Artificial Intelligence 2026 Comparisons

Germany has no standalone national AI law. The EU AI Act applies directly, with national implementing legislation expected to be finalised during 2026. The following areas of general law remain central.

Contract Law

Procuring AI services raises questions about output quality and service levels; under German law, fault is presumed once a contractual breach is established. Contracting through AI – where a system concludes agreements autonomously – raises attribution questions that are largely unresolved, particularly for agentic AI systems that independently negotiate and enter contracts on a user’s behalf.

Tort and Product Liability

Establishing defects, causation and damages is complicated by AI opacity. The revised EU Product Liability Directive, in force since December 2024 and to be transposed by December 2026, explicitly includes software and AI as products, enabling strict liability claims. Germany published a draft implementing act (ProdHaftG-E) in September 2025. Agentic AI poses additional evidential challenges given the distance between design decisions and autonomous harmful actions.

Privacy and Data Protection

The General Data Protection Regulation (GDPR) and Germany’s Federal Data Protection Act (Bundesdatenschutzgesetz – BDSG) remain the primary frameworks. Key issues include the legal basis for AI training on personal data, deletion and correction obligations within trained models, and the application of Article 22 GDPR to automated decision-making. Agentic AI compounds these challenges through continuous, multi-step data collection, engaging purpose limitation and retention obligations.

Intellectual Property (IP)

Generative AI is the focal point. Training on copyrighted material, IP compliance during inference, and the protectability of AI-generated outputs remain contested. The EU text and data mining exception provides some basis for training use, but its limits continue to be tested. Liability across the value chain for outputs reproducing protected works is unsettled.

Consumer Protection

AI consumer products attract documentation, transparency and update obligations under German law. Defining what constitutes a “defect” in an AI system – and when an update is legally required – remains unclear. The EU AI Act’s prohibited practices, including social scoring, apply since February 2025. Agentic AI acting autonomously on consumers’ behalf raises additional concerns around informed consent and user control.

Employment Law

Works councils hold co-determination rights over AI tools monitoring employee behaviour. AI systems used in recruitment, performance assessment and monitoring are classified as high risk under the EU AI Act, with full compliance required from August 2026. Employers remain the primary liable party and must ensure human oversight, employee training, and regular bias audits under the Act on Equal Treatment (AGG).

Criminal Law

German criminal law’s focus on individual fault and foreseeability sits uneasily with autonomous AI behaviour. Debates centre on the standard of care, permissible risk and attribution of harm. These challenges are sharpest for agentic AI, where the causal distance between human decision-making and harmful outcomes is greatest – an area likely requiring legislative or judicial clarification.

AI continues to revolutionise German industries. Predictive machine learning remains indispensable for preventing machine failures in manufacturing, optimising data centre cooling, and detecting fraud. Healthcare relies heavily on it for diagnostics.

Generative AI accelerates code creation, drives customer service chatbots, and generates marketing content.

Retrieval-augmented generation (RAG) is the established enterprise standard to mitigate hallucination risks, frequently deployed on premises due to strict data protection standards. By 2026, focus has expanded to agentic AI systems.

In Germany’s manufacturing sector, multi-agent systems autonomously monitor supply chains and execute alternate production routings in real time. Telecommunications firms use voice agents for technical troubleshooting. These “systems of action” challenge existing compliance frameworks regarding traceability and human oversight.

Germany supports AI innovation through public funding, infrastructure investment and strategic programming. The AI market is forecast to grow from around EUR9 billion in 2025 to EUR37 billion by 2031.

National Strategy and Funding

A EUR5.5 billion policy initiative funds next-generation AI models, increased computing capacity and data infrastructure with a focus on industrial applications. Flagship knowledge transfer projects covering key industries – automotive, chemicals, biotechnology, cleantech, medicine and agrifood – are being launched from 2026 onwards.

Infrastructure

Germany adopted a National Data Centre Strategy in March 2026, targeting a doubling of overall data centre capacity and a fourfold increase in AI and high-performance computing capacity by 2030, with data centres designated as critical infrastructure. A planned “AI gigafactory” with 100,000 high-performance GPUs forms part of this offensive.

SME and Start-Up Support

Dedicated programmes such as KI4KMU provide grants for SMEs to pilot and adopt AI. The German AI start-up scene grew 35% year on year in 2024, with over 40% of new companies being scientific spin-offs. The German Research Centre for Artificial Intelligence (DFKI) anchors research excellence nationally and internationally.

Private and EU-Level Investment

Public initiatives are complemented by major private commitments – Google announced EUR5.5 billion in German infrastructure investment for 2026–2029, including new AI-focused data centres. At EU level, Germany participates in EUREKA clusters and GAIA-X, while national implementation of the EU AI Act – including supervisory authority designation – is expected to be finalised during 2026.

Germany’s regulatory philosophy relies entirely on the EU AI Act rather than creating national AI-specific legislation. The overarching EU framework dictates a capabilities- and risk-based approach, avoiding technology-specific rules for architectures such as agentic AI. Germany deliberately avoids national “gold-plating”, applying EU life cycle rules uniformly.

Regarding implementation, prohibitions and general-purpose AI (GPAI) rules are applicable. The market is preparing for the August 2026 high-risk system obligations, although an ongoing EU “Omnibus” procedure may delay this timeline. Nationally, enforcement structures are established via the AI Market Surveillance and Innovation Promotion Act (KI-MIG).

The substantive AI legislation in Germany is the EU AI Act. It imposes stringent requirements for high-risk AI systems, including mandatory risk assessments and human oversight. GPAI and foundation models are governed by Title VIII. Nationally, Germany focuses solely on administrative enforcement via the AI Market Surveillance and Innovation Promotion Act (KI-MIG). Progressing through parliament in spring 2026, the KI-MIG formally designates the Federal Network Agency (BNetzA) as the primary market surveillance authority and establishes a co-ordination and competence centre (KoKIVO). Formal enforcement actions remain limited until high-risk obligations fully apply.

Several public bodies have issued non-binding guidance to help organisations navigate AI responsibly in Germany.

Federal Administration Guidelines

In March 2025, the Federal Ministry of the Interior published guidelines for the use of AI within the federal administration, setting out rules that federal authorities must observe when deploying AI in their work. This is the most significant domestic soft-law development for public sector AI use to date.

BNetzA Guidance and AI Service Desk

The Federal Network Agency (BNetzA), set to become Germany’s central AI market surveillance authority, published a guidance paper on AI literacy requirements under Article 4 of the EU AI Act in June 2025, and launched an AI Service Desk in July 2025 as a low-threshold point of contact for regulatory questions.

Data Protection Authorities

In December 2025, the BfDI (the German federal data protection regulator) issued guidance for public authorities aimed at helping them identify data protection issues when using AI tools – in particular large language models (LLMs) – with the goal of developing a structured approach to AI projects and reducing legal uncertainty. Earlier in 2025, the BfDI also launched a public consultation on the data-protection-compliant handling of personal data in AI models, focusing on memorisation risks in LLMs. The German Data Protection Authorities’ (DPAs) 2024 position paper on generative AI and data protection remains relevant as a broader framework. Additionally, the Data Protection Conference published guidance on the data protection implications of generative AI systems using the RAG method.

Outlook

Further guidance is expected as Germany finalises its AI implementing legislation (KI-MIG) during 2026, with the BNetzA expected to issue additional sector-facing guidance alongside its market surveillance role.

Regarding EU AI Act implementation, Germany’s administrative focus is on testing and certification infrastructure. Under the KI-MIG framework, the BNetzA acts as the national notifying authority. It is responsible for formally designating “Notified Bodies” (such as technical inspection agencies like the TÜV or DEKRA) to conduct third-party conformity assessments for certain high-risk AI systems. High-risk systems will be assessed against harmonised CEN/CENELEC standards currently undergoing EU-level finalisation. Additionally, German authorities are structuring the mandated national regulatory sandboxes to foster innovation and facilitate compliance testing.

This topic is not applicable.

Copyright/Content Law

To implement Articles 3 and 4 of the EU’s Digital Single Market Directive, the German government introduced Section 44b and amended Section 60d of the German Copyright Act (UrhG) on text and data mining (TDM). These rules allow AI developers to scrape data from the internet for training under specific conditions:

• the data must be lawfully accessible; and
• the rights-holder must not have opted out in an appropriate manner (eg, in a machine-readable format such as the robots.txt file).

The Higher Regional Court of Hamburg, in its December 2025 Kneschke v LAION ruling, expressly confirmed that the general TDM exception under Section 44b UrhG (Article 4 CDSM Directive) covers the training of generative AI models, referencing Article 53 EU AI Act as legislative confirmation that the Union legislature presupposes the exception’s applicability to GPAI providers. On the opt-out requirement under Section 44b(3) UrhG, the court held that a reservation expressed in natural language (here, in the stock agency’s ToS) was not machine-readable in 2021, but adopted a technology-neutral, dynamic standard: whether a plain-text opt-out qualifies as machine-readable depends on the state of reliably available technology at the time of use, leaving room for natural-language reservations to suffice today or in the future.

Data Protection

The GDPR establishes strict requirements for the collection and use of personal data from the internet to train AI models. The German DPAs have not eased the interpretation of the GDPR to facilitate AI training on personal data. The European Data Protection Board’s (EDPB) opinion on AI models from December 2024 also did not ease requirements.

Germany has not proposed any national AI-specific legislation. The legislative focus remains exclusively on establishing the administrative framework for the EU AI Act via the KI-MIG. At the EU level, the Commission has refrained from proposing new regulations for agentic AI, viewing the EU AI Act as sufficient. Regarding supply chain accountability, obligations are managed through broader frameworks such as the Corporate Sustainability Due Diligence Directive (CSDDD). Following the withdrawal of the AI Liability Directive in 2025, no replacement proposal has been introduced, leaving liability to general tort law and the Product Liability Directive.

In Germany, the first copyright-related decisions concerning generative AI were handed down in recent years. Higher court rulings and established case law on copyright and generative AI are currently being established.

CJEU (EuGH), Court Docket No C-250/25 – Like Company v Google (Pending)

The CJEU addresses four questions on generative AI:

• whether LLM training is unauthorised reproduction;
• whether the TDM exception applies;
• whether generating identical text is a “communication to the public”; and
• who is liable for user-prompted outputs.

Court of Appeal Hamburg, Court Docket No 5 U 104/24 – Kneschke v LAION

The court held that downloading images to build the LAION-5B training dataset is permitted under both the scientific research and general TDM exceptions. The plaintiff’s natural-language opt-out, buried in the Terms of Service, was invalid: it did not meet the statutory “machine-readable” standard, judged against the state of the art at the time of download in 2021.

District Court Munich, Court Docket No 42 O 14139/24 – GEMA v OpenAI

In a landmark first-instance ruling, the court held that ChatGPT’s reproduction of protected song lyrics constitutes unauthorised copyright infringement. The court rejected the TDM exception, arguing that the AI “memorised” rather than merely analysed the texts, marking a significant initial test of how standard copyright applies to generative AI outputs and model weights in Germany.

Higher Regional Court Düsseldorf, Court Docket No I-20 W 2/26

In a preliminary injunction over an AI-generated comic-style variation of a copyrighted dog photograph, the court denied relief. AI outputs can qualify for copyright where human input “shapes” them enough to reduce the AI to a mere tool, applying the CJEU’s Mio standard of “free and creative choices” in prompting, refinement or post-processing. On infringement, the court replaced the overall-impression test with element-by-element comparison: the variation kept only the unprotected motif (dog in a given pose) while transforming the protected elements – perspective, bokeh, lighting, textures – into a comic aesthetic.

Federal Patent Court, 11 November 2021, 11 W (Pat) 5/21

Only natural persons can be inventors, so AI cannot be an inventor under German patent law.

Federal Court of Justice, X ZB 5/22 (June 2024)

AI-generated inventions can be protected under patent law if a human contributor is named as the inventor. Listing AI itself as the inventor is not allowed, but naming a person who influenced the AI system is sufficient.

