The Zhitong Finance app learned that according to a research report released by Everbright Overseas, overall, copyright protection technology and policies for model training data at home and abroad have yet to be refined, and future copyright provider stock prices are expected to be intensively catalyzed. Looking ahead, the maturity of data attribution technology is closely linked to the amount of content generated by AI. As the commercial potential of AIGC's downstream applications is unleashed, it is expected that it will continue to drive the increase in licensing revenue for copyright providers. The development of AIGC is more beneficial to copyright holders than bad, and poor expectations are driving the stock price back up. It is optimistic that the copyright protection of domestic model training data will continue to improve, driving the performance growth of copyright providers in various information media such as news media, film, and television. Recommended attention: 1) AI+ Publishing: China Publishing (601949.SH), China Science Biography (601858.SH); 2) Image copyright library: Visual China (000681.SZ); 3) Film and television copyright library: Jebsen (300182.SZ), Huace Film (300133.SZ).

Everbright Overseas's main views are as follows:

Data sets used for AI model training have a number of ways to avoid infringement or directly compensate copyright holders.

Proprietary data: 1) Copyright cooperation agreements: Shutterstock, Axel Springer and many other copyright providers have established cooperation with AI companies; 2) Paid API access: Some highly specialized data providers will charge for API access. In '23, API access on Reddit and Twitter changed from free to paid.

Open source data: 1) Open license agreements: including CC, ODC, CDLA, etc.; 2) Specific data scraping strategies: such as complying with the website's Robot.txt file; 3) Community monitoring: improving data set transparency.

Direct compensation to creators: 1) Prior compensation: The technical difficulty is low, but it is difficult to define a reasonable compensation amount, such as the contributor fund established by Shutterstock; 2) Ex-post compensation: traceability of AI-generated content. The pricing is reasonable but the technology is not yet mature, such as the attribution model algorithm published by Carnegie Mellon University.

Private data sets: Data sets suitable for model training are sold directly, or packaged as MaaS services.

The profit model of overseas copyright cooperation agreements is stable, and the prospects for commercialization are initially showing.

The content generated by AI may pose a threat to the traditional business of copyright providers. Cooperation between copyright providers and AI companies is mutually beneficial and win-win. 1) Multimedia material library Shutterstock: Generates revenue through the sale of model training materials, launches an AI-generated image section, and provides AI tools supported by OpenAI; 2) Publisher Axel Springer: sells his publications to OpenAI as training materials, and jointly uses AI technology to enhance the user experience.

Looking at the cooperation between copyright banks and AI companies from Shutterstock: the overall benefits of AIGC are stronger than the disadvantages.

1) Benefits: Shutterstock's data licensing revenue is already clearly reflected in the performance side, driving valuation repairs and stock price recovery. Revenue from selling model training materials in 23Q3 accounted for 19.5% of the company's total revenue;

2) Unfavorable: The downturn in Shutterstock's traditional business since 23 years has been more affected by competition in the same industry. Similar company Getty Image has performed steadily, and AIGC's threat and replacement to the copyright storage industry is not obvious yet.

Copyright regulations for model training data at home and abroad have yet to be improved, and the stock prices of copyright holders are expected to be intensively catalyzed.

A 22M12 paper shows that Stable Diffusion reproduces the details of famous paintings at the pixel level. Disputes over AIGC's copyright issues and related regulations can be divided into two main categories:

1) Copyright definition of AI-generated content: The US does not recognize that AI-generated content has copyright, while China tends to protect the copyright of AI-generated content;

2) Copyright regulations for model training data: The United States and the European Union clearly require the use of copyrighted materials to train models, while Japan believes that training data is not protected by copyright.

Risk warning: AI technology development and product iteration fall short of expectations; risk of increased competition in the AI industry; risk of commercialization not progressing as expected; domestic and foreign policy risks.