Artificial IntelligenceCryptocurrencyMetaverseCybersecurityTech Forward

Amazon Prime Day: Beware, Fake Reviews May Run Rampant

July 15, 2019, 5:45 PM UTC

According to Fakespot, a New York-based startup that tracks fraudulent reviews on Amazon and other e-commerce sites, Prime Day has become, well, a prime day for fake reviews. The company claims that fake, promotional, paid, or otherwise inauthentic reviews rose 1.13% across Amazon from June 2017 to July 2017. In June 2018 to July 2018, they surged 11.89%. Currently, 34.65% of Amazon reviews are suspect, it says.

The rates may be rising because every year the surge in Amazon shopping around Prime Day creates a higher incentive than usual for sellers to increase their visibility on Amazon Marketplace, the company’s e-commerce platform for third-party merchants. Unsurprisingly, there’s an entire cottage industry built around helping Marketplace sellers to rank higher in search results. And even less surprising, some of the companies in this space use unethical tactics, like offering fake, positive reviews for products.

The scourge of fake reviews are a potentially big problem for Amazon heading into 2019’s Prime Day, because, as a 2016 Pew study found, roughly half of all adults under 50 “always” or “almost always” read online reviews before making a purchase. And of those frequent review-readers, 65% say they trust what they read. Review fraud has even proven serious enough to catch the attention of legislators. On July 9, the House Committee on Energy and Commerce sent a letter to Amazon seeking answers to a series of questions about how the company deals with fake reviews.

“As online consumers, we’ve been trained by Amazon and other platforms to look at online reviews in a very trusting way,” says Fakespot CEO Saoud Khalifah. “The star rating system is somehow ingrained in human psychology.”

Fakespot says charging cables, bluetooth headsets, and other commodified electronics are the biggest target for fake reviews, likely because there’s little other way for sellers of those products to distinguish themselves from the competition. A particularly egregious example is this charging cable, which as of this writing has more than 1,000 customer reviews, each giving the cable five stars. Many of the reviews seem to be fraudulent, including one from a verified purchaser—a designation given to reviewers by Amazon itself—which claims the charging cable is “Very lovely. Looks great in my yard.”

Despite such examples, Amazon broadly rejects Fakespot’s claims. “Companies such as Fakespot that claim to ‘check’ reviews cannot concretely determine the authenticity of a review, as they do not have access to Amazon’s proprietary data such as reviewer, seller, and product history,” the company said in a statement to Fortune. Amazon concedes that it is aware of “bad actors that attempt to abuse our systems,” but says it invests significant resources into both detecting and blocking fake reviews, and taking legal action against fraudsters.

As for Fakespot, Amazon has review of its own, saying: “These companies have business models that are inherently biased towards instilling distrust in reviews on Amazon’s and other companies’ stores.”

Fakespot doesn’t monetize its consumer-focused tool, a web plugin that rates the trustworthiness of individual reviews. But the company does sell its services to manufacturers, who can better measure consumers’ feelings about their products by weeding out fake reviews.

Another independent review-vetting project, Reviewmeta, also disputes Fakespot’s analysis. Reviewmeta says that in 2017 and 2018, it found no rise in either overall average star-ratings, the percentage of deleted reviews, or the level of inauthentic reviews flagged by its own metrics. Reviewmeta does, however, believe that Amazon generally understates its fake-review problem, and puts the overall rate of fakes at “under 10%.” Amazon has generally maintained that less than 1% of its reviews are inauthentic.

The discrepancy between these numbers may be partly due to the nebulous definition of what qualifies as a “fake” review: Reviewmeta, Fakespot, and Amazon itself each have their own metrics, which can encompass everything from automated spam reviews to hand-crafted fakes bought from real Amazon customers. Reviewmeta and Fakespot each have their own complex algorithms that apply their rubrics on a massive scale, but even some honest behavior, such as a wave of positive reviews from a product’s loyal fans, could be identified as suspicious.

According to Khalifah, automated posts from bots are easy to spot algorithmically, and Amazon does tend to eliminate those quickly. “But the problem is much tougher than bot-generated reviews,” he says. “The reviews that are really sketchy are the ones that are biased and promotional.”

According to both Khalifah and a recent Buzzfeed News report, some of those biased reviews are the product of a secretive network of ‘black hat’ operators who help sellers recruit and pay real Amazon customers to write biased reviews, both paying them for their reviews and reimbursing them for their (genuine) purchases. Because of these tactics, Khalifah says Amazon’s ‘verified purchase’ tag on a review has become “worthless.”

To avoid falling for those flimflam reviews, Reviewmeta founder Tommy Noonan encourages shoppers to “actually read reviews instead of just relying on the average rating,” look at new reviews first, and be sure to read critical reviews.

The creative tactics used by review-fakers make the problem of inauthentic reviews particularly tough to solve. In many ways it’s similar to the challenge of detecting hate speech or harassment on social media sites like Facebook or Twitter: Because automated detection methods are trained on past examples, new tactics or language can often outmaneuver the filters. Even Facebook’s head of automated detection doesn’t think A.I. will ever effectively catch all bad behavior on that platform.

But Khalifah is much more optimistic. He claims that Fakespot’s own detection efforts, refined over the five years since he founded the company, have become good enough that they can detect “all” fraudulent reviews, including those written by humans. Both Fakespot and Reviewmeta use both automated language analysis and other data about both reviewer and seller behavior.

Fakespot also monitors several other major outlets, and Khalifah says Amazon is doing a better job at screening reviews than most. He says Wal-Mart, Sephora, and other outlets have fake-review problems that are at least as bad as Amazon’s, if not worse. And Amazon, as the biggest online retailer, attracts the most review manipulation.

But the rub is that once a fake review is detected, fixing it can be a slow process. Khalifah says that when Fakespot flags fake reviews, Amazon usually deletes them “three or four months later.” And Amazon has few effective tools for stopping deceptive reviews at the source, because Marketplace sellers can easily evade bans.

“If Amazon kicks you out, you pop up the next day with a completely different legal entity, selling the same thing,” says Khalifah. “Wash and repeat.”

In short: Buyer beware.

Related reading on Amazon Prime Day 2019:

What to buy (and avoid) on Amazon Prime Day 2019

—8 ways to track the best Amazon Prime Day deals

—Black Friday vs. Prime Day: Which event has better deals?

How to shop Amazon Prime Day without overspending

—Prime Day 2019: By the numbersCatch up with Data Sheet, Fortune‘s daily digest on the business of tech.

Catch up with Data Sheet, Fortune‘s daily digest on the business of tech.