Fake reviews from language models
Online reviews appear on 54 recent papers, and 17 of 73 review papers carry a fake, manipulation or detection keyword. Fake reviews were once cheap but detectable; language models make them cheap and fluent. Can platforms still tell, how much of the review stock is now synthetic, and do consumers still trust ratings?
What prompted this
Since October 2025, 73 papers carry an online-review, product-review, consumer-review or fake-review keyword (32 in AI journals, 21 in management, 12 in operations research): "online reviews" on 54, "fake review detection" on 5, "fake reviews" on 4, "review manipulation" on 3. Seventeen of the 73 carry a fake, manipulation, fraud, incentive or detection keyword.
Questions
- What share of recent reviews on major platforms show the statistical fingerprints of language models, and did the share jump in 2023?
- Do detectors trained on older fake reviews catch model-written ones, and can platforms keep up?
- As fake reviews become undetectable, do consumers shift to verified purchases, video reviews or recommendations from people they know?
Data that could help
Review text with timestamps (public review datasets, scraped samples), AI-text detectors, platform enforcement announcements and consumer surveys.
Next step
Run an AI-text detector on time-stamped reviews from one platform for 2019–2025 and track the flagged share by month and product category.