By Pascale Caron for EntrepreneurIA

The article “How AI is rewriting the rules for product review sites” (Mediacopilot.ai, Dec. 2025) offers a relevant, though incomplete, analysis of the issues at stake. AI, now trained to summarize, recommend, and compare, increasingly influences the purchasing journey. Conversational engines respond to product queries with near-instantaneous summaries, bypassing the maze of searches and traditional rankings of review sites.

While the effectiveness of this new interface is undeniable, a major question remains: what are these recommendations based on? And, more importantly, can we still trust the reviews it uses, or generates itself?

The Growing Weight of AI in Review Mediation

According to recent data published by Adobe Analytics (Nov. 2025), nearly 34% of traffic to American merchant sites during Black Friday was initiated by interactions with conversational AI. Algorithmic mediation is therefore gradually replacing proactive search: users consult AI as an advisor and accept a synthetic answer rather than a list of results or links.

In this context, online review sites, historically perceived as spaces for democratizing consumer voice, are seeing their influence erode. The promise of authenticity, carried by thousands of customer reviews, remains strong. But it faces two limitations: information overload on one hand, and the growing possibility for AI to generate fictional yet plausible comments on the other.

AI-Generated Reviews Indistinguishable from Human Ones

A 2025 study conducted by Joël Krueger and Lucy Osler (‘AI Gossip‘, Ethics and Information Technology) highlights the ability of large language models to produce persuasive texts mimicking real user reviews. Even more troubling: these fake recommendations, when integrated into platforms, are not detected by human readers or by current automated tools.

Another study, MAiDE-up (Multilingual AI-generated Deceptive Evaluation) published on arXiv in 2024, tested 10,000 AI-generated reviews against 10,000 authentic reviews in ten languages. Result: the rate of fake review detection by humans remains close to chance, even among informed users.

In other words, trust in review platforms now rests on unstable foundations. An online review, if not linked to proof of purchase, could just as easily be the work of a conversational agent as of a real consumer.

The Democratic Risk of Simulated Opinion

This confusion between the real and the plausible is not merely about marketing fraud. It questions the very nature of digital deliberation. If reviews aggregated by AI are themselves generated by other AI, we witness a self-feeding loop, where human judgment is gradually replaced by consensual simulation.

The study Large Language Models as Hidden Persuaders (arXiv, 2025) describes models capable of generating series of targeted reviews designed to influence a specific perception of a product, brand, or competitor. The result? An unprecedented ability to manipulate collective representations at scale through short, convincing, yet artificial texts.

A Regulatory Framework Taking Shape

Faced with this systemic threat, authorities are beginning to respond. In the United States, the Federal Trade Commission (FTC) published an explicit rule in October 2024 prohibiting the creation or distribution of fake reviews, including those produced by AI. This rule authorizes civil sanctions against platforms or companies that have purchased, distributed, or commissioned such content. (©FTC Press Release, 2024)

In Europe, the framework is more fragmented. The “Omnibus” directive, transposed into French law, already prohibits reviews without verification of purchase experience authenticity. But it is especially the European AI Act that could change the game: AI-generated content must be explicitly labeled as such. The question remains whether this obligation can be technically enforced.

The Difficulty of Proof: A Technical and Ethical Question

Identifying the real source of a review requires tracing the content production chain. However, in an open marketplace system, where sellers outsource their communication to agencies or automated tools, responsibility becomes diffuse.

In France, the DGCCRF is experimenting with lexical and behavioral analysis tools to identify clusters of suspicious reviews. But these methods remain challenged by increasingly adaptive generators.

The real challenge lies in creating verifiable systems linking reviews to actual transactions. Amazon, for example, highlights “Verified Purchase” mentions. But are these mentions sufficient in a universe where AI can also write a review for a real buyer without them being the author?

A Reassessment of Recommendation’s Market Value

The business model of review platforms relies on implicit monetization of trust: peer recommendations act as conversion factors. However, if this trust is eroded by suspicion of artificial influences, the entire value creation chain becomes fragile.

Platforms will need to meet a dual imperative: restore review traceability and clarify AI’s role in creating or synthesizing this content. Otherwise, they risk irreversible loss of credibility.

What Solutions to Restore Trust?

Three approaches seem priorities today:

  1. AI Content Labeling: make it technically explicit that content originates from a generative model (watermarking, cryptographic signatures).
  2. Purchase Verification + Identity: link every review to authentic proof of purchase, associated with a validated user profile.
  3. Generation API Framework: impose transparency obligations on AI providers regarding permitted uses (e.g., ban mass automatic review generation without disclosure).

Reviews as a New Object of Algorithmic Regulation

The digital economy is entering a phase where the boundary between authentic production and automatic generation becomes porous. A review is no longer just an expression of experience: it’s a persuasion vector that can be optimized, automated, and instrumentalized.

In this sense, regulating fake AI reviews is not merely about fighting fraud: it engages broader reflection on truthfulness in algorithmic interfaces.

Main sources: