Copyright, Personality Rights and transparency Compliance for AI Digital Human Content Production in the EU

Key Takeaways

• AI digital human compliance is not limited to copyright. Providers must also assess whether the digital human imitates as an identifiable person’s likeness, voice, gestures or other personal attributes
• Commercial text and data mining may fall under Article 4 of the DSM Directive only where the works are lawfully accessible and rightholders have not reserved their rights in an appropriate manner, including by machine-readable means for online content.
• Where a digital Human is based on a real person, providers should obtain clear likeness, voice, performance and data-processing permissions, with defined scope, duration, territory, platforms, advertising use and synthetic-replication rights,
• Under Article 50 of the AI Act, providers of AI systems generating synthetic audio, image, video or text content must ensure that outputs are marked in a machine-readable format, while deployers may have separate disclosure duties for deepfakes and certain AI-generated text. For general-purpose AI model providers, Article 53 of the AI Act requires a copyright policy and detailed summaries for model training.

1. Protected Content and Personal Attributes in AI Digital Human Content Production

The development of AI digital humans is a multi-layered process that intersects with several protected works under the Information Society Directive (Directive 2001/29/EC, InfoSoc Directive). To achieve AI digital human copyright compliance, technical founders must recognize that a single virtual presenter is a composite of different intellectual properties. Under Article 2 of InfoSoc Directive, authors have the exclusive right to authorize or prohibit direct or indirect, temporary or permanent reproduction by any means and in any form, in whole or in part. This broad protection covers the source materials used to train and render the digital human. A failure to audit these inputs can lead to significant liability for SaaS companies entering the European market.

1.1 Images, Photographs and Visual Materials

AI digital human copyright compliance begins with the visual training sets. Photographs used to train facial movements or skin textures are protected under Article 2(a) of Directive 2001/29/EC. As clarified by the Court of Justice of the European Union (CJEU) in Case C-145/10 (Painer), a portrait photograph is protected if it is an intellectual creation reflecting the author's personality through free and creative choices. If your SaaS uses celebrity images or stock photography without specific generative AI licenses, you risk infringing these reproduction rights.

1.2 Videos, Livestream Clips and Audiovisual Works

The use of existing videos and livestream clips to train motion models or 'lip-sync' algorithms constitutes a reproduction of audiovisual works. Under EU law, these works enjoy protection not only for the creative content but also for the related rights of the producers. Digital human service providers must ensure that the motion capture data or video training sets do not incorporate proprietary footage from variety shows, sports broadcasts, or competitor livestreams without explicit consent, as these are high-risk materials.

1.3 Music, Voice-Over, Sound Effects and Performances

Voice cloning raises a broader set of issues than copyright in sound recordings. The source material may be protected as a musical work, phonogram, performance, or audiovisual recording. In addition, where a cloned voice remains identifiable as a real person’s voice, personality rights and data protection rules may also be relevant. Providers should therefore distinguish between the right to use a recording, the right to imitate or synthesize a person’s voice, and the right to deploy that synthetic voice in commercial content. A licence for the audio file does not necessarily authorize voice cloning, model training, advertising use, or cross-platform distribution.

1.4 Texts, News Reports, Scripts and Course Materials

The scripts that AI digital humans read are often protected literary works. While facts are not copyrightable, the specific expression in news reports, educational course materials and original scripts is protected under InfoSoc Directive. If an AI system ingests these texts to generate 'new' scripts for a virtual presenter, it may be infringing the original author's rights unless it falls under a specific exception, such as the quotation exception in Article 5(3)(d).

1.5 Character Designs, Virtual Humans and Brand IP

Creating an AI digital human that resembles an existing character—whether fictional or a real-world brand mascot—implicates brand IP and copyright. Protecting AI avatar copyright involves ensuring that the visual design does not borrow the 'creative expression' of existing virtual humans. This is particularly relevant for MCNs and brand owners who must ensure their virtual presenters do not cross the line into derivative works of protected character designs.

