Scientists from the University of Chicago (USA) conducted a study, during which they showed how AI can be used to write complex fake reviews. Such reviews are indistinguishable from real ones by modern methods, and unsuspecting readers find them highly reliable.
Restaurant reviews were generated using recurrent neural networks (deep learning techniques) that previously trained on the thousands of real-life reviews available online.
According to the researchers, the generated reviews were virtually indistinguishable from the real ones. So, the authors of the work showed that users not only did not recognize fake reviews, but also considered them as useful as real ones written by people.
The latter is probably the most alarming. Since this essentially means that reviews written by AI are fulfilling their main function - they purposefully influence the opinion of people.
It is noted that plagiarism was rarely found in such reviews (using software). This is due to the fact that the AI generated them by character, rather than plucking words from real reviews.
Today there is a fairly large underground industry where people write fake reviews (for money). However, as noted by University of Chicago professor Ben Y. Zhao in an interview with Business Insider, the introduction of AI could undermine it. The scientist says that he does not yet know about the use of such algorithms in this industry. But there is no guarantee that someone will not come up with something similar and will not use it for personal gain.
At the same time, the researchers write that the responses of the neural network were still not ideal. It turned out that the algorithm used a smaller set of characters - and this was not difficult to see. However, according to the authors of the work, future neural networks may be even more complex, and, accordingly, the feedback they generate will be more difficult to detect.
Jiao notes that it's not just about fake restaurant reviews: such technologies can generally shake our beliefs about what is real and what is not.
Promotional video:
The text of the study can be found here.