How does #genAI affect creative professions? Does domain training also provide an #AI edge? In "Expertise elevates AI usage: experimental evidence comparing laypeople and professional artists" we ran controlled tasks w 50 artists + laypeople & a #GPT4o agent: arxiv.org/abs/2501.12374 Results... ⏬ - ThreadSky

andreskarjus.bsky.social • 32 days ago

How does #genAI affect creative professions? Does domain training also provide an #AI edge?
In "Expertise elevates AI usage: experimental evidence comparing laypeople and professional artists"
we ran controlled tasks w 50 artists + laypeople & a #GPT4o agent: https://arxiv.org/abs/2501.12374
Results...
⏬

1 / 2

Comments

andreskarjus.bsky.social•32 days ago

...are both surprising and also not. We ran 2 tasks, copying & creative, emulating artistic practice using #stablediffusion. Trained artists outperform laypeople in both as expected. GPT does well in copy.. but outperforms even professionals in the creative task 😯
Wait what?

andreskarjus.bsky.social•32 days ago

The setup was: participant looks at reference image and either writes a prompt to recreate it, or in creative: to create something as different as possible (but we fixed the first part of the original generation prompt e.g. "A photo of a mini cooper", so they had to get creative). The top results:

andreskarjus.bsky.social•32 days ago

Various tricks worked in creative, like hiding the original object or changing style.
Now, below are the best (left) and worst (right) by GPT4o, given the same input images and instructed to write prompts.

andreskarjus.bsky.social•32 days ago

Quantitatively (we used CLIP embeddings+cosine similarity) GPT does better than the average artist (but not top artists!), but since we used a single instruction prompt its strategy is usually the same: make everything "vibrant" and underwater 😁
What does all this mean?

andreskarjus.bsky.social•32 days ago

Anyone can now create artistic imagery (of some quality). Gap between laypeople vs artists using these tools is there but not huge. LLMs excel in instructing other AIs & output is cheap.
Seems like we should teach students (of arts &everybody else) to use these tools...

andreskarjus.bsky.social•32 days ago

Sentiments differ though: we also carried out a small survey among the participants, and many artists are less than enthusiastic about AI (not without reasons as we discuss in the paper).

margen.bsky.social•32 days ago

This study seems very unethical. You let artists and other people use models that already have the works of trained professional artists optimized into them without consent, literally plagiarism machines, and then compare the results like making art has something to do with typing some words.

andreskarjus.bsky.social•32 days ago

Ethics&exploitation are important topics, but to study how (by now, billions of) people use these tools, we need to run realistic experiments. The options the way I see it are: A) study nothing bc everything AI-related may be from some angle problematic or B) study how the world works best we can.

andreskarjus.bsky.social•32 days ago

We did apply, according to academic regulations, and received approval to conduct the study from the ethics board of the main host institution.

margen.bsky.social•32 days ago

It's basically: "Look, artists, we took your life's work and build AI models that offer the same service as you did, but you are no longer needed. And, oh, look, random lazy persons can now make art!" The reality is that the internet is flooded with garbage and artists are not sharing new works.

andreskarjus.bsky.social•32 days ago

Reading the actual paper before commenting might be fruitful here. This is not what we're saying, rather almost the opposite.

margen.bsky.social•32 days ago

Of course I read it (most of it). One thing I noticed is that artists were mostly critical towards using AI. And you are still concluding that AI could "enhance" making art, but making art is not something to speed up, it's a personal journey. Artists don't need AI, AI needs artists to scrape from.

Comments

Posting Rules

Reply