Privacy

Higher Regional Court Cologne, 15 UKl 2/25 (Meta case)

In April 2025, Meta announced that it would begin training its AI on public first-party data from Facebook and Instagram users in the EU. The German consumer association Verbraucherzentrale NRW sought an emergency injunction before the Higher Regional Court of Cologne, alleging violations of the GDPR and the Digital Markets Act (DMA). On 23 May 2025, the court dismissed the injunction in its entirety, ruling that Meta’s approach complied with both instruments, in line with the Irish DPC’s regulatory assessment and the EDPB’s December 2024 AI Opinion. The decision confirmed that AI training on public user data can be based on legitimate interests under Article 6(1)(f) GDPR without requiring user consent, provided GDPR requirements are carefully examined and appropriate safeguards are in place. The Hamburg DPA, which participated in the proceedings, raised reservations about the adequacy of Meta’s de-identification measures – these are likely to remain relevant in the main proceedings, as the injunction ruling is not the final word on the merits.

CJEU, 7 December 2023, C-634/21 (SCHUFA case)

Article 22 GDPR prohibits automated analysis of data if the result decides whether a contract is made, executed or cancelled, unless data controllers can rely on limited justifications such as consent and contractual necessity.

Employment

Labour Court Hamburg, 16 January 2024, 24 BVGa 1/24

Employers generally do not need works council consent to allow employees the optional use of AI tools with private accounts, where employees were previously authorised by the employer to use the tool.

Germany has not designated a single national “AI regulator”; instead, oversight is distributed among multiple agencies/authorities. Under KI-MIG, which is progressing through parliament in spring 2026, the BNetzA is designated as the central market surveillance and notifying authority. The BNetzA has already established an AI Service Desk and will operate a co-ordination and competence centre (KoKIVO) to support other authorities. The BNetzA also serves in parallel as the national authority responsible for implementing the Digital Services Act, which likewise intersects with AI-related issues such as dark patterns and deepfakes.

The DPAs remain critical de facto AI regulators. Because AI systems inherently process personal data, the DPAs actively enforce GDPR compliance. The KI-MIG requires close co-ordination between the BNetzA and the DPAs to prevent regulatory overlap.

Germany operates a decentralised regulatory landscape for AI, with oversight distributed across multiple authorities depending on the context of use.

DPAs

German DPAs remain the most prolific source of AI guidance. Their 2024 position paper on generative AI and GDPR compliance is the baseline framework, expressing scepticism about many generative AI deployments while stopping short of prohibition. In December 2025, the BfDI issued guidance for public authorities on identifying data protection issues when using AI tools – particularly LLMs – aimed at supporting a structured compliance approach. The Data Protection Conference also published guidance on the data protection implications of generative AI systems using the RAG method.

BaFin – Financial Sector

On 18 December 2025, BaFin issued guidance on managing ICT risks when using AI in financial entities, intended to help firms implement DORA requirements. It is particularly aimed at banks under the Capital Requirements Regulation and insurers under Solvency II. BaFin described the document as a “living” framework intended to evolve alongside technological and regulatory developments.

BNetzA

The BNetzA published guidance on AI literacy requirements in June 2025 and launched an AI Service Desk in July 2025. Once formally designated as Germany’s central AI market surveillance authority under the pending KI-MIG legislation, broader operational guidance is expected. A co-ordination centre (KoKIVO) is also planned within the BNetzA to support sector-specific authorities in their AI Act tasks.

BSI – Cybersecurity

The BSI has issued an AI Cloud Services Compliance Catalogue and highlighted generative AI cybersecurity risks. Its role was strengthened by the NIS2 Implementation Act, which came into force in December 2025, giving the BSI extended supervisory powers over AI-related cybersecurity requirements.

ChatGPT/OpenAI

Several German state DPAs – including those in North Rhine-Westphalia, Baden-Württemberg, Hesse, Rhineland-Palatinate and Schleswig-Holstein – launched co-ordinated investigations into ChatGPT’s data processing practices in 2023, focusing on legal basis, transparency, processing of minors’ data, and data accuracy. No final German enforcement decision has been issued. When OpenAI established its European headquarters in Ireland in February 2024, the GDPR one-stop-shop mechanism was triggered, and the EDPB ChatGPT Taskforce acknowledged that, for continuing infringements, pending proceedings should be transferred to the Irish Data Protection Commission as lead supervisory authority – effectively absorbing the German national investigations.

Meta AI Training

Shortly before Meta began AI training on EU user data in May 2025, the Hamburg DPA initiated urgent proceedings against Meta, intending to prohibit AI training on German data subjects for at least a further three months. The Hamburg DPA questioned whether Meta’s processing of such large volumes of data was necessary, and expressed scepticism about the adequacy of Meta’s de-identification measures as a risk-mitigating step. The outcome of those proceedings, and any follow-on regulatory action, remains to be seen.

DeepSeek

German regional DPAs began co-ordinating on DeepSeek’s data processing practices in early 2025, particularly regarding its storage of user data in China and apparent lack of GDPR compliance infrastructure. No formal German enforcement decision had been concluded at the time of writing.

EU AI Act Enforcement

Formal enforcement under the EU AI Act has not yet commenced in Germany. Prohibition provisions have applied since February 2025, but the national implementing legislation (KI-MIG) remains in draft. Most substantive obligations for high-risk AI systems apply from August 2026 (subject to the Omnibus trilogue), meaning that material AI Act enforcement is still on the horizon.

Enforcement Trends

German DPAs have been active investigators but have yet to issue a concluded enforcement decision against an AI provider. The one-stop-shop mechanism has been a structural constraint, channelling major cases involving global AI providers to the Irish Data Protection Commission (DPC). Given the limited resources relative to the pace of AI deployment, the DPAs are expected to rely in part on consumer groups and media reports to identify violations going forward.

In Germany, standard setting emphasises adapting to EU frameworks rather than creating isolated national rules. The German Institute for Standardisation (DIN) and the German Commission for Electrical, Electronic and Information Technologies (DKE) lead these efforts. Their primary role is channelling German industry interests into the European CEN/CENELEC harmonised standards mandated by the EU AI Act. Domestically, DIN facilitates early-stage sector frameworks and actively adopts international standards, such as ISO/IEC 42001, helping local companies operationalise compliance ahead of the AI Act’s 2026 deadlines.

