Dissecting AI-generated fake reviews: Detection and analysis of GPT-based restaurant reviews on social media

Alessandro Gambetti, Qiwei Han

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Recent advances in generative models such as GPT may be used to fabricate indistinguishable fake customer reviews at a much lower cost, posing challenges for social media platforms to detect this kind of content. This study addresses two research questions: (1) the effective detection of AI-generated restaurant reviews generated from high-quality elite authentic reviews, and (2) the comparison of out-of-sample predicted AI-generated reviews and authentic reviews across multiple dimensions of review, user, restaurant, and content characteristics. We fine-tuned a GPT text detector to predict fake reviews, significantly outperforming existing solutions. We applied the model to predict non-elite reviews that already passed the Yelp filtering system, revealing that AI-generated reviews typically score higher ratings, users posting such content have less established Yelp reputations and AI-generated reviews are more comprehensible and less linguistically complex than human-generated reviews. Notably, machine-generated reviews are more prevalent in low-traffic restaurants in terms of customer visits.
Original languageEnglish
Title of host publicationInternational Conference on Information Systems (ICIS) 2023
Number of pages17
Publication statusPublished - 11 Dec 2023
EventInternational Conference on Information Systems - Hyderabad, India
Duration: 10 Dec 202313 Dec 2023
Conference number: 43
https://icis2023.aisconferences.org/

Conference

ConferenceInternational Conference on Information Systems
Abbreviated titleICIS
Country/TerritoryIndia
CityHyderabad
Period10/12/2313/12/23
Internet address

Fingerprint

Dive into the research topics of 'Dissecting AI-generated fake reviews: Detection and analysis of GPT-based restaurant reviews on social media'. Together they form a unique fingerprint.

Cite this