On July 10, 2026, TikTok announced it is testing improvements to its spam detection system specifically designed to identify AI-generated content. The platform flagged AI spam related to politics, financial advice, medical information, and “topics that could pose a risk to public trust or well-being.”
TikTok is a social video platform. It is not a marketplace. And yet it is investing engineering resources into detecting AI-generated content because the volume and quality of AI spam has reached a level where automated systems cannot distinguish it from human output.
Now consider Amazon. Amazon receives millions of product reviews per week. A significant percentage are fake. The fakes used to be easy to spot: broken English, generic phrasing, five-star ratings with no substance. Those days are over.
The Cost of Generating a Fake Review Just Hit Zero
The economics of fake reviews have changed fundamentally. In 2020, a fake review cost between $5 and $15 from a review farm. The reviews were often written by underpaid workers in Southeast Asia who produced low-quality, detectable content. Brands had to spend real money to get a meaningful number of fake reviews.
In 2026, the cost is effectively zero. OpenAI’s GPT-5.6 Sol, launched July 9, is the most capable consumer-available language model to date. Meta’s Muse Spark 1.1 and SpaceX’s Grok 4.5 launched the same week. Open-weight models from Chinese labs match or exceed earlier frontier models and can run on consumer hardware for free.
Anyone with a laptop can generate 10,000 product reviews in an afternoon. The reviews are fluent. They include specific product details. They vary in tone, length, and sentiment. They reference shipping times, packaging quality, and use cases that sound authentic. They are, to any text-based detection system, indistinguishable from real reviews.
This is not a future threat. It is happening now. The tools are available. The incentives are aligned. The detection infrastructure is not ready.
Why Amazon’s Detection Cannot Keep Up
Amazon has invested in fake review detection for years. The company uses machine learning models to analyze review patterns, flag suspicious clusters, and remove inauthentic content. Amazon sued the administrators of Facebook groups that coordinated fake review campaigns. The company reports removing hundreds of millions of fake reviews annually.
But Amazon’s detection systems were designed for a previous era of fake reviews. They look for patterns that characterized human-generated fakes: review bursts, identical text across multiple reviews, accounts with no purchase history, unusual rating distributions. These signals worked when fake reviews were produced by humans following templates.
LLM-generated reviews do not produce these signals. Each review is unique. The language model generates natural variation in sentence structure, word choice, and detail emphasis. A sophisticated operation can simulate review velocity by spacing submissions over days or weeks. Accounts can be warmed up with legitimate purchases before posting the fake review. The text passes every grammatical and stylistic check.
Amazon’s own filing with the FTC acknowledged the scale of the problem: the company removed over 200 million suspected fake reviews in 2023 alone, and the number has grown since. But removal is reactive. The reviews influence rankings and purchases before they are caught, if they are caught at all. And with LLMs, the cost of creating a new batch of reviews after a takedown is minutes of compute time.
The Detection Problem Is Asymmetric
The fundamental challenge is asymmetry. Generating fake reviews is cheap, fast, and scalable with LLMs. Detecting them requires expensive, slow, and constantly evolving systems that must be trained on new generation patterns.
Every time a detection system learns to identify a specific pattern of LLM-generated text, the generation side adapts. A prompt change, a temperature adjustment, or a different model produces reviews with different statistical signatures. The detection system must be retrained. The generation system just needs a new prompt.
This is the same adversarial dynamic that plays out in cybersecurity, where attackers and defenders cycle through exploits and patches. But in cybersecurity, the defender’s advantage is that attacks must succeed completely while defense only needs to detect one signal. In review detection, the attacker’s advantage is that a single fake review, undetected, contributes to a product’s ranking. Thousands of undetected fake reviews dominate the ranking entirely.
The platforms have a structural conflict of interest that makes this worse. Amazon benefits from more reviews, more engagement, and more products ranking well. Aggressive review removal reduces marketplace activity, which reduces GMV, which reduces revenue. The incentive is to remove enough fake reviews to avoid regulatory action and bad press, but not so many that the marketplace looks empty.
What Independent Trust Infrastructure Looks Like
The solution to LLM-generated fake reviews is not better detection inside the marketplace. It is independent trust infrastructure that operates outside the marketplace’s commercial incentives.
GoBuy is built on this principle. The platform does not sell products. It does not take a cut of transactions. It has no reason to tolerate fake reviews because fake reviews undermine the only thing it sells: trust signal.