International standards are critical for operationalising compliance in Germany while EU harmonised standards remain under development. German enterprises increasingly adopt ISO/IEC 42001 (AI Management Systems) alongside risk management standards to establish auditable governance. While international standards do not independently grant a legal “presumption of conformity” under EU law, they deliberately align with EU requirements to prevent conflicts. For transatlantic operations, the NIST AI Risk Management Framework is also frequently utilised.

Modernisation Agenda

The German government is expanding AI use across public administration under Chancellor Merz’s “Modernisation Agenda”, targeting a 25% cut in bureaucracy costs by 2029 and deploying AI for visa reviews and export documentation. The Federal Ministry of the Interior published guidelines for AI use within the federal administration in March 2025.

Public Administration

The federal AI portal KIPITZ is being significantly expanded: in the second half of 2026, the Digital Ministry plans to extend access to state and municipal authorities via the Deutschland-Stack and a sovereign AI cloud. A planned KIPITZ 2.0 will introduce an app-store model for approved AI functionalities and workflows, including agentic AI capabilities. At municipal level, cities such as Cologne and Munich are already testing AI for traffic management and waste collection.

Judiciary

Germany has mandated electronic file management across all court proceedings by 2026, providing the data infrastructure for AI-assisted case processing. Operational tools are already in use: Baden-Württemberg deploys OLGA for case categorisation, and the Frankfurt District Court has piloted Frauke to assist in drafting repetitive judgments. Widespread deployment remains pending, with human oversight maintained throughout.

Legal Requirements

Government AI systems must comply with the GDPR, the federal administration guidelines, and the EU AI Act. Many public sector deployments – particularly in visa processing, benefits and law enforcement – will qualify as high risk under the AI Act, triggering documentation and human oversight obligations from August 2026. Real-time remote biometric identification in public spaces remains broadly prohibited.

Federal Constitutional Court – Police Data Analysis (2023)

The leading case remains the Federal Constitutional Court’s decision of 16 February 2023 (1 BvR 1547/19, 1 BvR 2634/20), which held that statutory authorisations for automated police data analysis in Hesse and Hamburg were unconstitutional. The court confirmed that automated state data analysis interferes with citizens’ right to informational self-determination – a fundamental right under German law. The broader the analytical capabilities of a system, the heavier the legislative justification required. This remains the constitutional benchmark for government AI deployments involving personal data.

Regional Court of Darmstadt – AI Expert Reports (2025)

On 10 November 2025, the Regional Court of Darmstadt held that, if a court-appointed expert relies extensively on AI when preparing a report without disclosing it, their fee can be reduced to zero and the report rendered inadmissible. The Code of Civil Procedure requires experts to prepare reports personally; undisclosed AI use violates this requirement. This has direct implications wherever expert reports inform administrative or judicial decisions.

Outlook

The EU AI Act classifies AI systems assisting judicial authorities as high risk, requiring risk management for bias and opacity, with final decision-making remaining human-driven. These obligations apply from August 2026 (Omnibus pending) and are likely to prompt further judicial scrutiny of government AI use.

The EU AI Act’s national security exception (Article 2(3)) excludes AI used exclusively for military or defence purposes from its scope. Consequently, Germany’s use of AI in national security remains governed by general constitutional principles and sector-specific laws rather than AI-specific statutes.

Under the Zeitenwende policy, the Bundeswehr has accelerated AI integration, particularly in sensor-to-shooter systems and cyber defence. Germany strictly adheres to the NATO AI Strategy, focusing on “meaningful human control” to prevent fully autonomous lethal systems. While the planned KI-MIG establishes surveillance for dual-use and civilian AI, military applications remain under the exclusive jurisdiction of the Ministry of Defence, prioritising ethical frameworks over rigid civilian regulations.

The emergence of generative AI technologies raises new legal complexities under German and EU law. (See 16. Intellectual Property and 17. Data Protection for detailed analysis.)

Foundation Models/GPAI

As of August 2025, Title VIII obligations for GPAI models are fully applicable. Providers must maintain technical documentation and provide detailed summaries of training data. Models presenting systemic risks face heightened duties, including adversarial testing and incident reporting. In Germany, the KoKIVO (within the BNetzA) serves as the technical competence hub supporting the enforcement of these requirements.

Copyright in Training Data

The central flashpoint is the scope of the TDM exceptions. Courts confirm that Section 44b UrhG covers generative AI training (pending final CJEU ruling); the battleground has shifted to the technical details of TDM, in particular the form requirements for an effective opt-out. On the output side, courts increasingly treat memorisation as a copyright-relevant act outside Section 44b UrhG (see 4.1 Precedent-Setting Judicial Decisions).

Copyright in Outputs

AI outputs lack IP protection unless the deployer sufficiently controls the generative process. Under the CJEU’s Mio standard (C-580/23 & C-795/23), German courts require the human’s “free and creative choices” to dominate the result so that the AI functions as a mere tool – exercised in prompting, iterative refinement or post-processing.

Data Protection

The GDPR and generative AI are generally compatible but create specific difficulties: data subject rights (rectification/erasure in AI models due to the “black box effect”), purpose limitation, and legal basis for training on personal data. German DPAs have issued comprehensive recommendations on generative AI and data protection (May 2024). In December 2025, the BfDI issued guidance for public authorities on identifying data protection issues when using AI tools, particularly LLMs, aimed at supporting a structured compliance approach. The Data Protection Conference also published guidance on the data protection implications of generative AI systems using the RAG method.

Transparency

The EU AI Act sets transparency requirements for AI systems directly interacting with persons (disclosure of non-human nature), synthetic content (machine-readable marking) and deepfakes (specific disclosure mandate).

Contractual and Liability Issues

Generative AI outputs in services and AI service procurement require new contractual frameworks. The revised Product Liability Directive now applies to AI-based products. The AI Liability Directive was withdrawn in February 2025.

The legal profession has universally adopted Generative AI, rendering traditional rule-based tools obsolete. Law firms and legal departments are currently transitioning into the “agentic era”. While initial AI adoption focused on reactive, prompt-and-response LLMs, emerging agentic systems can autonomously orchestrate and execute complex, multi-step legal workflows, such as comprehensive due diligence or sophisticated contract analysis.

