Artificial intelligence seems to face a peculiar paradox when it enters the territory of creativity: we can prefer what it writes and, at the same time, penalize it for having been written by a machine. A new study published in Judgment and Decision Makinga journal from Cambridge University Press, has found that readers rated short stories generated by ChatGPT as higher quality and more absorbing than those written by human authors. However, regardless of the actual origin of the text, scores improved when participants were told that there was a person behind it.

The research, titled “Bot or not”, was carried out by Sydney Sears and Deena Skolnick Weisberg, from Villanova University, and includes three studies carried out with adults. The first had 1,682 participants; the second, with 424; and the third, with 481. The samples were selected to be demographically representative of the United States in age, gender, and race. The researchers used three previously published human stories and three corresponding texts generated using ChatGPT 4.0, each of approximately 1,000 words.

The results draw a significant distance between the ideas about artificial creativity and the response to the result when it is unknown who produced it. In the words of the researchers, the participants “they actually prefer AI-generated stories while believing these stories are worse”.

Human authorship continues to provide value

The first of the experiments included 1,682 adults, recruited to form a representative sample of the American population in age, gender and race. Each participant read one of six stories of approximately 1,000 words: three had been written by human authors and previously published in literary magazines or collections of stories, and another three had been generated with ChatGPT 4.0 following themes and characteristics similar to the first.

But there was a second variable. Some participants were correctly told who had written the story; others were deliberately given false authorship. In this way, the researchers were able to separate two issues that usually appear mixed: the perceived quality of the text and the effect caused by knowing that it has been generated using AI.

The result was that, on a scale of -3 to 3, the stories generated by ChatGPT reached an average rating of 1.42 in absorption, compared to 1.00 for the human stories. In perceived quality, the difference was even greater: 1.54 for AI texts compared to 0.97 for humans.
However, the perception changed when attributed authorship came into play. Stories presented as written by a person obtained an average of 1.32 in absorption, compared to 1.10 when they were presented as generated by AI. In quality, the scores were respectively 1.40 and 1.12. That is, the content produced by ChatGPT was liked more, but the human label also made any content more liked.

Knowing that there is an AI behind it modifies perception

The finding is consistent with what academic literature calls algorithm aversion: the tendency to evaluate certain results less well when we know that they come from an algorithm.

The study itself recalls previous research in which this effect has been observed in areas such as counseling, poetry or the visual arts. But the research by Sears and Weisberg allows us to observe it together with another phenomenon: the ability of generative models to produce certain content that, evaluated by themselves, can be especially attractive to the public.

A possible explanation proposed by the researchers is the ease of processing. Previous literature has observed that AI-generated content tends to be more fluid and easier to interpret, characteristics that people tend to value positively. Human stories, on the other hand, can introduce greater ambiguity, complexity, or interpretive difficulty.

Thus, an easy-to-absorb text is not necessarily better literature. The authors themselves warn that quality and absorption capacity are different concepts, although in their experiment they have evolved in a practically identical way. Literary fiction may precisely require the reader to confront less obvious meanings, while an extremely absorbing story may have little literary quality.

When you need to recognize ChatGPT, the reader fails

Studies 2 and 3 removed the authorship information and posed a more direct test: can we really know what a person has written and what ChatGPT has written?

Participants received two related stories, one human and the other generated by AI, and had to identify which one corresponded to each authorship.
424 adults participated in the second study. Only 40% were correct, a result lower than the 50% that would be expected by chance. The researchers later replicated the experiment with another 481 participants. This time they got 52% correct, statistically indistinguishable from chance.
Thus, the two experiments do not provide general evidence that readers can reliably distinguish human fiction from that generated by ChatGPT. In fact, the paper highlights that the participants did not exceed chance in any of them.

On the other hand, the researchers studied whether previous experience could improve the ability to detect the origin of texts. And they found a difference between knowledge of literature and knowledge of artificial intelligence.
In the second study, greater self-reported experience with AI systems was significantly related to a greater likelihood of correctly identifying the stories. Experience with fictional literature, on the other hand, showed no relationship with success.

The third study again found the same pattern using a previously validated artificial intelligence literacy scale: those who scored higher were more likely to answer correctly, while again no significant relationship with literary experience appeared.

Even some of the clues that readers intuitively used seemed to work against them. In the second study, those who reported noticing the language were more likely to get it wrong. In the third, the same thing happened with those who relied on the language or how much they had enjoyed the story. Symbolism, on the other hand, was marginally associated with a higher probability of getting it right in the latter experiment.
The conclusion raised by the researchers is that perhaps we are looking for the wrong signals when we try to recognize text generated by AI.

A more pleasant, but also more homogeneous, creativity

Why might we like these stories more? In addition to its ease of interpretation, the authors propose a hypothesis related to the functioning of the models. ChatGPT builds new content from huge amounts of previous material, and, the study argues, that process could in some sense average out existing trends and smooth out some of the irregularities of human works, producing results that are palatable to a greater number of people.

But precisely there could also be one of its limitations. The study includes research that indicates that the creative results produced by AI tend to be relatively homogeneous. An individual human response may be less creative than one generated by AI, but the set of human responses has significantly greater variability.

The researchers also distinguish two moments of creativity: generating possibilities and subsequently selecting which ones are worth developing. The models used in this research were especially oriented to the former, while the human process incorporates reflection and selection. In fact, previous research cited in the paper shows that when a person intervenes by selecting between different results generated by AI, the evaluation of the resulting work can improve even more.

The question, therefore, transcends the competition over who “writes better.” The study points to deeper differences between what kind of results each creative system produces and by what process it arrives at them.
The authors are particularly careful to point out the limits of their results. The texts analyzed were short stories, of around 1,000 words and belonging to realistic fiction. We don’t know if the pattern would hold up in a full-length novel, where developing characters, plots, and narrative coherence over tens of thousands of words makes very different demands.

We also do not know if the same would happen in other literary genres or cultures. The participants were American adults between the ages of 18 and 81, and the authors expressly note that perceptions of AI and narrative traditions may vary between countries. Furthermore, in the identification studies, participants were expressly informed that one of the texts was human and another was artificial, information that we normally do not have when we read content outside of a laboratory. And the experience measures used were self-declared, another limitation recognized by the researchers.

Therefore, the most relevant conclusion of the work may not be that artificial intelligence has proven to be “more creative” than people. The thing is that when we judge only certain results, our preferences can directly contradict what we believe about who should be able to produce them. And this distance between product and origin introduces a new variable in the debate on creativity and artificial intelligence. A machine can already generate something similar to a human creation, but how much does our judgment about a work change when we discover who – or what – is behind it?

More info: Bot or not