Ideal content raises suspicions of being generated by neural networks

The experiment involved respondents who were asked to identify the author of materials in four formats: text, images, audio, and video. Within each category, different media types were used—from product cards and articles to voice messages and musical compositions. After each block, participants explained their decisions, which allowed them to identify the criteria they relied on. Additionally, respondents' confidence in their recognition abilities was recorded. The highest percentage of correct answers was recorded in the video category—77%. Participants indicated that neural network origins are indicated by violations of physical laws and logic, such as incorrect object movement or desynchronization of audio and video. Text materials ranked second with a 74% success rate. Messenger messages that mimic face-to-face communication were the most easily recognized (82%): respondents noted excessive politeness and perfect formatting as signs of AI, while conversational word order and informal elements betrayed human speech. Product cards were correctly identified 76% of the time. Informational articles proved the most challenging (64%)—in these cases, AI, according to participants, disguises itself behind generalized statements, while humans reveal themselves through irony or self-criticism. In the image category, the overall recognition rate was 65%. Artistic paintings were recognized better (68%)—respondents focused on texture, brush marks, and historical authenticity, while stylistic inconsistencies were immediately attributed to neural network errors. Realistic landscape photographs presented more difficulties (61%): excessively perfect images are often mistaken for artificial imagery, which is due to the habit of professional processing. Music was recognized reliably (75%), with AI-influenced lyrics and unnatural rhymes being indicators. Only 44% of participants correctly identified voice messages. A human voice is distinguished by pauses, breathiness, and hesitation, while the neural network produces sterile diction and speech without natural pauses. Some respondents admitted they couldn't tell the difference and were guessing. The study documented a shift in evaluation criteria: imperfections—typos, shaky camera, hoarse voices, or uneven brushstrokes—are now more often perceived as signs of human authorship. Conversely, hyperrealism, flawless diction, and perfect composition raise suspicions of being artificially generated. Respondents are not opposed to AI as a tool for routine tasks, but they react negatively to attempts to portray it as human in tech support, automated services, and advertising. In emotional domains—music, art, and personal communication—the use of neural networks evokes greater rejection. Participants indicated that they encounter AI content daily and often don't realize it before participating in the test. According to Natalia Kuzmina, Marketing Director of Kokoc Group, the results reflect a shift in markers: while previously smoothness and grammar were considered a sign of quality, now the same characteristics can signal originality. She notes that the essence of authorship is linked not to the origin of the material, but to responsibility and the expertise invested. Around 30% of respondents consider themselves confident experts and rely on technical indicators. About half identified themselves as "pragmatic skeptics," acknowledging that technology is advancing faster than their vigilance. The remaining 20%, after the test, reported a complete loss of confidence in their ability to distinguish AI from humans.