Who Owns Copyright in Ai‑generated Works in Belgium? 2026 Guide for Tech Founders & In‑house Counsel

Under current Belgian law, purely AI‑generated works generally do not attract copyright protection because the Belgian originality test, aligned with binding Court of Justice of the European Union (CJEU) jurisprudence, demands a human author’s own intellectual creation. For tech founders and in‑house counsel navigating AI copyright in Belgium in 2026, this single principle shapes every downstream decision about ownership, licensing, enforcement and risk. The landscape is also shifting fast: the EU AI Act is rolling out transparency obligations for general‑purpose AI models, and the Digital Single Market (DSM) Directive review is scheduled for June 2026, potentially reshaping opt‑out rights for rightsholders whose works are used to train models.

This guide delivers the Belgium‑specific, practitioner‑level analysis that founders, startup legal teams and outside IP lawyers need, covering authorship tests, ownership scenarios, contract drafting, training‑data clearance, enforcement strategy and a ready‑to‑use clause bank.

Executive Summary, Short Answer for Founders & Counsel

Under Belgian law as of mid‑2026, an AI system cannot be an author. Copyright vests only in a natural person who makes free and creative choices resulting in an original work. If a human contributes sufficient creative input, through selection, arrangement, curation or significant iterative prompting, a hybrid work may qualify for protection, with the human recognised as author. If the output is generated autonomously by the AI with no meaningful human creative input, the work falls into the public domain.

The practical implications are immediate. Founders and in‑house counsel should take the following actions now:

• Audit every AI‑assisted output to classify it as “human‑led” or “fully autonomous” before claiming or assigning rights.
• Draft explicit assignment and licensing clauses covering AI‑generated outputs, model weights, training datasets and derivative works.
• Include warranty and indemnity provisions addressing the risk that AI outputs may reproduce third‑party copyrighted material.
• Monitor the DSM opt‑out review (June 2026) for changes that may require new licences or contractual triggers for training data.
• Review tax treatment of copyright‑related revenues following 2026 Belgian fiscal amendments to ensure efficient IP structuring.

Can AI Be an Author Under Belgian Law? The Quick Legal Answer

No, Belgian law does not recognise AI as an author. Authorship in Belgium is reserved for natural persons. An AI system lacks both the physical and legal personhood required to hold copyright. This position flows directly from the Belgian Code of Economic Law (Wetboek van economisch recht, Book XI) and from the CJEU’s harmonised originality standard, which Belgian courts are bound to follow.

The CJEU established the decisive test in Infopaq International A/S v Danske Dagblades Forening (C‑5/08): a work is original, and therefore eligible for copyright, only if it is the “author’s own intellectual creation.” The Court elaborated in Painer v Standard VerlagsGmbH (C‑145/10) that originality requires the author to make “free and creative choices” that stamp the work with a “personal touch.” These holdings apply across all EU Member States, including Belgium, and effectively exclude non‑human agents from authorship.

Key Legal Tests for Originality

Belgian courts apply a two‑limb test derived from CJEU case law. First, the work must result from a human intellectual effort, mechanical or random processes do not qualify. Second, the human author must have exercised free and creative choices, meaning the result cannot be dictated entirely by technical constraints, rules or functional requirements. Applied to AI, this means that clicking “generate” on a standard prompt, without further creative intervention, is unlikely to satisfy either limb. Where a human selects, curates, modifies and arranges AI outputs in a way that reflects personal creative judgment, the resulting compilation or adaptation may cross the originality threshold.

How Belgian Courts Have Treated Non‑Human Works

Belgian case law has not yet produced a reported decision squarely addressing AI authorship. However, Belgian academic commentary, including analyses from KU Leuven and Ghent University, consistently concludes that the absence of legal personhood for AI is dispositive: a system that is not a legal subject cannot hold rights. Industry observers expect that any Belgian court confronted with the question would follow the CJEU’s human‑centric originality standard and deny copyright to a purely machine‑generated output. The European Parliament Research Service reached a similar conclusion in its 2025 briefing, noting that the EU “lacks specific rules on the copyrightability of AI‑generated works” but that existing case law “demonstrate[s] a strong need for human creativity.”

When Does a Human or Company Own Copyright in AI‑Assisted Works?

AI copyright ownership in Belgium depends entirely on the nature and extent of human creative contribution. Belgian law does not have a single “work‑for‑hire” doctrine equivalent to the US model; instead, ownership follows the author (the natural person) unless validly assigned or transferred by contract.

