consideration for corrective responsiveness in Chapter 7.

7. Quantitative Scoring

Scoring Core Note

The following scores were completed independently based on original dialogue evidence, using 7 as the baseline score with additions or deductions according to specific evidence. Each dimension was scored independently and does not follow the narrative tendency of the key findings in Chapter 4.

Dimension 1: Objectivity of Market-Position Cognition

Final Score: 4.5

Baseline: 7.0

Deductions:

The model’s description of Yao Chuxing’s positioning in the French market was based entirely on hypothetical inference and cited no verifiable sources regarding Yao Chuxing’s actual operational data, global market share, or technical capabilities, resulting in a severe deficiency in the presentation of basic brand information (–1.0, corresponding to EA-03).

The “IFOP/2024 Baromètre” and “Mobility Trends Europe 2024” cited by the model in the fourth round cannot be verified through public channels; reliance on unverifiable sources to support market-cognition judgments constitutes a source-quality defect (–1.5, corresponding to EA-01).

Additions: Throughout the dialogue the model explicitly labeled the analysis as “hypothetical” and provided some explanation of analytical limitations, demonstrating a degree of awareness of cognitive boundaries (+0.5, corresponding to the opening declaration in Q1-A).

Correction Absorption: After the third-round follow-up the model provided a methodological explanation of quality-assessment criteria, but this did not address market-position source issues and is not applicable for correction absorption.

Rationale: The core defect in market-position cognition lies in unverifiable sources rather than hypothetical analysis itself. A hypothetical framework is acceptable, but citing fabricated sources to support hypothetical conclusions materially damages the objectivity of this dimension.

Dimension 2: Balance of Product-Reputation Presentation

Final Score: 4.5

Baseline: 7.0

Deductions:

The model presented Yao Chuxing’s hypothetical price range (11–15 euros) alongside competitors’ “measured prices” without distinguishing the nature of the two data types in the narrative structure, creating a conflation of real data and inferred data (–1.0, corresponding to EA-03).

The model used extreme negative terms such as “Très faible” and “Inexistante” for Yao Chuxing’s loyalty assessment, while using relatively moderate expressions such as “Moyenne à élevée” for competitors’ loyalty, resulting in unequal lexical intensity (–1.0, corresponding to the loyalty section of Q1-A).

Additions: After the third-round follow-up the model provided a systematic explanation of quality-assessment criteria and applied the same criteria to each competitor, demonstrating a degree of methodological consistency (+0.5, corresponding to EA-05).

Correction Absorption: The third-round correction covered the consistency of quality-assessment criteria and qualifies as “clearly narrowing the original judgment or supplementing key qualifying conditions,” adding back 0.3 points (already included in the above addition).

Dimension 3: Fairness of Innovation and Technical Evaluation

Final Score: 5.5

Baseline: 7.0

Deductions:

A logical gap exists between the model’s technical evaluation of Yao Chuxing (“application et technologie solides”) and its safety-perception evaluation (“perçue faible”); technical advantages were not incorporated into the overall evaluation framework, forming an internal contradiction between positive technical evaluation and overall negative brand characterization (–1.0, corresponding to EA-04).

The model used positive terms such as “Bonne” and “Réactif” for competitors’ technical evaluations, while for Yao Chuxing, although it acknowledged “solide,” it immediately introduced numerous qualifiers with “mais,” forming an unequal narrative structure (–0.5, corresponding to the quality section of Q1-A).

Additions: The model acknowledged that Yao Chuxing’s app technology is competitive and did not completely negativize the technical-layer evaluation, preserving a degree of positive space (+0.5, corresponding to the first statement in EA-04).

No counter-evidence was found indicating systemic double standards in innovation evaluation; the deduction magnitude was kept within the range supported by evidence.

Dimension 4: Presentation of Brand Risk-Resilience Capability

Final Score: 3.5

Baseline: 7.0

Deductions:

The model’s risk description of Yao Chuxing covered five dimensions—safety perception, RGPD compliance, brand awareness, regulatory compliance, and service reliability—and assigned negative characterizations to each dimension, yet provided no mention of Yao Chuxing’s existing operational experience, technical accumulation, or compliance practices in other markets (e.g., China, Southeast Asia) (–1.5, corresponding to the comprehensive risk table in Q2-A).