A notable market tension currently exists between bespoke, law-specific AI platforms and powerful general-purpose foundation models from major AI labs, which increasingly offer robust legal functionalities. The market is fragmented, and it remains to be seen which approach will ultimately dominate.

Regarding professional law, the adoption of cloud-based AI tools is permissible but requires careful structural compliance. The key regulatory anchor in Germany is Section 43e(1) of the Federal Lawyers’ Act (BRAO), which governs the use of external IT service providers and safeguards client confidentiality. While AI significantly augments capabilities, lawyers retain ultimate accountability for all submissions, necessitating strict human oversight to mitigate hallucination risks and ensure ethical compliance.

Under German law, as AI itself is not a legal person, liability for damages caused by AI systems must be attributed to the operator or others in the supply chain.

Liability Through Contract Law

Proving breach of duty and causality can be challenging, especially when the inner workings of an AI system are not accessible (black box problem).

Product Liability

Claims face difficulties due to the complexity and opacity of AI systems, including establishing a defect, damage and causal link. The revised EU Product Liability Directive (in force December 2024, transposition deadline December 2026) now addresses AI-based products, introducing enhanced disclosure powers and updated burden-of-proof rules to address information asymmetries. Germany published a draft implementing act (ProdHaftG-E) in September 2025.

Tort Liability

Hindered by the lack of regulatory rules for AI safety, complexities arise in proving fault and causation, and challenges in assessing non-human AI systems. The AI Liability Directive, which would have introduced fault-based tort liability rules with presumptions for AI victims, was withdrawn by the European Commission in February 2025. This moderates the overall impact of EU liability reform – with no dedicated EU-level rules specific to AI fault-based claims, supply chain actors remain subject primarily to the existing strict liability approach under the revised Product Liability Directive.

Insurability

The blurring of the line between human and machine behaviour makes it challenging to allocate responsibility and determine insurability. This has sparked debate on the need for separate AI insurance.

There are no local German governmental initiatives specifically addressing AI liability. The regulatory landscape is shaped primarily by EU-level action.

The revised EU Product Liability Directive entered into force on 9 December 2024 (the German draft implementing act (ProdHaftG-E) was published in September 2025). It maintains strict liability for manufacturers, holding them responsible for harm caused by defective products, including AI-based products. Key changes address information asymmetries through enhanced disclosure powers and altered burden of proof, including presumptions of evidence.

The AI Liability Directive proposed by the EU Commission was withdrawn in February 2025 in a rare move to reduce regulatory density. Critics argue that this undermines adequate protection for victims of AI-related harm. It remains unclear whether the Commission will introduce an updated proposal.

As a result, the liability framework in Germany and the EU will see relatively incremental change – primarily through the revised Product Liability Directive and existing national law.

AI Act Regulatory Framework and Governance

The EU AI Act does not establish a separate legal category for “agentic AI” or autonomous agents. Instead, they are regulated based on their technical components, typically qualifying as an “AI system” (Article 3(1)) that integrates a GPAI model (Article 3(63)). Consequently, standard AI Act rules apply.

Prohibitions and high-risk domains

Agentic systems must incorporate design safeguards against prohibited practices, particularly harmful manipulation or exploitation of vulnerabilities (Article 5). If an agent operates in a high-risk domain (eg, employment or critical infrastructure), it is subject to strict Chapter III requirements from August 2026.

Systemic risk in multi-agent systems

For the underlying GPAI models, advanced capabilities such as tool use and high autonomy are decisive factors for being designated as models with “systemic risk”. Providers of these models face strict risk management obligations that specifically address autonomous capabilities and complex, multi-agent configurations.

GDPR Framework and Governance

Under German law, the primary applicable privacy framework for agentic AI is the GDPR and BDSG. Agentic AI is particularly challenging from a data protection perspective because its defining characteristics – autonomy, open-ended task execution, multi-system access, and unpredictable data flows – sit in direct tension with the GDPR’s foundational principles of purpose limitation, data minimisation, transparency and human accountability.

Organisational responsibility

Agentic AI systems have no legal personality; organisations deploying them remain fully responsible for GDPR compliance regardless of the system’s level of autonomy. Governance, oversight and accountability mechanisms must be in place even where agents act independently.

Controller/processor roles

Determining controller and processor roles is particularly challenging in multi-vendor agentic ecosystems. Organisations must clarify accountability across every layer of the agentic chain, with contracts allocating responsibilities for security and data rights compliance.

Purpose limitation and data minimisation

Open-ended agent tasks create pressure to define purposes too broadly. Organisations must establish specific purposes for each processing activity and apply least privilege, granting access only to data necessary for defined tasks.

Automated decision-making

Where agentic systems make decisions with legal or similarly significant effects, organisations must inform affected individuals, enable them to contest decisions, and ensure meaningful human intervention. Article 22 GDPR applies in full.

Special category data and accuracy

Agents pursuing open-ended goals may incidentally infer special category data, triggering Article 9 GDPR requirements. Hallucinations can cascade across tools and agents, compounding accuracy risks at scale – a compliance challenge that has no clear solution under existing GDPR frameworks.

Logging and oversight

Organisations may need standalone monitoring systems to log, interpret and intervene in agent activity to demonstrate accountability under Article 5(2) GDPR.

Intellectual Property

Agentic AI poses a distinct copyright challenge: a categorical shift to autonomous systems that plan, act and iterate with minimal human input. This may collide with the CJEU’s Mio/Konektra standard, requiring human “free and creative choices” to dominate the output, with the AI as a mere tool. The more agentic the deployment, the more choices are delegated to the machine, and the less likely outputs qualify as protected works. Narrow exceptions remain where agentic AI is tightly constrained – eg, deterministic reproduction of a pre-cleared corpus.

As set out previously, German liability law relies on contractual and tort frameworks not designed for autonomous AI. The revised EU Product Liability Directive extends strict liability to software and AI by December 2026.

