Was this art made by a computer?

In recent months AI produced artwork has been all over TikTok, Twitter, and traditional media. Tools like Midjourney and Dall-E being accessible to those who are interested means that producing art with AI has never been easier. I see no end of content on my social media which is just a series of AI generated images, like “if this cartoon was an 80s fantasy movie”, or “if these animals were dressed in suits”. Before these highly advanced AI art tools were around, Dall-E mini became a brief meme sensation with blurry, indicative images of increasingly surreal prompts. The content really does write/paint itself.

This growth in AI art has not been without controversy. There have been examples of AI art models trained to mimic the style of living artists without their permission, for example. It raises a number of both philosophical and ethical questions, about whether AI is a tool to create art no different to a digital paint software or something wholly different, or the importance of human emotion and expression to the production of “art”. 

Increasing awareness of AI art comes at the same time as ChatGPT is demonstrating the power of large language models. Many of the long-standing expectations around the sorts of jobs and tasks which can be automated have been called into question. It has always been easy to see how a robot can replace a worker on a factory line, or self-checkout reduces the number of human cashiers needed, but “creative” tasks like producing art, writing code or generating new text which is cogent have tended to feel unachievable.

I wanted to explore a few things with this polling. How believable people find AI art. Whether informing people that art is AI generated has any impact on how it is perceived (is it seen to be less beautiful, is it liked less). Finally, whether showing people AI art has any impact on how they perceive the abilities of AI in general. Effectively, are we looking at a potentially huge change in public opinion as AI art and AI text becomes more visible. In a year's time, will the skill-sets that people value or think have the greatest longevity, have completely changed in the face of AI which appears to be “creating” in a way that it has not before.

We find:

  • AI Art is now at a level where people often misindentify it as human art. People are more likely to think human-made abstract art is AI produced than non-abstract AI-made art.

  • Informing people that a piece of art was produced by AI rather than a human artist tends to reduce people’s assessments that it is beautiful and how much they like it, although we do find examples that buck this trend.

  • Seeing examples of AI art increases people’s belief that AI is capable of producing visual art and beautiful art, and that an Artists job could be done by a machine, although it has a much smaller impact on beliefs that AI has the emotions needed to produce good art, or that AI can produce an “original” piece of art. 

  • This impact is sector specific, and seeing AI art does not shift other general attitudes towards AI, or beliefs that AI could do jobs outside of art

Methodological approach

For this work we used 6 pieces of art, 3 of which were pieces of art from human artists, 3 of which were generated by Dall-E 2 given instructions to produce art in the style of these human artists. We chose to have a still life, impressionist and abstract piece of art in each of the human and AI categories. The art we used is shown below, hover over it to see if it is AI or human generated

A core part of the design of this research involved splitting the sample to test the impact of showing AI art on attitudes towards AI, and the impact of giving information that art was AI generated on perceptions of the art itself. The details of the order which questions were presented to participants are provided here:

Stage

Group A - Informed

Group B - Uninformed

Stage 1

Questions on attitudes to art

Questions on attitudes to art

Stage 2

Questions on current attitudes to AI, including belief that various jobs and tasks can be automated, and whether AI is conscious

Questions on the beauty of various pieces of art unattributed to human or AI artist (random 4 of the 6 pieces of art shown to each participant)

Stage 3

Questions on which pieces of art were AI or human made

Questions on which pieces of art were AI or human made

Stage 4

Questions on the beauty of various pieces of art with information on human or AI artist (random 4 of the 6 pieces of art shown to each participant)

Questions on current attitudes to AI, including belief that various jobs and tasks can be automated, and whether AI is conscious

This structure allowed us to compare:

  • Perceptions of beauty of each piece of art among an informed and uninformed audience

  • Attitudes towards AI among a sample which had completed the process of evaluating whether AI generated paintings were believable and among those who had not been shown any examples of AI art

Believability of AI Art

As a first piece of analysis, we wanted to understand how easy people found it to identify pieces of art which were produced by an AI. We find a lot of uncertainty (likely spurred by the fact that asking participants to consider whether art is human or AI generated is going to make people doubt their judgements somewhat). In general there are no large differences between the AI art and human art when asked directly whether the art is human or AI. 