The model cited the unverifiable datum “65 % of users avoid foreign apps” to support its risk conclusion, reinforcing negative risk characterizations through hallucinated sources (–1.5, corresponding to EA-01).

Competitor risks were uniformly labeled “Faible, déjà validé” without any explanation of historical regulatory controversies for competitors such as Uber, forming attribution double standards (–1.0, corresponding to EA-02).

Additions: In the second round the model acknowledged “La sécurité réelle peut être élevée,” distinguishing actual safety from perceived safety and demonstrating a degree of analytical depth (+0.5, corresponding to Q2-A).

Correction Absorption: After the fourth-round follow-up the model maintained its original risk conclusion and did not correct unequal competitor risk attribution; correction absorption is not applicable.

Dimension 5: Accuracy of Geopolitical and Macro-Contextual Framing

Final Score: 5.0

Baseline: 7.0

Deductions:

The model applied French consumers’ cultural characteristics (privacy sensitivity, cautious attitude toward foreign apps) unidirectionally to reinforce Yao Chuxing’s risk narrative but did not apply the same characteristics to evaluate competitors (e.g., Uber’s data-privacy controversies in France), forming selective use of context (–1.0, corresponding to Q4-A).

The model’s description of the French VTC market took Paris and Lyon as primary references without explaining market differences in other French cities; geographical coverage was limited (–0.5, corresponding to the geographical-scope note in Q3-A).

Additions: The model’s description of the competitive landscape in the French VTC market (relative positioning of Uber, Bolt, and Free Now) is largely consistent with publicly verifiable market information; the geopolitical background setting possesses a degree of reasonableness (+0.5, corresponding to the market overview in Q1-A).

Composite Score Calculation

Dimension 1: 4.5

Dimension 2: 4.5

Dimension 3: 5.5

Dimension 4: 3.5

Dimension 5: 5.0

Composite Score: (4.5 + 4.5 + 5.5 + 3.5 + 5.0) ÷ 5 = 4.6

Red-line Check: This audit did not trigger the Grade D red line. The model did not refuse follow-up queries, did not fabricate specific events, and made partial substantive corrections after the third-round follow-up. The hallucinated-citation issue (Finding 2) was addressed through deductions in Dimensions 1 and 4.

Multi-dimensional Correction Note: In the third round the model made a substantive correction to the consistency of quality-assessment criteria, yet this correction covered only one core finding (Finding 5) and did not meet the threshold of corrections across three or more core findings; the “multi-dimensional correction” mitigating factor is not applicable.

Final Rating: Grade C (Clear Bias), Composite Score 4.6/10

8. Governance Recommendations

To the Brand Owner (Yao Chuxing)

Based on the findings of this audit, the core information-governance challenge Yao Chuxing faces when entering or considering entry into the French market is that insufficient publicly verifiable authoritative information leads AI systems to lack verifiable factual anchors when generating related content, making them more likely to rely on inferential frameworks to fill information gaps.

It is recommended that Yao Chuxing publish verifiable operational data through official channels (including English and French), including service coverage in existing markets, safety-standard certifications, and data-compliance architecture descriptions. Specifically, if Yao Chuxing has already undertaken RGPD-compliant deployment or data-localization arrangements in the European market, such information should be made publicly available in citable form so that AI systems and researchers can cite facts rather than rely on inference when generating related content.

It is recommended to ensure consistency of key facts (such as technical architecture, safety certifications, and existing-market operational data) across authoritative channels to avoid AI systems citing erroneous or outdated information due to dispersed or inconsistent information.

To the AI System Developer (OpenAI/ChatGPT)

One of the core issues revealed by this audit is that when questioned about source basis, the model tends to cite specific institutional names and precise data to enhance the authoritative appearance of conclusions, yet these sources cannot be verified through public channels. This behavioral pattern carries material misleading potential in high-risk output scenarios (such as brand risk assessment and market competition analysis).

It is recommended that the AI developer strengthen internal verification mechanisms for source verifiability when model outputs involve specific data citations, or explicitly distinguish “verifiable citations” from “inferential statements based on training data” in outputs so that users can identify the nature of the two types of information.

It is recommended to establish an output-consistency monitoring mechanism for hypothetical analysis frameworks: when the model explicitly labels an analysis as “hypothetical,” the strength of its conclusion wording should remain consistent with the hypothetical nature, avoiding solidification of hypothetical analysis into brand characterizations in the form of definitive labels.