1.6 Likeness, Voice and Digital Representation

A digital human compliance should not focus only on copyrighted works. Where a digital human reproduces or imitates a real person’s face, body, voice, gestures or performance style, the provider should also assess personality rights, data protection law and commercial likeness rights. In particular, a realistic synthetic representation may function as a sign of the person, even where the scenario is fictional. This means that clearance should cover not only the source image, video or audio file, but also the person’s permission to use their likeness or voice in generative outputs.,

2. Confirm Whether the Materials Can Be Used

Before deploying any AI digital human service in the EU, providers must verify the legality of their data sources. The assumption that data is available for use because it is accessible on the open web is the most common legal error made by technical founders.

2.1 Publicly Available Does Not Mean Free for Commercial Use

The distinction between 'publicly accessible' and 'free for use' is critical for AI digital human copyright compliance. Under the Information Society Directive, any reproduction of a work requires authorization unless an exception applies. The fact that a photograph or video is visible on social media does not grant a license for AI training. SaaS providers must distinguish between materials available for viewing and those licensed for the commercial development of AI models.

2.2 Material Use Should Be Assessed by Purpose

EU law differentiates between research and commercial use. While Article 3 of Directive (EU) 2019/790 (the DSM Directive) provides a mandatory exception for text and data mining (TDM) for scientific research, this does not apply to commercial SaaS companies. For-profit entities must rely on Article 4 of the DSM Directive, which allows TDM for commercial purposes only if the rights holder has not reserved their rights. Understanding the 'purpose' of your data processing is essential for determining which legal exception applies.

2.3 Establishing a Rights Clearance and Provenance Register

To prepare for enterprise procurement and regulatory audits, companies should maintain a Right Clearance and Provenance Register. This register should document not only the origin of training data (images, voice, text) and production materials, but also the legal basis for using each relevant category of rights. For AI digital humans, this should include copyright licences for images, videos, scripts, music and sound recordings; neighbouring rights and performer permissions where applicable; likeness and voice permissions of identifiable persons; TDM opt-out checks; platform terms; and evidence of AI Act marking or provenance controls.

3. Review Copyright Risks in Training, Production and Generation

The training phase is where the most significant copyright risks reside. The EU's 'text and data mining exception EU' framework is sophisticated and requires active monitoring of rights reservations.

3.1 The TDM Exception Should Not Be Treated as Free AI Training

Article 4(3) of the DSM Directive (Directive (EU) 2019/790) allows for the reproduction of lawfully accessible works for the purpose of TDM. However, this is not a 'free pass.' It is a default permission that can be revoked by the rights holder. If you are training AI models to generate digital human content, you must proactively check if the source content is subject to an opt-out. Failure to respect these reservations can invalidate the TDM exception and lead to massive copyright infringement claims.

3.2 Website Terms, Machine-Readable Rights Reservations and Platform Restrictions

Rights holders can reserve their rights through 'machine-readable means' for online content, such as via robots.txt files or metadata. Article 4(3) of the DSM Directive explicitly mentions this. Furthermore, many platforms' Terms of Service specifically prohibit the scraping of data for AI training. EU regulators expect AI providers to respect these technical and legal barriers as part of their 'duty of care' and copyright compliance policy mandated by Article 53(1)(c) of the AI Act (Regulation (EU) 2024/1689) .

3.3 Reviewing Whether the Digital Human Design Borrows from Existing Characters, Virtual Humans or Brand IP

During the production phase, the design of the AI avatar must be checked for 'substantial similarity' to existing IP. If the digital human's appearance is too close to a protected character, it may be deemed an unauthorized derivative work. The CJEU's Painer criteria suggest that if the AI's output is purely the result of an algorithm without human creative control, it might not be copyrightable itself, but it can still infringe upon the copyright of the works it was trained on.

3.4 Clarifying Rights Ownership for Outsourced Design, AI-Generated Content and Stock Materials

Ownership of the final digital human content is often complex. If a SaaS company outsources the design of an avatar, the contract must explicitly transfer copyright. Furthermore, purely AI-generated content may not be eligible for copyright protection in the EU due to the lack of a human author's 'creative choices.' Companies should use a hybrid approach, where human designers make significant creative decisions to ensure the resulting avatar can be legally protected as an intellectual creation.

