The Federation of Screenwriters in Europe (FSE) and the International Affiliation of Writers Guilds (IAWG) express their concerns regarding the second draft of the Code of Practice on AI
The FSE and IAWG advocate for the inclusion of the following key points in the Code:
Before addressing technical measures and safeguards, it is essential to recognize that the current framework of the Code of Practice is fundamentally flawed. By failing to acknowledge that the Text and Data Mining (TDM) exceptions outlined in Articles 3 and 4 of the DSM Directive were not intended to authorize the reproduction of works for generative AI purposes, the draft perpetuates a systemic misunderstanding of the legal framework. This omission forces stakeholders to engage in damage control rather than addressing the root cause of unauthorized data usage. Without rectifying this legal misinterpretation, any proposed measures risk being perceived as legitimizing unlawful practices.
1. Clear mechanisms for rights reservations (TDM)
It is essential that authors can indicate that their works must not be used to train AI models, as provided by European law. This requires simple and accessible tools to express these rights reservations in a way that is understandable and respected by technology companies
Current issue: The current draft does not propose any concrete or standardized solution to allow authors to effectively express their reservations in a recognized manner.
The AI Office’s recommendation to use robots.txt files as a way for rightsholders to opt-out of data scraping highlights a significant gap in understanding the realities of content usage in AI training. The robots.txt protocol was designed to manage web crawlers, not to protect creative works, and it requires authors to specify each crawler they want to block individually. This is an overly burdensome and ineffective solution, as it places the entire responsibility on rightsholders rather than on the companies that use their work.
Moreover, recent investigations (The Atlantic, The Ankler) revealed that generative AI systems are often trained using offline data sets, such as subtitle files extracted from DVDs and streaming services. These data sets, used by companies like Meta, Apple, and Nvidia, show that the issue goes beyond simple web crawling—it involves large-scale aggregation of creative works from sources that are inaccessible to the robots.txt mechanism.
2. Recognition of authors’ moral rights
The use of copyrighted works for training purposes undermines their integrity, as these works were never intended to be disassembled and used to establish correlations with other works for the generation of new content. The Code should explicitly address moral rights to align with international copyright law and prevent unauthorized uses that distort the original intent and artistic vision of creators.
3. Regulation of open-source models used for commercial purposes
Open-source AI models are often shared freely, but this does not mean they should be exempt from rules. Copyright law must be respected by all users of these models, regardless of their commercial or non-commercial nature, except when a lawful exception applies in compliance with the Three-Step Test. Regulation must apply at the distribution stage of these models to establish safeguards for protecting authors’ rights before they are integrated into any application. The use of an open-source model must not be a pretext for bypassing copyright obligations.
Current issue: The second draft does not provide any guarantees in this regard.
Concrete example: Meta’s open-source models, such as LLaMA and ImageBind, are used by third parties for commercial applications without any oversight or transparency, illustrating a significant risk of circumventing creators’ rights.
In fact, recently unredacted documents in the Kadrey v. Meta copyright lawsuit suggest that Meta may have knowingly used a dataset of pirated books for training and created a script to remove copyright identifiers.
4. Safeguards on the scientific research exception
The exception that allows works to be used for scientific research should not become a loophole for abuse. Currently, a model developed “for research purposes” can be sold to commercial entities without any control.
Current issue: The Code must include strict rules to prevent protected works from being subsequently used in commercial products without respecting authors’ rights.
Proposal: Include audits or mechanisms for tracing the datasets used during the training of models, focusing on newly released models and significant updates to ensure transparency. For example, some research initiatives already use ‘data logs’—similar to a logbook—to record the sources used and ensure that the origin of the data is documented.
5. Increased oversight of SMEs benefiting from exemptions
The second draft exempts small and medium-sized enterprises (SMEs) from certain obligations due to cost considerations. However, some SMEs can have significant resources, with thresholds of up to €50 million in annual turnover or €43 million in total balance sheet
Current issue: No entity, regardless of size, should be exempt from respecting copyright obligations. These thresholds demonstrate that many SMEs have sufficient financial capacity to adhere to basic compliance measures. Allowing such entities to bypass their responsibilities risks normalizing the unauthorized use of creators’ works. Exemptions in the name of innovation must not be used as a shield for circumventing accountability.
Concrete example: Meta’s LLaMA model has reportedly reached 650 million downloads, making it nearly impossible for any authority, including the understaffed AI Office, to track how these instances are used—whether by small companies, commercial developers, or large-scale businesses. This highlights the urgent need for stricter oversight, even for entities classified as SMEs
6. Creation of a globally accessible rights database
A centralized database would allow authors to declare their works and rights reservations in a visible and accessible manner worldwide. This would enhance transparency and prevent situations where works are used without authorization due to a lack of clear information.
Proposal: This database, managed by a trusted authority such as the EUIPO, should be open to rightsholders worldwide to prevent the fragmentation of rights reservations and ensure interoperability of rights management systems on a global scale.
Conclusion
The FSE and IAWG call on all stakeholders to integrate these adjustments into the Code of Practice to ensure a balance between innovation and the protection of creators’ rights.
“To concede rights today in the name of progress is to invite oblivion tomorrow in the name of profit.”
Let us not forget that behind every algorithm lies the irreplaceable creative work of authors that gives meaning to our shared cultural fabric. We urge policymakers and industry leaders to uphold their duty: to ensure that innovation does not come at the expense of integrity, equity, and respect for those who shape the stories that inspire humanity.
FSE (Federation of Screenwriters in Europe) is a network of national and regional associations, guilds and unions, established in 2001. It comprises 29 organisations from 26 countries, representing more than 10,000 screenwriters in Europe.
For further information, please contact:
Denis Goulette, Délégué Général
Email: d.g@federationscreenwriters.eu
IAWG (International Affiliation of Writers Guilds) is a global network of national associations, guilds and unions, established in 1986. It comprises 15 member organisations, representing 60,000 screenwriters worldwide.
For further information, please contact:
Sarah Dearing, Secretariat
Email: sarah@iawg.org