含羞草社区

At the centre of the present debate over artificial intelligence (AI) lies concern about large language models (LLM) and their training data. Many legal questions arise, such as whether LLMs may use copyright data for training; whether that process is infringement or fair use, and whether the output is transformative or merely reproductive. New disputes occur daily regarding creator rights, innovation and AI. The Ministry of Commerce and Industry has now joined the fray, issuing its (paper).

The paper examines whether Indian copyright law resolves concerns raised by generative AI. It focuses on two key issues, the input of copyright works used for training and the output of AI-generated content. In part 1, the paper deals with the first, pointing out that the lack of access to copyright content raises “major public interest concerns”. This, of course, leaves unanswered the basic question as to how requiring access to copyright training data for private AI models may be of major public interest. However, this anomaly is nothing compared to the problems of the hybrid model proposed in the paper.

Mandatory blanket licence for AI

This recommends the introduction of a mandatory blanket licence for AI systems using copyright-protected works in exchange for statutory compensation. This model is not only unique but also untested because it expressly rejects proven systems in other jurisdictions.

The model rejects the popular opt-out mechanism, which permits rights-holders to choose whether to allow their works to be used in training. Opting out is burdensome, it reasons. With no transparency of the data used, opt-outs will not in reality help rights holders. Voluntary, or direct, licensing agreements are dismissed for being logistically inconvenient because of scale and transaction costs. The paper speculates that, should negotiations fail, models would lose access to content, and therefore introduce bias.

Extended collective licensing, backed by the United States Copyright Office, is vetoed because it is a voluntary framework with negotiable terms. The text and data mining exception, adopted by the United Kingdom, is seen as lacking, because rights holders may not allow works to be used for training, thus degrading the quality of LLMs by hindering their need to access large representative datasets. Under the model, however, developers will need to obtain no separate permission to “lawfully access” copyright material, including that of rights holders outside India.

Mandatory AI licensing: Promise and flaws

The model appears to balance the needs of LLMs with creator rights, by rejecting a zero price licence, that is one that makes copyright content freely available as training data. It proposes that a non-profit government body be established to collect royalties from developers. These would then be disbursed to rights holders. Developers would not have to negotiate individual terms with multiple parties.

For all its good intentions, the model is flawed. Its flat revenue-sharing arrangement means that rights holders receive royalties at government-fixed rates. Such rights holders must register with the official body, and accept, without objection, the royalties so determined. As well as a lack of clarity about the royalty setting mechanism and its implementation, the proposal does not take into account the detrimental economic effects on creative industries. Royalties will be a percentage of revenue generated by LLMs. No revenue earned denies recompence, even if rights holders’ works are used for training.

The model requires developers to self-declare the data used for training, in return for which they gain access to all copyright works. This would seem to be of little use.

Overreaching model needs practical reform

The model may be innovative, but is also overambitious and overreaching. It is inconsistent with the exceptions to copyright under article 9(2) of the Berne Convention and article 13 of the Trade-related Aspects of Intellectual Property Rights Agreement.

In trying to balance the rights and needs of all stakeholders, the model succeeds only in adding greater complexity and more burdensome compliance. Because it is in a working paper, future inclusive stakeholder consultations should result in a more practical assessment of existing statutory frameworks. They should produce solutions that are principles-based, flexible, agnostic and workable.

Essenese Obhan is the managing partner and Sumathi Chandrashekaran is a consultant at Obhan Mason