Readers unaccustomed to poetry are unable to reliably differentiate that generated by artificial intelligence (IA) than that written by authors such as Emily Dickinson or Shakespeare, and they even rate the first better, according to a study of more than 1,500 people.
Research from the University of Pittsburgh (USA) indicates that this tendency to value poetry generated by AI may be due to the fact that Readers confuse the complexity of verses written by humans with the inconsistencies that this technology can sometimes incur.
Another reason may be that participants may underestimate how human generative AI can be, indicates the study published in Science Advances.
Furthermore, volunteers who were not informed whether the text had been artificially produced or came from a person rated the former more favorably, which may respond to the fact that they were simpler and more accessible than the work of prominent poets.
The researchers tested the ability of 1,634 people to distinguish between poetry generated by AI and that written by a poet.
Overall, the group’s level of experience with poetry was low: 90.4% said they read this genre a couple of times a year or less; 55.8% described themselves as ‘not very familiar with poetry’ and 66.8% as ‘not at all familiar’ with the assigned poet.
The group worked with ten poems in random order, five written by poets such as William Shakespeare, Lord Byron, Emily Dickinson and TS Eliot, and the rest generated by ChatGPT3.5 in the style of those authors.
The result was that participants were more likely to attribute the AI’s poems to having been written by a human.and the five considered least likely to be of human production were all written by authentic poets.
In a second experiment, a different group of 696 people evaluated the writings based on 14 characteristics such as quality, beauty, emotion, rhythm and originality.
The authors observed that IA poems were rated more favorably for qualities such as rhythm and beautywhich contributed to their misidentification as written by humans.
The volunteers were randomly assigned to three groups. One was told that the texts had been written by a human being, another that they were produced by AI, and the third was not given any information about authorship.
Those told they were created by AI gave lower scores on 13 characteristics than participants who were told they were written by humans, regardless of whether the poems were created by AI or humans.
Participants who were not told anything about authorship rated AI-generated poems more favorably than those written by humans.
The authors suggest that participants preferred AI-generated poems because they were simpler and more accessible than the work of prominent poets.
In addition, volunteers expect to prefer poetry written by humans and, since AI-generated poetry is easier for them to interpret and understand, they mistakenly understand this preference as an indication that the poem was written by a human, write the authors of the investigation.