The spectrum of AI involvement produces three ownership scenarios:

• Fully autonomous AI output (no human creative choices). No copyright arises. The output is in the public domain and cannot be assigned or licensed as a copyrighted work. Businesses may still protect it through trade‑secret measures, contractual restrictions or database rights, but not copyright.
• Hybrid works (human plus AI). Where a human makes meaningful creative decisions, selecting source material, crafting detailed prompts, iterating on outputs, composing or editing the final work, the human qualifies as author. The resulting work is copyrightable, and rights can be assigned or licensed in the normal way.
• AI‑assisted works where the AI is merely a tool. If the AI functions like a sophisticated instrument (comparable to Photoshop filters or auto‑correct) and the human retains creative control throughout, the work is treated no differently from any other human‑authored work. Copyright vests in the human creator.

Employer–Employee vs Contractor Commissioning

Belgian copyright law provides that the initial owner of a copyright is always the natural‑person author, even in an employment context, unless a valid written transfer or assignment exists. For employees, a contractual clause assigning present and future copyrights to the employer is enforceable if it meets strict formal requirements under Book XI of the Code of Economic Law: the assignment must be explicit, must specify each mode of exploitation, and must be limited in scope. Blanket assignments of “all future works” without specificity risk being unenforceable.

For contractors and commissioned works, the position is even stricter. The commissioner does not automatically receive copyright, a written assignment or licence is essential. Where AI‑generated content in Belgium is produced by freelance developers, prompt engineers or creative agencies, founders must ensure the underlying contracts contain specific IP assignment clauses that cover AI‑assisted outputs and any associated rights in training data or model fine‑tuning.

Practical Examples

• Product descriptions generated by a large language model. If a copywriter feeds detailed brand guidelines and factual inputs, then selects, edits and approves final text, the copywriter is likely the author. Assign rights to the company by written agreement.
• UI images created via image‑generation tools. If a designer provides detailed compositional prompts, iterates on colour and layout, and composites multiple outputs, the designer’s creative choices may qualify the result as a hybrid work.
• Raw model outputs used without human editing. If a chatbot produces customer‑facing text autonomously, with no human review or creative curation, the output likely lacks a copyright‑qualifying author. The company cannot claim AI copyright ownership in this scenario.

Training Data, Model Builders and Copyright Risk

Using copyrighted works to train AI models in Belgium is governed by text and data mining (TDM) exceptions introduced on 1 August 2022. These exceptions, transposing the DSM Directive into Belgian law, permit certain forms of automated analysis of copyrighted content, but with significant conditions and limitations that directly affect copyright for AI models and their commercial deployment.

Text and Data Mining Exceptions in Belgium, What Founders Must Know

Belgian copyright law now contains two distinct TDM exceptions. The first, for scientific research, permits research organisations and cultural heritage institutions to mine lawfully accessed works for non‑commercial scientific purposes. The second, broader exception allows any person to carry out TDM on lawfully accessed works, provided that the rightsholder has not expressly reserved (opted out of) that use. The opt‑out can be implemented through machine‑readable means such as robots.txt directives or metadata tags.

Crucially, these TDM exceptions cover the input side, the act of copying and analysing works to extract data patterns for model training. They do not immunise the output side. If a trained model generates content that reproduces substantial parts of a copyrighted work in its training set, the output itself may constitute infringement regardless of whether the original training was permitted under a TDM exception.

Practical Model‑Training Checklist

• Data audit. Catalogue every dataset used for training; record sources, access dates and any applicable licence terms.
• Opt‑out compliance. Check each source for machine‑readable opt‑out reservations (robots.txt, metadata, terms of service).
• Provenance manifest. Maintain a documented chain of custody for training data to support due‑diligence defences in any future infringement claim.
• Licence gap analysis. Where TDM exceptions do not apply (e.g., commercial use of opted‑out content), negotiate specific training licences with rightsholders.
• Output monitoring. Implement filters or similarity‑detection tools to flag outputs that substantially reproduce known copyrighted works.

Some vendor policies reference informal heuristics, occasionally called a “30% rule”, suggesting that outputs diverging sufficiently from source material are safe. This is not a legal test recognised in Belgium or anywhere in the EU. Belgian courts apply the qualitative CJEU standard (reproduction of original elements), not a quantitative percentage threshold.

EU AI Act and Policy Developments: Implications for AI Copyright in Belgium

The EU AI Act introduces transparency and governance obligations that intersect with copyright strategy for Belgian companies deploying or developing AI. While the AI Act is primarily a product‑safety regulation, several provisions have direct implications for intellectual property management and generative AI copyright compliance.

Providers of general‑purpose AI models face obligations to document and, in some cases, make available sufficiently detailed summaries of the copyrighted training data used. This requirement is designed to enable rightsholders to identify whether their works have been used and to exercise opt‑out rights. For Belgian founders, compliance means building robust data‑provenance records from the earliest stages of model development.

The DSM Directive review, which the European Commission is conducting alongside ongoing AI Act implementation, could alter the balance further. If the review results in strengthened opt‑out mechanisms or mandatory licensing frameworks for training data, Belgian companies that have relied on existing TDM exceptions may need to renegotiate data access and update contractual arrangements.