Allocation Between Developers, Deployers and Users

Developers bear primary exposure for defective system design; deployers are liable for inadequate oversight, while users may bear responsibility where they act on obviously flawed outputs without review. For agentic AI, this allocation is particularly difficult: where an agent acts outside the scope of its instructions, attributing liability along this chain is largely uncharted territory under German laws.

Evidentiary Challenges

The complexity, autonomy and opacity of AI systems make it difficult and expensive for victims to identify the liable party and prove causation. These challenges are acute for agentic systems, which may take sequences of autonomous decisions across multiple tools before harm materialises. The EU AI Act’s logging requirements for high-risk systems will assist claimants from August 2026. Disclosure rights to access that evidence are, however, limited: under German law, they will exist going forward primarily in the context of product liability claims under the revised EU Product Liability Directive and its German implementing act, but not for fault-based tort claims more generally.

Contractual Liability Allocation

Developer contracts typically limit liability for model outputs heavily, leaving deployers exposed. For agentic AI specifically, contracts should address autonomous action scope, escalation triggers, and indemnification for harms caused by agent decisions outside explicitly authorised parameters.

Cascading Failures in Multi-Agent Systems

Where autonomous agents interact across organisational boundaries, identifying which agent caused harm and which party is responsible becomes highly complex. No specific framework addresses cascading failures; liability will be assessed under general causation principles, with joint and several liability potentially applicable. This remains one of the most significant unresolved areas of AI liability law in Germany and across the EU.

Germany’s primary tools for addressing algorithmic discrimination are the AGG, the GDPR and – going forward – the EU AI Act. The AGG applies where algorithms produce discriminatory outcomes on protected grounds such as race, gender or disability, including in automated recruitment decisions. Under the GDPR, algorithmic bias is relevant in two ways: Article 22 provides protection against fully automated decisions with significant effects, requiring human oversight and the right to contest outcomes; and the fairness principle under Article 5(1)(a) GDPR independently requires that personal data is not processed in a manner that is unjustifiably detrimental or discriminatory to individuals – a requirement that applies to any AI system processing personal data, regardless of whether a fully automated decision is involved.

The EU AI Act introduces the most substantive bias obligations to date. AI systems used in recruitment, performance assessment and credit decisions are classified as high risk, requiring bias testing, data governance and human oversight from August 2026 – subject to possible delay under the Digital Omnibus proposal. Harmonised technical standards for bias testing remain under development.

No concluded German enforcement actions on algorithmic bias have been published. Individuals face significant challenges in proving bias, resulting in a lack of case law on compensation claims. More active enforcement is expected once the BNetzA assumes its market surveillance role. Industry best practice includes algorithmic impact assessments, regular bias audits, representative training data, and documented human review for high-stakes decisions.

The EU AI Act distinguishes between “post” and “live” biometric identification methods – each associated with different levels of risk and regulatory requirements.

Post-biometric identification is classified as a high-risk application under the EU AI Act, requiring a comprehensive set of regulatory requirements. The only exception is biometric verification used solely to confirm an individual’s claimed identity.

Live biometric identification faces a general prohibition, especially when applied in real-time in publicly accessible spaces for law enforcement purposes. Exceptions are narrowly defined:

• search and rescue operations targeting victims of serious crimes;
• preventing imminent threats to safety or foreseeable terrorist attacks; and
• localising individuals suspected of serious criminal offences.

Emotion Recognition

The EU AI Act contains specific prohibitions on AI systems that infer emotions of natural persons in workplace and education contexts, and on AI systems used for biometric categorisation to deduce sensitive characteristics. Under the GDPR, biometric data is treated as sensitive special category data (Article 9), requiring explicit consent or another specific legal basis.

Code of Practice (Article 50 AI Act)

Providers of AI systems generating synthetic media must mark outputs in machine-readable form; deployers must disclose deepfakes and AI-generated content. The implementing Code of Practice is in its second draft (5 March 2026), with finalisation expected in early June 2026, ahead of Article 50’s entry into application on 2 August 2026. Because no single marking technique meets all four statutory requirements (effectiveness, interoperability, robustness, reliability), providers must combine tamper-evident provenance metadata under open standards (eg, C2PA) with robust, imperceptible watermarks that persist when metadata is stripped. A voluntary common “AI” icon for deployers is envisaged, with a later interactive second-layer version.

Platform Liability (DSA)

Very Large Online Platforms (VLOPs) must proactively assess and mitigate systemic risks including manipulative deepfakes (Articles 34, 35), comply with ad-transparency duties (Article 26) and accept user challenges to moderation decisions via certified out-of-court bodies (Article 21).

German Law (German Unfair Competition Act (UWG) and Criminal)

Misleading AI-generated advertising is unlawful under Sections 5, 5a UWG. Malicious deepfakes are prosecuted as insult (Section 185 StGB), defamation (Section 186) and incitement (Section 130), with a 2026 push to explicitly criminalise deepfake pornography.

Civil Remedies (Germany)

Victims can rely on the constitutional right of personality to obtain preliminary injunctions for immediate removal (Section 1004 BGB analog) and sue creators or deployers for damages and pain-and-suffering compensation (Section 823 BGB).

The EU AI Act and German law set strict transparency and disclosure requirements across different risk levels.

Direct Interaction

AI systems such as chatbots must disclose their non-human nature to users unless evident under normal circumstances (with narrow safe-harbour exceptions, such as authorised law enforcement use).

Synthetic Content and Deepfakes

Providers must mark outputs in a machine-readable format to distinguish AI-generated from human-created media. Deployers of deepfakes face specific disclosure mandates. As of early 2026, the second draft of the Transparency Codes of Practice is under review. The final version, expected shortly, will provide the essential technical and practical framework for operationalising these complex labelling and watermarking requirements.

High-Risk Systems

Providers must ensure explainability by supplying downstream deployers with detailed documentation regarding system capabilities, limitations and proper usage.

Foundation Models (GPAI)

Transparency obligations are fully applicable as of August 2025. GPAI providers must publish technical documentation and detailed summaries of the training data used.

Manipulative Practices

AI systems deploying subliminal techniques or exploiting vulnerabilities are prohibited under the AI Act (effective since February 2025). Furthermore, under the UWG, using AI for aggressive commercial practices or emotional manipulation is unlawful.

