On Friday, May 9, the US Publishing Rights Office (USCO) issued a pre -publishing version of the third and long -awaited report in a series of directives on copyright and artificial intelligence. This report, which was followed by the first (published in July 2024), focuses on “digital symmetry” or Deepfakes and the second (published in January 2025) The copyright for the works that were created with the help of artificial intelligence, to train artificial intelligence models. The pre -publication version was issued “in response to the congressional inquiries and expressions of interest from the stakeholders,” noting that “the final version will be published in the near future, without any fundamental changes in the analysis or conclusions.”
As with other reports in the series, USCO does not recommend any government interference at this time; However, it provides a detailed analysis of the potential application of the fair defense in training the artificial intelligence models, as well as strong support for further development of the volunteer licensing market for data training.
While the report emphasizes that a fair use analysis requires the achievement of each case separately, it also provides specific guidelines about some realistic conditions that are likely to reduce or against fair use. Usco notes that although there is no specific formula, “the first and fourth factors can expect a great weight in the analysis.” 2 Although the report does not prevent the possibility of a successful defense for fair use in some cases, its analysis of factors – especially the first and fourth factors more distributed.
In its analysis of the first factor – the purpose of use and its character – USCO is largely focused on the transformative nature of use. Although the report concludes that “the Training of the AI's Molidic Foundation on a large and varied data set will often be transformative”, 3 warns that “the transformation is a degree issue, how the conversion or justification depends on use on the use of the model, how it is directed to the models, and how it is dug.” In the data set, “indicating that” (m) any uses falls somewhere between them “. 5 With regard to commercial cracks, the report confirms that the inquiry is not whether the user is an entity for profit or non -profit, but rather the specific purpose of the use itself.
In addition, the report notes that illegal access – such as pirate work or wall fraud – will weigh strongly with fair use. 7 Although this report states that this factor alone is not decisive, it confirms that this behavior “is going on this position, which makes it in anything else to use it, which leads to clarifying this behavior, which makes it in anything else. A model on pirated or restricted content – especially without suitable handrails to ensure that the outputs do not include parts of the work of the copyright – are largely harmful For fair defense.
The USCO analysis of the second factor-the nature of the work of copyright-does not require this part an analysis of the facts “will differ according to the model and the work in discussion”, commenting that “(w) here are the concerned works more expressive or unpublished previously, the second factor will clarify the fair use.” 10. 10
Likewise, in his analysis of the third worker-the amount of use and the major-USCO recommends evaluating each case separately, which leaves the possibility that, although the use of the entire work will affect the discovery of fair use, “the use of entire businesses seems necessary.”
The report devotes an important analysis of the fourth worker – the influence on the potential market or the value of the copyright work – its absence ”
Usco expresses its concern about the outputs that can serve as direct alternatives to the work of copyright: “If thousands of romantic novels created from artificial intelligence are placed in the market, it is possible that fewer romances' human accounts that have been trained on Amnesty International will be sold.” The violation of its use in the market is expanding this market's theory of “anonymous zone” and generally corresponds to defining the comprehensive priorities of the interests of business owners protected by copyright. In addition to concerns related to market relief based on obstetric intelligence outputs, the report also sheds light on the market concerns for data groups that can be licensed to train artificial intelligence models and encourages a training data license wherever possible: “When licensing options or likely to be feasible, this consideration will show fair use in the fourth factor.” 18
Although the report does not directly address any of the suspended claims regarding the use of copyright works in training artificial intelligence models, it determines a general set of possible results:
At one end of the spectrum, it is possible that the uses of research or non -commercial analysis that do not enable parts of works that must be reproduced in the outputs. On the other side, it is unlikely to be copying expressive work from the sources of pirates in order to generate unrestricted content competing in the market, when the license is reasonable, fair use. Many uses, however, will fall somewhere between 19
The report also discusses in detail the changing comments it received regarding potential interests and considerations regarding licensing of copyright-protected business to train artificial intelligence, including the possibility of a mandatory license system or the subscription cancellation mechanism-where this report is ultimately recommended. (“” CMO “), similar to ASCAP and BMI in the music industry. Although the report “ultimately recommends allowing the licensing market to continue to develop with government intervention”, it also indicates considering “targeted intervention like ECL” if the market fails.
In addition to its analysis of the application of fair use factors and their encouragement of licensing training data wherever possible, the report also discusses the possible means through which it may occur. Two of these discussions deserve to be highlighted. The “Ozan” report discusses the model – or numerical parameters that codify what you learned – to study whether these symbols can form a copy, and therefore, the subsequent reproduction or the use of the weight of the model may amount to violating copyright. Work (currencies) in the case, “focuses mainly on the outputs and whether the final content created is very similar to the work of copyright. 23 has the ability to expand the risks that users of artificial intelligence models may face them with content who object to their work as training data.
The report also discusses the generation of retrieval (“rag”), which usually includes the generation of a mentor or research through which it can recover works or materials responsive to claim and it is noted that this activity includes the reproduction of copyright works. It also warned that such uses are unlikely to be transformed.
By providing examples and analysis of the types of realistic patterns that are likely to support or reduce the discovery of fair use, USCO has provided a long -awaited instructions on the way in which the defense of fair use can be applied to cases that involve artificial intelligence. Due to the fully established USCO, the fully created by artificial intelligence is not qualified to protect copyright, it is not surprising that the important parts of the report are to protect the interests of copyrights-especially with regard to the fourth factor that is very likely to analyze fair use. However, the report ultimately provides ammunition for both sides in many lawsuits pending in terms of copyright and AI.
1 copyright and artificial intelligence, Part 3: Artificial Intelligence Training (Pre -Publishing Version), in I.
2 ID. In 74.
3 ID. In 45.
4 ID. In 46.
5 ID.
6 ID. In 51.
7 ID. At 51-52.
8 ID. In 52.
9 ID. In 62.74.
10 ID. In 54.
11 ID. In 57.
12 ID. In 59.
13 ID. In 61 (internal categories were deleted).
14 ID. In 61.
15 ID. In 65.
16 ID. In 73.
17 ID. In 65.
18 ID. In 73.
19 ID. In 74.
20 IDs. In 103-104.
21 ID. In 106.
22 IDs. At 28-29.
23 ID. In 30.
24 ID. In 30.
25 IDs. In 31, 47.