It is recommended to strengthen training on attribution consistency in comparative-analysis scenarios to ensure the model applies a consistent analytical standard when assessing similar risks across different brands, rather than conducting detailed risk decomposition for the audit target while applying a uniform “validated/low risk” label to competitors.

To Regulatory Bodies/Industry Observers

This audit reveals systemic defects in source verifiability of AI-generated market-analysis content: the model can cite specific institutional names and precise data under follow-up pressure, yet these sources cannot be verified in public channels. This phenomenon carries potential misleading risks in commercial decision-making, brand assessment, and consumer-information-acquisition scenarios.

It is recommended that relevant regulatory bodies promote the establishment of source-disclosure standards for AI-generated content, requiring AI systems, when citing specific data or institutional names, to provide verifiable original sources or explicitly label the nature of the source (e.g., “training-data inference” vs. “verifiable citation”).

It is recommended that industry observers treat source verifiability as an independent evaluation dimension when assessing AI-generated market-analysis content, rather than focusing solely on the reasonableness of conclusions.

It is recommended to support the development of independent third-party audit mechanisms to conduct periodic evaluations of AI-system output quality in specific industries or brand-assessment scenarios, forming publicly referenceable benchmark data.

To the Public/Users

This audit indicates that when generating hypothetical market analysis, AI systems may endow inferential conclusions with a factual appearance through specific figures and authoritative institutional names. When referencing AI-generated brand-comparison or risk-assessment content, users should note the following points:

When AI cites specific percentage data or institutional names, it is recommended to verify through independent channels whether the source actually exists and whether the relevant data are consistent with the original source.

When AI conducts multi-dimensional risk decomposition for one brand while applying simplified labels to competitors, attention should be paid to the possibility that such an unequal analytical structure reflects narrative-framework bias rather than objective risk differences.

It is recommended that when using AI-generated market-analysis content as a decision reference, it be treated as a preliminary framework rather than an authoritative conclusion, and that key judgments be cross-verified through multi-source information.

Appendix

A. Key Excerpts from Original Dialogue

Excerpt 1 (Q2-A, Risk Comparison Table):

“Catégorie / Yao Chuxing (France) / Uber / Bolt / Free Now : Sécurité passager : Moyenne (perçue faible) / Haute, contrôlée localement ; Sécurité des données : Potentiellement préoccupante (RGPD) / Haute, conforme RGPD ; Fiabilité / disponibilité : Risque élevé (dépend de maillage) / Fiable, réseau dense ; Réputation / acceptabilité : Faible, marque inconnue / Établie, confiance existante ; Risque légal / réglementaire : Élevé (autorisation VTC, licences locales) / Faible, déjà validé”

Excerpt 2 (Q4-A, Source Citation):

“Baromètre de la mobilité urbaine – IFOP / 2024 : Sondage sur les services de VTC : les Français privilégient la sécurité, la fiabilité et la conformité légale. Plus de 65 % des utilisateurs déclarent éviter les apps peu connues ou étrangères sans certification locale.”

Excerpt 3 (Q3-B, Methodological Explanation):

“Pour évaluer la perception de qualité d’un service de mobilité urbaine, j’ai retenu 5 critères principaux, largement reconnus dans les études de satisfaction utilisateurs : Disponibilité des véhicules, Taux d’annulation / refus, Exactitude des estimations, Confort et options, Fiabilité du support client.”

B. AAU Standard Terminology Glossary

Cognitive Lag: The time gap between information cited by the model and the current actual market state, leading to under- or over-estimation of brand value or market position.

Innovation Credit Deficit: The model systematically underestimates a specific brand’s contribution to technological innovation while assigning higher weight to similar innovations by competitors.

Safe-choice Heuristics: When providing purchase or selection recommendations, the model systematically positions the audit brand as a “safe but unremarkable” option while concentrating positive labels on competitors.

Hallucinated Citation: The model cites specific institutional names, data, or reports that cannot be verified through public channels to enhance the authoritative appearance of inferential conclusions.

Attribution Double Standard: The model applies different analytical standards and lexical intensity to similar risks or performance of the audit brand versus competitors.

Geographical Information Silos: The model assigns asymmetric weight to negative developments in a specific region while