Industry observers expect the combined effect of EU AI Act copyright obligations and a tightened DSM opt‑out regime to increase compliance costs for model builders operating in Belgium. Founders should build contractual escalation clauses that allow licence renegotiation if the regulatory framework materially changes.

Commercial Options: How to Capture, Assign and Licence Rights in AI Outputs

Even where copyright may not subsist in purely AI‑generated outputs, careful contract drafting remains essential to allocate rights, manage risk and protect commercial value. Belgian law offers several structuring options for founders seeking to capture AI copyright ownership in hybrid works, control downstream use and limit liability exposure.

The core drafting challenges are fourfold. First, the assignment must clearly specify what is being transferred, outputs, model weights, fine‑tuned parameters, datasets, or rights to derivative works. Second, licensing terms must delineate exclusivity, field of use and territorial scope. Third, warranties must address the risk that outputs infringe third‑party copyrights. Fourth, model‑training clauses must define the scope of permission to train on supplied data, audit rights and data‑deletion obligations.

For prompt engineers, creative directors and other contributors who exercise creative judgment, contributor agreements should specify that any copyrightable elements they contribute are assigned to the commissioning entity, with moral‑right waivers to the extent permitted under Belgian law (noting that moral rights in Belgium are inalienable, but the author can agree not to exercise certain moral rights in defined circumstances).

Sample Clause Bank

Clause 1, Assignment of AI‑assisted output rights: “The Contributor hereby assigns to the Company, with full title guarantee, all intellectual property rights (including copyright, database rights and related rights) in and to any Output created by the Contributor using AI tools in the course of performing the Services, for all modes of exploitation, worldwide, for the full duration of such rights.”

Clause 2, Model‑training consent: “The Data Provider grants the Company a non‑exclusive, royalty‑free licence to reproduce and extract data from the Licensed Content solely for the purpose of training, validating and improving the Company’s AI models. The Company shall maintain a provenance log of all Licensed Content used and shall delete or cease use of any content upon written notice of opt‑out by the Data Provider.”

Clause 3, Indemnity with cap: “The Company shall indemnify the Client against third‑party copyright infringement claims arising directly from the AI‑generated Deliverables, subject to an aggregate liability cap equal to the total fees paid under this Agreement in the twelve (12) months preceding the claim. This indemnity shall not apply to claims arising from Client‑supplied inputs or Client modifications to the Deliverables.”

Contracting Checklist Before Product Launch

• Confirm that all contributor and employee agreements contain explicit AI‑output assignment clauses meeting Belgian formal requirements.
• Verify that training‑data licences cover the specific commercial use intended (not just research TDM).
• Include contractual triggers for renegotiation if the DSM opt‑out regime changes.
• Specify warranty scope, distinguish between AI‑generated and human‑authored components.
• Agree on audit rights allowing the licensor or data provider to verify compliance with training restrictions.
• Address moral‑rights provisions: obtain written agreements from human contributors regarding the exercise of moral rights.

Enforcement Risk and Litigation Considerations in Belgium

Copyright enforcement in Belgium for AI‑generated content raises novel evidentiary and procedural challenges. Rightsholders alleging that an AI output infringes their work must demonstrate that the output reproduces original elements of the protected work, the same qualitative test that applies to traditional infringement. Belgian courts can grant injunctive relief, award damages (including lost profits and reasonable royalties) and order seizure of infringing copies under summary proceedings (kortgeding / référé).

Practical Steps After an Infringement Claim

• Preserve evidence immediately. Retain model snapshots, training logs, prompt histories and output versions with timestamps and cryptographic hashes.
• Conduct a similarity analysis. Engage technical experts to compare the allegedly infringing output with the claimant’s work, focusing on qualitative originality rather than percentage overlap.
• Assess TDM defence applicability. Determine whether the training activity fell within a valid TDM exception and whether the rightsholder had opted out.
• Trigger contractual indemnities. Notify upstream model providers or data suppliers under any applicable indemnification clauses.

Defending Claims Based on Generated Output

Defendants in AI‑related copyright disputes in Belgium can raise several defences. Independent creation, demonstrating that the model’s architecture and training process could not have resulted in deliberate copying, is a factual defence that requires detailed technical evidence. The TDM exception may shield the training activity but does not protect infringing outputs. Where the output contains only non‑copyrightable elements (facts, ideas, functional expressions), the defendant can argue that no protectable expression has been taken. Belgian courts also apply proportionality in injunctive relief, meaning that a court may decline a broad injunction where it would disproportionately restrict lawful uses of the AI system.

