Cracking the "anti-AI" sentiment, Germany's Max Planck reveals: Humans prefer self-customized AI art

Source: “Xin Zhiyuan” (ID: AI_era), author: LRS

Since the release of high-quality AI painting tools such as DALL-E 2, Stable Diffusion, and Midjourney last year, the debate on “AI artists” has never stopped, such as whether AI works can participate in competitions, model training infringes copyright, hard-working There are many problems such as the reduction of painters to corpus.

However, most people have reached a consensus that if an exquisite AI work is put together with a master’s handwriting, we will still think that human works are better, and we are willing to understand the stories behind the works, the emotions expressed, etc.

But, are AI works really inferior to human works?

If it were an AI work specially customized for your emotions and experiences, would you turn around for it?

AI art can be beautiful, but means little to humans

Researchers from Duke University, the University of Waterloo, the University of Cambridge, and more have explored whether and why humans dislike AI-generated artwork.

Paper link:

The researchers first recruited 150 participants from the online platform, and asked the participants to evaluate 30 AI artworks selected from the art platform Artbreeder in different categories. Some of the works were marked as “human creations”, including 15 Abstract art, 15 pieces depicting concrete objects or scenes.

The results of the experiment showed that artworks marked as “created by humans” received significantly higher ratings than artworks marked as “created by AI”, which shows that participants have potential “anti-AI bias”, and they think that AI works are not so valuable , and not profound.

The classification results of emotional value such as storytelling and emotion show that artworks marked as created by humans also receive higher evaluations.

The researchers believe that when artworks were labeled as human creations, participants were more likely to believe and value the stories behind the artworks, which in turn provided positive support for the artworks’ sheer sensory enjoyment.

The research results also emphasize that human appreciation of art not only includes technology, but also needs to consider emotion, intelligence, and the story behind the artwork.

Let AI works draw into your heart

Although humans have their own psychology of “anti-AI art”, researchers from the Max Planck Institute of Empirical Aesthetics, Erlangen-Nuremberg University and other institutions published an article in the journal Psychological Science The paper reveals the principle behind the “art aesthetics”. They believe that it is not that AI works are not good enough, but that they are not customized according to the viewer.

Paper link:

Research results show that the aesthetic appeal of a work of art is closely related to the viewer’s personal preferences.

When faced with similar pictures, viewers are more likely to choose works of art related to their past experiences and cultural background.

Offline Experiment Settings

To draw preliminary correlations between self-relevance assessments and aesthetic ratings, the researchers recruited 33 German-speaking experimental participants (29 women, 4 men), all aged between 18 and 55. All had normal vision and no neurological diseases.

The researchers selected 148 lesser-known works of art from museum collections, covering works of various time periods, styles and genres of American, European and Asian cultures. Participants were required to rate the works of art according to their personal subjective pleasure .

List of works:

Participants were also asked to rate the artwork’s relevance to themselves, that is, the degree to which they relate to themselves, past experiences, identity, etc.

In addition to subjective ratings, before the experiment, the researchers also collected basic information about the participants, such as educational background, age, gender, left-handedness, sexual orientation, mental illness diagnosis, etc., and conducted some art and aesthetics-related education. .

Online Experiment Settings

In order to expand the size of the participants, the researchers also recruited 208 English-speaking participants online (135 men, 70 women, 2 other gender, 1 unspecified), with an age range of 18-74 years old, each People viewed 42 artworks and were asked to answer a series of questions.

There are two subjective questions that are more critical, “To what extent did the image move you?” (To what extent did the image move you?) and “How much beauty can you feel from the picture?” (How much did it you get the feeling of beauty?)

In the self-relevance assessment session, participants were asked to answer “How self-relevant is the image to you?” after re-viewing the image.

Experimental results

After calculating the correlation between aesthetic ratings and self-evaluation, the researchers used a linear mixed model with three different conditions to predict the ratings:

1. Use only participant autocorrelation scores as cutoff

2. Add autocorrelation slope

3. Add image-related truncation and slope

It was found that model 3 had the best prediction effect.

When using G*Power for posterior statistical testing, the effect size (power) reached 0.89, indicating that there is a relatively strong correlation between autocorrelation and aesthetic ratings.

Personalized style migration

Supported by the above conclusions, the researchers conducted a second experiment to verify the source of aesthetic differences between AI artworks and human works.

Experimental Settings

The researchers recruited 45 participants online (28 men, 15 women, 2 unknown), fluent in German and aged between 18 and 55 years.

The researchers first selected 20 artworks from previous data, covering natural, artificial content, indoor, outdoor scenes, and different architectural structures, and then used 3 transfer styles to generate AI works, totaling 80 images.

Migration model paper:

The specific migration style is related to the participants’ personal life experiences and the places, objects, food, animals and cultural relics mentioned in the cultural background survey.

During the test, participants did not know which works were generated by AI and needed to intuitively choose whether each work was unfamiliar, familiar, or definitely recognized.

### Experimental results

After using a similar experimental process as before, it can be found that the participants’ aesthetic ratings are obviously concentrated on the autocorrelation dimension, further strengthening the correlation between the two.

By customizing style transfer works, we can see that self-relevance is an important determinant of aesthetic scores. For example, if a participant once spent a perfect vacation in Helsinki, he will be more inclined to give Helsinki a similar style. High marks for images.

When analyzing the aesthetic differences between autocorrelation and other works of art, it can be seen that the style score of autocorrelation even exceeds that of “real works hand-painted by masters”.

Summarize

The experience of viewing artworks is just a typical example of how people interact with the external world and are deeply affected. Moreover, aesthetic taste is highly individualized and everyone has different experiences.

The conclusions of these two studies are not contradictory. The aesthetic rating of visual art has a very high correlation with self-relevance, and human experience makes us more inclined to choose human works.

Cem Uran, author of the second study, said that with real art, it all depends on what visual elements the viewer can spot, and may not even be aware of the specific elements, so just like certain art without knowing why.

However, the power of self-relevant reference information to attract viewers also highlights the serious dangers of misuse of personalized content.

With the development of various content recommendation platforms using recommendation algorithms, this trend is becoming more and more obvious, and users who are deeply involved in it may not be aware of this problem.

References