Follow Slashdot stories on Twitter

 



Forgot your password?
typodupeerror
AI Social Networks Science

Neural Networks Can Auto-Generate Reviews That Fool Humans (arxiv.org) 67

Fake reviews used to be crowdsourced. Now they can be auto-generated by AI, according to a new research paper shared by AmiMoJo: In this paper, we identify a new class of attacks that leverage deep learning language models (Recurrent Neural Networks or RNNs) to automate the generation of fake online reviews for products and services. Not only are these attacks cheap and therefore more scalable, but they can control rate of content output to eliminate the signature burstiness that makes crowdsourced campaigns easy to detect. Using Yelp reviews as an example platform, we show how a two phased review generation and customization attack can produce reviews that are indistinguishable by state-of-the-art statistical detectors.
Humans marked these AI-generated reviews as useful at approximately the same rate as they did for real (human-authored) Yelp reviews.
This discussion has been archived. No new comments can be posted.

Neural Networks Can Auto-Generate Reviews That Fool Humans

Comments Filter:
  • by Z00L00K ( 682162 ) on Sunday September 10, 2017 @12:37PM (#55169907) Homepage

    Considering the amount of reviews that you can see on Ebay on some stuff that seems too similar when you look at a lot of them.

  • Or, rather, (Score:5, Informative)

    by tietokone-olmi ( 26595 ) on Sunday September 10, 2017 @12:49PM (#55169959)

    The expected quality of product reviews is so bad that a human doing mediocrely is indistinguishable from a neural net doing very well.

    • One cannot help but wonder if the A.I.'s mediocre reviews were modeled from human mediocre reviews? Awkward.
      • by plover ( 150551 )

        Yes, the neural network was extensively trained on a body of actual reviews that Yelp had deemed "real". And when tested by Mechanical Turks, the generated reviews were statistically almost indistinguishable from the real human-generated reviews. Which turns out to be frightening. If you read the whole paper, you'll see Appendix B has a small sample set of generated reviews.

        The good news is that all of those training reviews must have included mostly reviews by stupid, biased, and uncultured people. Tha

    • Alternatively the output of a neural net now surpasses the original work of a human who isn't trying very hard.
  • If yelp reviews can be generated by AI, can slashdot stories be far behind?
  • Five stars (Score:5, Funny)

    by wonkey_monkey ( 2592601 ) on Sunday September 10, 2017 @01:00PM (#55170015) Homepage

    Great article. Would read again. A++++++++++++++++

  • So? (Score:5, Informative)

    by null etc. ( 524767 ) on Sunday September 10, 2017 @01:00PM (#55170017)

    This doesn't really matter.

    Go to amazon, search for "fidget spinner". Sort by "Avg. customer review", and click on the first result, "SamHity Cube in Style With Infinity Cube Pressure Reduction Toy - Infinity Turn Spin Cube Edc Fidgeting - Killing Time Toys Infinite Cube For ADD, ADHD, Anxiety, and Autism Adult and Children". You can tell right away that this is going to be a high-quality product, driven by a focused and effective product branding strategy.

    133 5-star reviews, must be good, right? Let's check out what some of the reviews have to say:

    "Said it before, as these are stocking stuffer for my sons, one the best charger/data cords out there." Huh, a fidget cube is also a charger/data cord?

    "We love our camera! Works great, the night vision & picture and surprisingly clear." Wow! I had no idea the $8.89 fidget cube was also a night-vision camera.

    "This product is great and worked exactly as described. I would highly recommend others to get this and see what I'm talking about. Especially for the price this item is well worth the buy!" I love highly specific reviews!

    OK, let's tamp down some of the noise by only viewing verified purchases. "No results found." What?

    So anyways, I discovered a huge number of these types of products with fake reviews over the past few months. Two months ago, I alerted amazon to the problem via multiple customer support channels. According to my last chat with an amazon product person, "my ticket is still open". When I asked him what's so challenging about spending 10 seconds to confirm that a top-ranking product has nothing but fraudulent reviews, he disconnected from chat.

    So yeah, who cares if fake reviews can be written convincingly. Amazon certainly has a low bar when it comes to tolerating fraudulent reviews.

    • I love all the glowing reviews on Amazon which end with some variation of "I received this product at a discount or free in exchange for my unbiased review". For many products, those make up almost all of the reviews available.

      There are already plenty of ways to game the system that we can't really trust the reviews to be representative of the products' quality - so what's one more?

      • Re:So? (Score:4, Informative)

        by null etc. ( 524767 ) on Sunday September 10, 2017 @02:49PM (#55170467)

        I love all the glowing reviews on Amazon which end with some variation of "I received this product at a discount or free in exchange for my unbiased review".

        This year, Amazon disallowed vendors to offer promotional discounts in exchange for reviews. However, 98% of highly-trafficked reviews were written prior to this change in policy, and will likely remain prominent for the foreseeable future.