While AI is predominantly procured via Software as a Service (SaaS) – sharing similarities with traditional cloud agreements regarding availability and support – the rise of generative and autonomous agentic AI systems necessitates highly tailored contractual frameworks.

SLAs and Performance

Traditional uptime metrics are no longer sufficient. Contracts increasingly incorporate AI-specific SLAs addressing model drift, hallucination rates and, particularly for agentic systems, the accuracy of autonomous task execution.

Data Rights and IP Allocation

A critical negotiation point is strictly restricting providers from using customer data to train broad foundation models. Contracts must explicitly allocate IP rights for generated outputs and determine ownership or licensing of fine-tuned model weights.

Liability and Indemnities

With agentic AI executing autonomous actions, risk allocation is complex. Robust indemnification clauses for third-party IP infringement are standard, while liability caps are heavily negotiated to reflect the unpredictable nature of AI outputs.

Audit and Exit

Comprehensive audit rights are essential to verify data usage and regulatory compliance. Exit provisions must securely address the portability of fine-tuned models or mandate the strict deletion of vector databases.

Despite these evolving practices, a universal market standard has not yet solidified, requiring bespoke legal solutions.

EU AI Act Value Chain Framework

Accountability frameworks for AI supply chains are primarily driven by the EU AI Act’s value chain provisions. A critical risk in procurement is the shift in responsibility: if a downstream operator substantially modifies a high-risk AI system or places it on the market under its own name, it assumes the strict legal obligations of the original provider.

Consequently, procuring AI systems requires rigorous due diligence and the contractual cascading of obligations. To manage liability for third-party AI components, companies must implement mandatory flow-down clauses, strict indemnities and comprehensive audit rights. Furthermore, upstream providers are legally required to supply downstream operators with sufficient technical documentation, logging capabilities and transparency regarding system provenance to enable compliance at the deployment level.

Data Protection

From a GDPR perspective, supply chain due diligence requires mapping controller and processor roles across all layers of the AI supply chain and ensuring that data processing agreements are in place with all processors. Where an AI system processes personal data, the deployer as controller remains fully responsible for GDPR compliance regardless of upstream provider conduct, including in respect of training data provenance and data governance practices.

AI technologies have a profound impact on personnel decisions, offering advantages in processing large amounts of data quickly for tasks such as pre-selecting job applicants and creating scoring tables for employee dismissals and performance reviews.

The GDPR restricts exclusively automated decisions in employment relationships. Decisions with legal implications (hiring, transfers, dismissals) should generally involve human review unless the narrow justifications under Article 22 GDPR can be complied with.

Pre-selection and support measures utilising AI are permissible but require careful examination. AI can be used in generating job descriptions, pre-selections, reference letters and relocation support.

Discrimination Risks

Employers can be held liable under the AGG if they inadequately programme AI systems, use flawed data, or neglect regular quality checks. Liability applies regardless of whether the tools are internal or external. Indirect discrimination can occur when seemingly neutral criteria favour certain employee groups in practice.

Works Council Rights

Depending on the specific AI tools and set-up, works councils have significant co-determination rights, and detailed works agreements must be negotiated with employee representatives.

Performance evaluation using AI tools promises greater objectivity and efficiency. Tools include performance review systems, individual/group work activity analysis, and automated review of data and processes (eg, travel cost reimbursements).

Monitoring

Monitoring employees is subject to strict conditions based on case law. General total surveillance of employees without cause (including covert video or audio surveillance) is not allowed. Exceptions are limited to specific cases and suspicions.

Preventative or support measures are permissible as long as they do not create undue surveillance pressure. However, these principles may conflict with new technologies such as voice-based live evaluation of calls and transcription tool reviews.

Employers must ensure through regular checks and compliance systems that employees use AI in a safe and compliant way. When processing individual (log) data with AI tools, works councils have significant co-determination rights, and detailed works agreements should be negotiated.

Discrimination Risks

There is risk of violation of the law if programming and/or output is inadequate. An open feedback culture to improve systems used is advisable.

EU Digital Services Act

Article 27 requires plain-language disclosure of recommender parameters and user-modifiable options. Article 14 may impose transparency duties for restrictions on the use of AI-generated content, while Article 26 governs labelling and targeting disclosures for AI-generated ads. No Commission guidance yet exists on AI-driven dark patterns under Article 25.

Enforcement is accelerating: since the Delegated Act on Data Access took effect on 29 October 2025, vetted researchers – accredited in Germany by the Bundesnetzagentur – can request VLOP data under Article 40 for systemic-risk research.

EU AI Act

Since February 2025, Article 5 prohibitions apply directly to platforms: Article 5(1)(a) and (b) ban subliminal or manipulative AI techniques that distort behaviour and cause significant harm, overlapping with but extending beyond the DSA’s Article 25, with penalties up to 7% of global turnover. Interface manipulations such as dynamic pricing, adaptive nudges, or AI-generated scarcity cues targeting vulnerable users may be assessed under both regimes in parallel.

From August 2026, Article 50 requires providers of generative AI to mark outputs in machine-readable form (eg, C2PA) and deployers to disclose deepfakes and certain AI-generated public-interest text – but only where the platform itself provides or deploys the system, not for mere distribution of third-party AI content.

Competition Law

The Bundeskartellamt scrutinises platform AI conduct, which enables extended abuse control against companies with paramount cross-market significance, with recent enforcement against price-control mechanisms and anti-competitive practices in automotive services and maps.

Financial services companies increasingly rely on AI, subject to stringent and evolving regulatory expectations.

DORA and ICT Risk Management

DORA directly governs AI deployments. In early 2026, BaFin published crucial guidance (Orientierungshilfe) confirming that AI systems are formally classified as ICT assets under DORA. Consequently, financial institutions must fully integrate AI into their operational resilience and ICT risk management frameworks. This necessitates strict safeguards against cyber-risks, data poisoning, model drift and vendor lock-in when procuring cloud-based AI services.

