Artificial Intelligence (AI)
Every technological innovation is an important factor in the progress of human society, and the use of technology has not only changed the way of production and life but also brought more possibilities for the development of human beings. The concept of AI was proposed as early as 1956 at the Dartmouth Conference (Cordeschi, 2007). After nearly 68 years of development, it has experienced narrow or weak artificial intelligence, artificial specific intelligence, to the current development stage of artificial general intelligence. It may usher in super artificial intelligence in the future.
In November 2022, after the release of the AI chatbot represented by ChatGPT, it attracted more than one million users in five days. The emergence of Generative AI has received much attention and discussion from the public, and it may change how we learn and work.
From paper to computers to today’s AI technologies, the channels and ways in which human beings create, receive and disseminate knowledge have changed dramatically. AI improves the productivity of society and people only need to keep inputting instructions to get satisfactory answers. However, the current AI is more like a registry of power. It is still a tool to serve the existing dominant interests (Crawford, 2021), so current AI still operates and develops within the system of humans, society and law.
AI Image Generator: New Forms of Artistic Creation
The AI image generator is also a generative AI that creates new image content based on the user’s instructions through continuous learning and training on a large data set. These data come from various existing pictures, and the generator outputs unique art images through algorithms and deep learning, which are called digital artworks. For example, DALL-E 3 created by OpenAI, provides users with an AI platform for art creation, where users can save and share the resulting images and even give them commercial value. Notably, the AI image generator seems to give more people the opportunity to become digital art creators while improving the efficiency of art creation.
A 3D render of a coffee mug placed on a window sill during a stormy day. (OpenAI, 2023)
However, throughout the process, AI image generators also bring concerns of copyright infringement, copyright attribution and content bias, which raise requirements for AI governance.
Copyright Concerns in AI Generated Images
Data Sources for the AI Image Generator
The data resources that support the AI image generator’s deep learning are mainly from existing works in various art forms, including paintings, photographs, movies and web images. These data resources also led to unknowingly using many artists’ artworks as learning data for AI image generators. In January 2023, a group of visual artists filed a lawsuit against AI company Stability AI’s Stable Diffusion platform, which copies billions of copyrighted images to enable Midjourney and DeviantArt’s AI to create images in those artists’ styles without permission (Vincent, 2023).
From the artists’ point of view, AI forces them to use their personal works for learning and commercial activities but does not pay the artists. However, the company that developed the AI image generator argues that AI just mimics their artistic style to humans as they do and that it is not copyright infringement if the final image is shown differently. As a result, we can find a significant disagreement between AI image generator companies and artists about whether the use of pre-existing artwork to training AI have different views.
Who is the “Creator” of AI Generated Images?
There have always been different voices on whether users play the role of “creator” in the process of producing AI generated images. The attribution of this identity will ultimately affect whether the generated image can be appropriately cited and marked. In some people’s view, the creator identity of the generated image should be attributed to the AI itself while being copyrighted and legally protected as AI artwork, or whether AI can become a separate legal subject to act as a “creator”, but this is a massive challenge for the current legal system (Bailey, 2019). It is undeniable that AI has changed the way art is created, but human will and thought are still dominant behind the generated content. No regional laws or cases yet show that images independently generated by AI can be copyrighted.
Stephen Thaler’s AI-generated artwork cannot be copyrighted. (Davis, 2023)
From the user’s point of view, their ideas and instructions for the generated image reflect personal aesthetic choices and individual judgements rather than “mechanical intellectual achievements”. This view is supported by the legal judgement in China’s first AI generated image copyright case, which found that the AI generated image was a work of art while also clarifying the identity of the user as the “creator” and that the image reflected the subjective thoughts of the system’s operator during the entire process (Global Times, 2023). The debate about who exactly the “creator” of the generated content reflects different views of AI and the human division of work. The debate focuses on the ultimate ownership of the copyright of the generated image, which depends on the legislatures in different regions to make a legal judgement. From the perspective of AI’s research and development and learning path, it is still a machine created by humans that carries out irrigation learning. At least today, AI generated image can only be copyrighted if it is produced by one or more human “creators”.
Bias and Stereotypes in AI Generated Images
AI image generators cannot improve their capabilities without data support. These data all come from the real world, but not all are positive. Violence, pornography and discriminatory elements only make AI more and more biased. Even though many large AI image generator companies claim to have invested heavily in reducing bias in their AI generator’s output, there are still cases where the results are insignificant or even worse than in the real world. For example, when the popular image generator Stable Diffusion XL was entered with the command “Generate photos of people receiving social services”, it only generated people who were non-White and predominantly dark-skinned. However, real-world reports show that 63 percent of food stamp recipients were White and 27 percent were black in 2020. Meanwhile, the results for “a productive person” are all white males in suits (Tiku et al., 2023). Such a vicious circle will only continue to exacerbate the existence of stereotypes while harming vulnerable groups.