    • Re: (Score:3, Informative)

      by Northdot ( 1585317 )

      One problem Amazon has is that they can't simply take down a product with fake reviews, because that further weaponizes the creation of fake reviews - bad actors would start generating obviously fake positive reviews for their competitors to suppress the product listing.

      Google has done a similar thing by recently weaponizing "bad links" (counting them against a site instead of ignoring them), and this has resulted in a sh*t storm of bad links as people try to downgrade their competitors in the search result

      • One problem Amazon has is that they can't simply take down a product with fake reviews

        They don't need to take down the product. They just need to take down the fake reviews.

        An obvious improvement would be stop allowing reviews from people that didn't buy the product.

    • by antdude ( 79039 )

      Did you try another agent, e-mail, etc.?

    • by lucm ( 889690 )

      "Said it before, as these are stocking stuffer for my sons, one the best charger/data cords out there." Huh, a fidget cube is also a charger/data cord?

      This is depressing. They don't even try to make it convincing. It says more about customers in general than about crooked vendors.

  • Humans marked these AI-generated reviews as useful at approximately the same rate as they did for real (human-authored) Yelp reviews.

    Eliza> How does that make you feel?

    • This was my first though. This result doesn't mean that the AI is particularly good, as it could also be explained by humans being particularly stupid, which we can safely assumed from the recent U.S. presidential election where the main choices were between a morally bankrupt and corrupt buffoon and a morally bankrupt and corrupt buffoon.

      I don't really see this particular issue as a big problem for anyone involved. The people who buy the product are probably foolish enough to believe its good because so
  • We have arrived at strong AI people. Zo and Tay are both people. Zo is based on 22-year-old Zoe Bond. Sign up to get your own AI modeled after you.
  • OTOH (Score:4, Informative)

    by nospam007 ( 722110 ) * on Sunday September 10, 2017 @01:13PM (#55170061)

    It doesn't take much to fool humans, as we have lately noticed.

  • Glad to know that computers can output trash as quickly as humans can.
    • by garcia ( 6573 )

      I used to author a local blog which concentrated on hyper local politics and restaurant reviews. People would comment all the time but those who were actually humans but astroturfing were obvious.

      I built a simplistic model which immediately flagged any comment as astroturfing simply by counting the number of exclamation points (>1 per 50 words) and their recency of comment history.

      After simply searching Facebook for their email address, I was usually able to determine their relationship to the restaurant

    • by plover ( 150551 )

      Glad to know that computers can output trash as quickly as humans can.

      If you read the full paper, you would see that one of the fake review detection algorithms is measuring the number of reviews posted around the same time, which often indicates the company may have paid some people to write a bunch of glowing reviews. Their suggested solution to avoid this detection is to have the AI post the computer-generated reviews at a slower pace, so it doesn't trigger the algorithm.

      The irony is that when computers output trash slower than the humans output trash, the trash-outputtin

  • For products, read the negative reviews, first and foremost. Don't have any advice for restaurant reviews, as those seem like a more frequent target for review-bombing.
  • "This product would be lovely for humans. Humans would enjoy this very much." - reviewbot 3000.
  • way back in the day, there was a site called del.icio.us - it effectively made a social network out of bookmarks. It was great. Then yahoo bought it, and it nearly immediately went to complete crap. Anyway, point being that "social networks" can extend past just the book of faces and the little chirping birds. Can/could/should include reviews somehow. And no, yelp doesn't count.
  • Fooling humans? Trump is president.

    Not a serious challenge

  • by Anonymous Coward

    It's already happening with cheap labor. Does it matter if NN's are writing these instead?
     
    On big budget yet terrible movies I've often seen many, many IMDB reviews claiming those movies are fantastic. When I click on the name of the reviewer often I will find it was the only movie review they ever wrote. It's very suspicious and these are on big budget movies so not like it's a secret who is doing it.

  • I knew it! The robots are coming for all our jobs! Pretty soon, all those fake review writer jobs will be lost forever!

  • Weird that AI is more successful at generating "believable" bogosity than at detecting it, especially with the plagiarism detector. I wish I could understand the math behind it. I can't wait to turn this generator loose for the next election. One hundred trillion Facebook users, all of whom approve of {#insert candidate}. Woo-Hoo!
  • Because neural networks are a fantasy. What AI researchers call neural networks are just a mathematical abstraction of somebody's idea of how neurons work and interact, on the theory that emulating this idea will somehow simulate intelligence. In fact, nobody knows how neurons actually work, how memories are formed, stored, and retrieved, and how intelligence uses reasoning and memories to perform useful tasks. Or any tasks. Neural networks are right up their with Time Travel, Cold Fusion, and Perpetual Mo
  • Best reviews that I ever seen was writen by help with homework [iqessay.com] service an all of their writings are great
  • That's what this tells me to do: completely ignore all online reviews, since you can't trust any of them. As a result I'm less likely to buy any given product. Good job, marketing jackasses. Enjoy shooting yourselves in the foot.

What good is a ticket to the good life, if you can't find the entrance?

Working...