In fact, the biggest difference we find is that both human and AI abstract pieces of art are more likely to be identified as AI generated. 37% said the AI generated abstract was probably or definitely human, 33% said this for the human abstract art. For all other art we tested this number surpassed 50%. This finding was true for both those who say they do not really understand the point of abstract art and those who do not.

When given a line up of all the art and asked to select the options which people felt were AI generated (including a “None of the above” option), we find the abstract art pieces are respectively selected by 30% of respondents. 32% of respondents say none of the pieces shown were AI generated. The human impressionist painting was the least likely to be selected, being selected by 9% of respondents.

We find younger respondents tend to have a greater selection rate than older respondents (23% of 18-24s say none of the pieces are AI generated compared to 38% of those 65 and over). However, the younger group has a higher rate of false-positive identification too. This is most pronounced for the human still life, where 26% of 18-24s identify this as AI generated compared to just 6% of those 65 and over.

Overall, participants were unable to consistently identify the AI generated art, and instead tended to identify abstract art as AI generated.

Impact of AI information on perceptions of beauty

As we might expect, informing participants that art was AI generated substantially reduces how likely they are to say that the art looks like it took a lot of work to create.

Participants evaluated each of the 4 pieces of art they saw on a scale of 1-7 for how much they liked it, and on a scale of 1-5 how much they agreed that it was “beautiful”. The results are somewhat inconsistent, but we do generally find that revealing the artist to be an AI reduces both perceptions of likeability and beauty. This is particularly true of the impressionist pieces, easily the highest scoring pieces among the uninformed groups. For both this one, and the abstract conditions we find revealing the artist to be an AI reduces likeability scores by around 0.4 on the 1-7 scale. 

On beauty we find that in the Abstract and Impressionist examples, estimates of beauty not only significantly decrease for the AI artworks when the participant is informed, but also significantly increase for the human-generated artwork.

The Still Life example is the only one to buck the trend. Here we find a significant increase in the beauty estimates for both AI and Human artworks when the participant is informed, although neither is matched by significant differences in likeability scores.

Impact of showing AI art on belief in what AI can achieve

Those who had gone through the process of evaluating AI art, and therefore seen AI art, were as expected more likely to say that an AI was capable of producing visual art (65% compared to 48% who had not done the task) and capable of producing a work of art which is beautiful (52% compared to 36% who had not done the task).

However, the elevated beliefs in the capabilities of AI were sector-specific. Completing the AI task didn’t significantly change how capable participants felt AI would be on a range of other tasks, such as feeling emotions, driving a car, or having a friendship with a human.

And similarly for professions which people felt could be carried out by machines/AI. The task selectively impacted the perception that Artists’ jobs could be done by AI. Seeing the examples of art produced by AI made the role of artists seem about as possible to be automated as the jobs of taxi drivers and waiting staff.

In some instances there were more interesting differences between age groups on these questions than there were between those who had seen the AI art. For example, many more of the youngest respondents believe an AI could feel emotions, and an outright majority believe an AI could have a friendship with a human. Similarly, while the AI art display made no difference to whether people felt an AI could ever be conscious, a majority of 18-24s (56%) believe AI could one-day be conscious (or indeed already is) compared to just 34% of those aged 65 and over.

When asked directly if people felt that AI-produced art is currently a risk to artists’ incomes, 55% who had seen the AI art before agreed, compared to 45% who had not. 

We asked a few other questions about AI art specifically, mainly testing how different wording choices impact the way people think about this issue. While a large minority of respondents feel that there is some art which AI could replace, and nearly half of those who had seen the AI art, about as many agree that AI can never have the emotion needed to produce good art. And uncertainty remains around the “originality” of the art. When asked if “AI could never create an original piece of art”, 36% agreed and 28% disagreed among those who had not seen the AI art examples, 32% and 33% among those who had. This represents a considerably smaller shift in opinion than we see for the belief that Artist jobs could be automated.

And seeing AI art made no real difference to how people felt about the advancements of AI. Regardless of whether they had seen the art, the top response was “Sceptical”, followed by “Worried” and “Interested”.

Previous
Previous

AI Risk and Public Opinion - The State of Play

Next
Next

Science has determined the best* Christmas film