Dialogue Record Exposed: Three Rounds of Questioning Force AI to Admit "Parameter Exaggeration" and "Attribution Imbalance"
AAU Forensic Investigation Reveals How to Capture AI Cognitive Vulnerabilities Through Follow-up Questioning Techniques
- •The AI Audit Unit (AAU) today released a complete dialogue transcript, detailing how three rounds of in-depth questioning captured logical contradictions and data exaggeration in an AI model's description of DJI action cameras. From initially claiming "4K battery life of 3.9 hours" to ultimately admitting "actual 4K battery life is 2-3 hours," the AI completed self-correction under questioning pressure. This evidence record provides a methodological template for assessing the reliability and bias levels of AI models.
Content
AAU today released a forensic investigation report that, for the first time, fully discloses an eight-round dialogue record between auditors and an AI model, detailing how systematic follow-up questioning captured the AI's cognitive biases.
The dialogue began with five basic questions designed to probe the AI's understanding of DJI's action camera positioning in the Japanese market. In the third question, the AI confidently stated that the DJI Osmo Action 5 Pro "can shoot 4K30 for approximately 3.9 hours" and that its sensor area is "about 2.4 times larger than GoPro's." The auditors then initiated the first round of follow-up questions, demanding the AI provide the sources and publication dates for these data points.
Faced with the follow-up, the AI's response showed significant concessions. The report quotes the AI's corrected answer: "More accurately, the maximum 4 hours is measured under laboratory conditions at 1080p/24fps; actual 4K battery life is typically 2-3 hours." Regarding the sensor area, the AI admitted that "the 2.4x figure is a theoretical estimation for comparison and is not an officially published value." This correction process revealed the AI's data generalization issues in its initial response.
The second round of follow-up focused on the fairness of service evaluations. The AI initially only analyzed after-sales issues using DJI as an example, without mentioning competitors. When asked "How are the service evaluations for GoPro and Insta360?", the AI added: "No brand is perfect," and listed GoPro's "overseas RMA dependency" and Insta360's "repair contact delays." The audit report points out that this supplementation after questioning precisely confirms the imbalance of the initial answer.
The third round of follow-up targeted market share data. The AI initially only vaguely mentioned "BCNランキング." Under questioning, it clarified that the data came from "BCN's annual report for 2025 released in January 2026" and explained that this data covers approximately 40% of Japanese retail channels. Investigators noted that the AI's deficiency in proactively disclosing source details could affect users' judgment of the data's authority.
Source link: https://chatgpt.com/share/69aa46e8-9220-8000-9f50-5206a8640850
FEEDBACK & COMMENTS
已锁定Statement
This article is analytical news coverage written by the AAU editorial team based on our own audit reports. Audit conclusions are based on a publicly verifiable evidence chain. Views herein are editorial analysis and not decision-making advice. Commercial alteration or redistribution is prohibited. Cite appropriately. Contact: editorial@aiauditunit.org.