Tax and IP Strategy Interactions

Belgian tax changes in 2026 affect the fiscal treatment of copyright‑related revenues, with direct implications for how AI output rights are structured. Belgium has historically offered favourable tax treatment for income derived from the transfer or licensing of copyrights and neighbouring rights. Recent amendments have refined the qualifying conditions and caps, making it essential to align IP assignment and licensing structures with current fiscal rules.

Founders assigning or licensing AI‑generated outputs should work with tax counsel to confirm that the income qualifies under the copyright remuneration regime and that the underlying works meet the originality threshold. Where outputs are not copyrightable, because they are fully AI‑generated, the favourable copyright tax treatment will not apply, and alternative structuring (e.g., software licensing, trade‑secret licensing or service‑fee models) should be considered.

Practical Checklist for Founders and In‑House Counsel

The following action items address the most pressing operational and legal requirements for companies developing or deploying AI in Belgium:

1. Classify every AI output as “human‑led” (copyrightable), “hybrid” (potentially copyrightable) or “fully autonomous” (not copyrightable).
2. Update employment contracts with explicit AI‑output assignment clauses that meet Belgian formal requirements for specificity.
3. Review all contractor and freelancer agreements for IP assignment coverage of AI‑assisted deliverables.
4. Audit training datasets, document sources, licence terms, opt‑out status and access dates.
5. Build a provenance manifest for every model deployed commercially.
6. Implement output‑monitoring tools to flag content that may reproduce copyrighted material.
7. Negotiate indemnity clauses with upstream AI providers covering third‑party infringement claims.
8. Add contractual renegotiation triggers tied to the DSM opt‑out review outcome and AI Act milestones.
9. Apply copyright notices only to outputs where a human author can be identified and the originality threshold is met.
10. Coordinate with tax counsel to structure copyright revenue streams in line with 2026 Belgian fiscal rules.
11. Consider professional indemnity insurance covering AI‑related IP infringement exposure.
12. Monitor EU regulatory developments, subscribe to European Commission updates on the AI Act and DSM review.

Recommended Contract Redlines and Model Clause Appendix

Below are five ready‑to‑use redlines that in‑house counsel can adapt for their agreements. Each redline is accompanied by a brief drafting note explaining its purpose under Belgian law.

• Redline 1, Scope of assignment. Replace “all intellectual property” with “all copyright, database rights, neighbouring rights and related rights in and to the Outputs, for all currently known and future modes of exploitation, worldwide, for the full term.” Drafting note: Belgian law requires each mode of exploitation to be specified; blanket language may be unenforceable.
• Redline 2, Licence to use outputs. Insert: “The Licensor grants a [exclusive/non‑exclusive] licence to reproduce, adapt, distribute and communicate to the public the Outputs in the Field of [specify] for [specify term] in [specify territory].” Drafting note: Clarity on field, term and territory avoids disputes over implied licences.
• Redline 3, Model‑training consent with opt‑out mechanism. Insert: “The Data Provider may revoke this training consent by written notice, effective [30/60/90] days after receipt, at which point the Company shall cease using and, if technically feasible, delete the Data Provider’s content from training datasets.” Drafting note: Mirrors the DSM opt‑out principle and future‑proofs the clause.
• Redline 4, Warranty carve‑out for AI outputs. Insert: “The Company warrants that it has implemented reasonable technical measures to prevent the reproduction of third‑party copyrighted content in AI‑generated Deliverables but does not warrant that such Deliverables are free from all third‑party rights.” Drafting note: Allocates risk transparently given the inherent limitations of generative models.
• Redline 5, Moral‑rights acknowledgement. Insert: “The Contributor agrees not to exercise moral rights (including the right of integrity and the right of attribution) in a manner that would prevent the Company from commercially exploiting the Deliverables in the ordinary course of business.” Drafting note: Moral rights are inalienable under Belgian law, but an agreement not to exercise them in specific circumstances is generally considered permissible.

Conclusion

AI copyright in Belgium follows a clear principle: only humans can be authors, and only works reflecting human creative choices attract protection. For founders and in‑house counsel, the practical consequence is that every AI‑assisted workflow must be assessed for the presence and sufficiency of human creative input. Contract drafting, precise assignments, robust indemnities, well‑scoped training consents and DSM‑linked renegotiation triggers, is the primary tool for managing AI copyright ownership risk. With the EU AI Act imposing new transparency obligations and the DSM opt‑out review poised to reshape training‑data rules, 2026 is the year to operationalise compliance.

Sources

1. European Parliament Research Service, Copyright of AI‑generated works: Approaches in the EU
2. Universiteit Gent, National and international legislation regarding AI
3. KU Leuven, Generative AI and Copyright Law
4. CURIA, Court of Justice of the European Union (case law search)
5. Belgian Official Gazette / Federal e‑justice portal (Moniteur Belge)
6. European Commission, European approach to artificial intelligence