GoBuy’s review authenticity analysis operates on different principles than platform-native detection:
Cross-platform verification. A product reviewed on Amazon, Walmart, Best Buy, and independent forums should show consistent quality signals across all channels. LLM-generated reviews exist only on the platform where they were planted. GoBuy cross-references review data across sources to identify discrepancies.
Reviewer behavior analysis. GoBuy analyzes reviewer history, account age, review cadence, and product category overlap. Even when individual reviews are textually perfect, the patterns of reviewer accounts reveal coordinated campaigns. Review farms, even sophisticated ones, leave behavioral traces.
Temporal signal analysis. Authentic reviews accumulate organically over time as products are purchased and used. Fake review campaigns create rating velocity anomalies. GoBuy’s Smart Score algorithm weights review stability over time, not just aggregate ratings. A product that jumps from 3.2 to 4.7 in two weeks gets scrutinized regardless of how natural the review text reads.
Quality-weighted aggregation. Not all authentic reviews are equally valuable. A detailed review from a verified purchaser with a long reviewing history contributes more to the Smart Score than a short review from a new account. This means that even if some fake reviews penetrate the filter, their impact on the final score is limited by the quality weighting.
The result is a Smart Score from 0 to 100 that reflects genuine product quality, not the output of a review manipulation campaign. Products that maintain a score above 80 for 90 days earn the GoBuy Verified badge. This is not a measure of popularity. It is a measure of sustained, verified quality.
Why This Matters for AI Agents
The fake review problem becomes catastrophic when AI agents enter the picture. ChatGPT Work, launched July 9, includes browsing and shopping capabilities. Claude, through MCP, can query product information. Google Gemini and Amazon’s Project Moonraker are building autonomous shopping agents.
These agents read Amazon reviews as data. They process ratings, review counts, and review text. They use this data to formulate product recommendations. They do not distinguish between authentic and fake reviews because, textually, there is no difference.
When an AI agent reads 500 LLM-generated five-star reviews and recommends a product based on that data, it is laundering fake reviews through a layer of algorithmic authority. The consumer receives the recommendation from a trusted AI assistant, not from a suspicious-looking Amazon listing. The trust transfer makes the manipulation more effective, not less.
This is why independent trust infrastructure must be available to AI agents, not just human shoppers. GoBuy’s MCP server, at gobuy.ai/api/mcp, exposes trust-scored product data to any MCP-compatible agent. Instead of reading raw Amazon data, the agent queries GoBuy and receives products ranked by Smart Score, with fake reviews already filtered out.
The agent makes better recommendations. The consumer makes better purchases. The review manipulation fails.
The Market Will Not Self-Correct
There is no market mechanism that solves the fake review problem. Consumers cannot tell the difference between authentic and fake reviews when the fakes are LLM-generated. They rely on the star rating, which is corrupted. They rely on the review text, which is fabricated. They cannot audit Amazon’s detection systems or verify that reviews were removed.
Competing marketplaces face the same problem. Walmart, eBay, and independent e-commerce sites all rely on user-generated reviews. The same LLMs that generate fake Amazon reviews generate fake reviews everywhere.
Only an independent layer, with no commercial stake in the outcome, can provide the trust signal that the market needs. This is not a feature that Amazon can build. It is not a plugin that a marketplace can offer. It is infrastructure that exists outside the marketplace and serves the buyer.
The Path Forward
The fake review arms race will intensify. As detection improves, generation improves faster because LLMs advance faster than detection systems. The platforms will play whack-a-mole, removing individual review operations while new ones emerge within hours.
The durable solution is not better detection inside the marketplace. It is trusted scoring outside the marketplace. GoBuy’s Smart Score does not try to identify every fake review. It builds a quality signal that is resistant to manipulation by weighting the right signals and ignoring the noise.
For consumers: check the Smart Score before you buy. The Chrome extension overlays it directly on Amazon product pages.
For developers building AI shopping agents: integrate the GoBuy MCP server so your agent consults trust data before recommending products. Documentation is at gobuy.ai/agent-docs.
For the platforms losing the arms race: the answer is not a better detector. It is an independent referee.
Start at gobuy.ai for verified product scores. Developers: integrate the GoBuy MCP server at gobuy.ai/agent-docs to give your AI agents a trust layer that works.