3.5 Managing Consent and Licences for Likeness and Voice

Where a digital human is modelled on a real person, providers should not rely on general copyright licences alone. A separate likeness and voice release may be needed, particularly where the person’s face, body, voice, or performance style remains identifiable. The agreement should define the permitted use cases, commercial purposes, platforms, territory, duration, modification rights, sublicensing, post-termination use, and whether the person’s likeness or voice may be used for future training, fine-tuning, or synthetic content generation. Providers should also distinguish between transferable copyright licences and personal permissions, which may be subject to stricter limits, especially where personal data or personality rights are involved.

4. Review Output Publication and AI Transparency Obligations

Once the digital human content is generated, its publication triggers specific transparency and copyright rules under the AI Act and existing directives.

4.1 News Facts May Be Reported, but News Texts, Images and Video Footage Should Not Be Copied Directly

While news facts are not protected, the original expression in news reports is. Under Article 5(3)(d) of InfoSoc Directive, quotations are allowed for purposes such as criticism or review, provided the source is indicated and the use is 'to the extent required by the specific purpose.' AI digital humans used in news or commentary must avoid wholesale copying of news texts or footage and instead rely on the quotation exception where appropriate.

4.2 Music, Films, Variety Shows, Sports and News Footage Are High-Risk Materials

Including excerpts from movies, variety shows, or sports broadcasts in AI-generated content is high-risk. These works often have multiple layers of rights (directors, performers, producers). Distribution of such content by a digital human without licensing will likely lead to takedown notices or litigation, as these rights are strictly enforced across the EU.

4.3 Platform Licences Do Not Automatically Cover Cross-Platform Distribution, Advertising or Overseas Publication

A license to use a digital human on one platform (e.g., TikTok) does not necessarily grant the right to use it for paid advertising or on other platforms (e.g., YouTube) unless explicitly stated. Marketing teams must review the scope of their licenses to ensure that their digital human campaigns do not exceed the permitted usage, especially regarding geographical limitations within the EU.

4.4 AI Act Requirements on Synthetic Content Labelling, Deepfake Disclosure and AI-Generated Text Disclosure

Providers should distinguish between deceptive deepfakes and clearly fictional synthetic representations. Deceptive deepfakes raise acute risks because they may mislead viewers about real conduct or events. Clearly fictional digital humans may reduce deception risks, but they do not eliminate personality-rights concerns where the output remains recognizably linked to a real person. For this reason, transparency labelling under the AI Act should be treated as a minimum requirement, not as a substitute for rights clearance.

The AI Act introduces mandatory transparency obligations. Under Article 50(2), providers of AI systems generating synthetic audio, image, video, or text content must ensure the outputs are marked in a machine-readable format. Furthermore, Article 50(4) requires users of deepfakes—AI-generated content that looks like real people or events—to disclose that the content has been artificially generated. This 'deepfake disclosure EU' requirement is mandatory for marketing and content production, with only limited exceptions for artistic or satirical works, provided they do not infringe on the rights and freedoms of third parties.

4.5 Provenance, Reliability and Evidentiary Sensitivity

Beyond copyright and transparency, AI digital human providers should also consider provenance and reliability controls. Synthetic avatars may be used not only in marketing, but also in education, customer support, public communication, or even legally sensitive contexts. Recent debates on AI-generated evidence illustrate a broader regulatory concern: synthetic media must be traceable, explainable, and reliable where it is presented as representing real persons, events, or statements. Providers should therefore maintain logs of source materials, model versions, generation prompts, human review, consent records, and output labelling. These records can support enterprise due diligence and help distinguish authorized synthetic content from misleading or unauthorized deepfakes.

5. Conclusion

Achieving AI digital human copyright compliance requires a dual focus on intellectual property law and the emerging transparency mandates of the AI Act. For AI SaaS companies, the path to EU market entry involves auditing the entire supply chain—from the TDM exceptions used in training to the machine-readable labeling used in output. By respecting Article 4(3) of the DSM Directive and fulfilling the deepfake disclosure duties of Article 50 of the AI Act, providers can build trust with enterprise clients and avoid the significant penalties associated with non-compliance.