Ethical Challenges of AI-Based Consumer Product Review Analysis

Abstract

Online consumer reviews have emerged as a valuable source of online information, enabling prospective buyers to consider the experiences of earlier consumers. This information, however, can be supplemented by the application of machine intelligence on scale, including sentiment analysis, review deception detection, reviewer-behavior modeling, review summarization, review aggregation and ranking, and automated moderation. While these approaches have the potential to complement human moderation by mitigating information overload and targeted deception, they pose important ethical issues.

Automated review analysis could suppress genuine consumer grievances, amplify conventional stylistic preferences, erroneously discourage useful expression, infer attributes of anonymous consumers outside the scope of legitimate research, produce inscrutable scores of authenticity, and enable commercial stakeholders to bypass public scrutiny by artificially inflating review or comment prominence. This research develops a Responsible-AI framework for consumer product-review analysis that combines automated deception detection, protective filtering of authentic negative reviews, cross-style linguistic fairness, explainability, data minimization, manual review, and consumer-trust preservation. A composited deception audit containing 2,400 simulated review analysis cases, allocated at equal parts among Sentiment-Only AI, Deception-Aware AI, Explainable AI, and Responsible AI was implemented without use of any consumer or platform identity, purchase record, proprietary deployment data, or real reviewer data. The Responsible-AI framework realized an F1 score of 0.91, false-positive legitimate review removal of 4.1%, and manipulation leakage of 5.2%. In contrast, the deception-aware approach attained an F1 score of 0.90 while suffering a false-positive legitimate review removal of 10.9% and manipulation leakage of 10.4%. Logistic regression modeling of the efficacy scores inferred that rating-burst automation, commercial secrecy, and linkage to consumer identity significantly exacerbate ethical risks, whereas range of explanation, scrutiny of dubious review appeals, inclusion of evidence, and cross-standard fairness enactments are protective. Thus, responsible product-review analysis should not focus on spam removal alone but rather protect genuine experience-sharing, detect fraudulent coordination, provide rationale for conclusions, limit undue oversight, and establish procedures for evidence submission and review appeal