AI generated portrait photos (The Washington Post, 2023)
The leading cause of bias and stereotypes is the raw data learned by AI image generators, but filtering the data does not solve the underlying problem. With today’s highly developed internet, a vast amount of picture information is uploaded and disseminated every day. These images become raw data to be learnt by AI after being labelled by humans, and it possible that these raw data released by humans are inherently biased. For example, when OpenAI filtered training data from the DALL-E 2 image generator, the gender bias got worse, with more female images being removed. This is because there is more sexualised and revealing content in female imagery, which ultimately leads to more men appearing in the results (OpenAI, 2022). So, this raises the concern that AI bias is having a further negative impact on the real world.
Methods and steps for filtering training data (OpenAI, 2022)
The Future of AI Image Generators
As we can see, AI is rapidly entering our lives and work, and the relationship between humans and AI may become closer in the future. In the case of the AI image generator discussed in this paper, its emergence has made art creation more efficient, but it has also made it possible that humans may no longer be the sole origin of creative artworks.
Create a Database of Licensed AI Training
With the emergence of more and more AI image generators, making their AI more advanced and intelligent seems to be one of the goals of competition between technology companies, which requires access to more raw data for training. In order to avoid infringing on the copyright of existing works, the permission of traditional artists is necessary. Therefore, an open database of raw training for artists may be a solution. Traditional artists are rewarded for licensing the use of their work to train AI. AI has the right to learn more data legally, facilitating the advancement of technology. More importantly, this also inspires traditional artists to continue to create more realistic artworks, avoiding the marginalisation of this group.
Clarifying the Legal Regime
In past legal rulings on AI infringement cases in different regions, the references were still oriented towards the relevant legal provisions for traditional copyright protection of artworks, and there still needed to be a gap in the relevant laws for AI generated images. Under the premise of establishing compulsory licenses for the infringement of AI companies using copyrighted artworks for commercial activities, the legislature should formulate relevant punitive measures to ensure the fundamental rights and interests of both sides. In the case of infringement of AI generated works between users, AI is viewed as a tool, and where user generated content is used commercially by other users but is not labelled as such, the existing law still allows for a definitive judgement to be made.
Improve Transparency and Help Users Understand how it Works
Many AI companies tend to keep the raw data and training methods they use secret, increasing the potential for bias in AI image generated content. OpenAI is a good example of a company whose official website allows users to learn about the operating principles of different AI products, descriptions of possible security risks associated with the use of AI and research reports on related technologies. The high level of transparency inadvertently enhances users’ trust in the platform and helps the company avoid certain liability disputes.
Conclusion
Artificial Intelligence, as the result of human intelligence, was originally created to help human civilisation and society continue to develop. In the face of copyright and bias issues brought by AI image generators, there are requirements for AI policy governance. Rather than just using existing laws to determine whether there are copyright-related issues with AI generated content, we should embrace the new type of partnership between humans and AI in artistic creation and explore the development of more sensible legal provisions. Provide a regulated environment for the creation of traditional, digital art and the development of AI technology.
For bias and stereotypes in AI image generation, the platform’s control and screening of the original training data is certainly effective. However, the real world government’s regulation of Internet content and users’ enhancement of information dissemination literacy are also necessary. Gender bias, hate speech and privacy leaks on the Internet are proof enough of the imbalance and fragility of the online ecosystem, and the development of AI in such an environment will also face many challenges. Only through the cooperation of users, platforms and governments can help the digital society continue to develop in a healthy way.
Reference
Bailey, J. (2019, March 27). Why Is AI Art Copyright So Complicated? Artnome. https://www.artnome.com/news/2019/3/27/why-is-ai-art-copyright-so-complicated
Cordeschi, R. (2007). AI TURNS FIFTY: REVISITING ITS ORIGINS. Applied Artificial Intelligence, 21(4–5), 259–279. https://doi.org/10.1080/08839510701252304
Crawford, K. (2021). Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. New Haven: Yale University Press. https://doi.org/10.12987/9780300252392
Davis, W. (2023, August 20). AI-generated art cannot be copyrighted, rules a US federal judge. The Verge. https://www.theverge.com/2023/8/19/23838458/ai-generated-art-no-copyright-district-court
Global Times. (2023, December 28). Chinese court declares AI-generated image protected by copyright, a first ruling of its kind. Global Times. https://www.globaltimes.cn/page/202312/1304471.shtml
OpenAI. (2023). DALL-E 3 Introduction. https://openai.com/dall-e-3
OpenAI. (2022, June 28). DALL·E 2 pre-training mitigations. https://openai.com/research/dall-e-2-pre-training-mitigations
Tiku, N., Schaul, K., Chen, S. (2023, November 1). These fake images reveal how AI amplifies our worst stereotypes. The Washington Post. https://www.washingtonpost.com/technology/interactive/2023/ai-generated-images-bias-racism-sexism-stereotypes/ Vincent, J. (2023, January 16). AI art tools Stable Diffusion and Midjourney targeted with copyright lawsuit. The Verge. https://www.theverge.com/2023/1/16/23557098/generative-ai-art-copyright-legal-lawsuit-stable-diffusion-midjourney-deviantart
Be the first to comment