EU AI Act and Credit Decisioning

The EU AI Act explicitly classifies AI systems used to evaluate creditworthiness or establish credit scores as “high risk” (Annex III). Deploying these models requires rigorous compliance structures. Furthermore, AI-driven credit decisioning must strictly adhere to Article 22 of the GDPR, requiring human-in-the-loop mechanisms and guaranteeing the right to explanation for automated individual decision-making.

AI in healthcare in Germany is governed by the GDPR, the EU AI Act and – for AI as a medical device – the Medical Device Regulation (MDR) and In Vitro Diagnostic Regulation (IVDR). Health data constitutes special category data under Article 9 GDPR, requiring explicit consent or another specific legal basis.

EU AI Act Classification

Under Annex III, AI systems deciding access to health services, assessing health insurance risk and pricing, and used in emergencies are classified as high risk, subject to obligations on risk management, data quality, transparency and human oversight from August 2026. AI used as a medical device or safety component is separately classified as high risk under Article 6(1).

MDR/IVDR and AI: Dual Framework

AI medical devices (MDAI) face dual compliance obligations under both the AI Act and the MDR/IVDR. A medical device is automatically a high-risk AI system under the AI Act where a notified body is involved in its MDR conformity assessment, broadly covering class IIa devices and above. In June 2025, the Medical Device Coordination Group published MDCG 2025-6, the first guidance document clarifying the interplay between the AI Act and the MDR/IVDR for MDAI systems. Harmonisation remains challenging given differing classification criteria, with particular open questions for self-learning AI systems.

In December 2025, the European Commission published a proposal that, if adopted, would harmonise AI Act requirements for medical devices within the MDR/IVDR framework and remove medical devices from the AI Act’s high-risk scope – potentially finalised by 2026 or 2027.

Germany-Specific Developments

The BfArM has been designated to operate a central data access office under the Health Data Use Act (2024), mediating access to statutory health insurance data for public-interest research under pseudonymisation safeguards. The BfArM and Federal Ministry of Health are actively engaged at EU level on MDAI classification guidance. No concluded enforcement actions specifically targeting healthcare AI in Germany have been published.

National Autonomous Driving Regulation

Germany remains a pioneer under its Autonomous Driving Act (2021), permitting SAE Level 4 vehicles on public roads within designated areas without a human driver, provided that a remote technical supervisor can intervene. While Levels 1 and 2 follow standard traffic laws, SAE Level 5 (unrestricted automation) remains prohibited. In the event of accidents, strict liability applies to car owners under the German Road Traffic Act, though apportioning product liability for AI failures remains legally complex.

EU AI Act Integration

AI systems acting as safety components in vehicles are classified as “high risk” under the EU AI Act (Article 6(1) and Annex I, Section B). However, pursuant to Article 2(2), the AI Act’s obligations do not apply directly to these systems. Instead, these requirements must “flow down” and be integrated into existing sectoral legislation, specifically the EU Motor Vehicle Type Approval Regulation. As of early 2026, the European Commission has not yet adopted the necessary delegated acts, meaning that direct AI-specific obligations have not yet materialised in the automotive sector.

Market Trends

Two AI trends are reshaping retail: GenAI for marketing content (text, image, video), and agentic commerce/smart retail – autonomous sales agents handling advice, upsell and after-sales, plus Smart Mirrors for in-store visualisation.

Core Legal Risks

The integration sits at the intersection of consumer protection, unfair competition, IP and AI-specific rules.

Liability and IP

Open questions remain on liability for hallucinated promises by autonomous agents, and on third-party IP infringement in and protectability of GenAI marketing content.

Consumer protection

AI products fall under German consumer law, triggering documentation, transparency and warranty obligations.

Unfair competition

AI-generated ads and recommendations must not mislead on price, availability or product features (Section 5, 5a UWG).

EU AI Act

Customer-facing AI (eg, chatbots) may trigger Article 50 transparency duties.

Germany’s manufacturing sector is a leading adopter of industrial AI, with applications across assembly, logistics, autonomous mobile robots (AMRs) and collaborative robots (cobots).

Regulatory Frameworks

Two EU-level frameworks apply in parallel. The EU Machinery Regulation (EU) 2023/1230 applies from January 2027, replacing the current Machinery Directive. It explicitly addresses AI and robotics, requiring risk assessments covering the potential evolution of AI behaviour in self-learning systems, documentation of data governance and training data for traceability, and extending health and safety concerns to include psychological stress from human-robot interaction. The EU AI Act applies alongside it: where an AI system is a safety component of a machine requiring third-party conformity assessment, it is automatically classified as high risk.

Dual Compliance and AI Act Interplay

Unlike medical devices, no proposal exists to carve industrial robotics out of the AI Act’s framework. Cross-regulatory compliance therefore remains unavoidable, with high-risk requirements for systems under Annex I of the AI Act applying from August 2027. Political discussions on the AI Omnibus are ongoing: while a large majority of EU member states oppose moving Annex I Section A items to Section B, there is growing openness to clarifying Article 6 high-risk classification specifically in relation to the Machinery Regulation, potentially reducing compliance burdens for lower-risk industrial applications. Trilogue negotiations on this point remain active at the time of writing.

No concluded German enforcement actions specifically targeting industrial AI or robotics have been published. Workplace safety obligations under the German Occupational Health and Safety Act (ArbSchG) apply alongside EU-level requirements.

The AI technology itself can be protected under several IP regimes, though recent case law and the shift towards agentic systems have begun to reshape where commercial value actually resides.

Software/Copyright

The AI model and AI algorithm may be protected as software under Section 69a UrhG (transposing the Software Directive 2009/24/EC), and training datasets may be protected as databases under Section 87a UrhG where they reflect a substantial investment.

Training Data

Individual training data is often protected by third-party copyright (texts, images, audio, video). Its use for AI training must be justified either by a licence or by one of the TDM exceptions under Sections 44b, 60d UrhG. For more details, see 4.1 Precedent-Setting Judicial Decisions.

Inputs/Prompts

Typical short prompts are generally too simple or too technically constrained to meet the originality threshold. More detailed prompts exercising genuine creative discretion may, in principle, qualify.

Outputs

AI outputs may lack IP protection unless the deployer’s creative decisions sufficiently control the generative process. Under the CJEU’s Mio standard, an output qualifies as a protected work only where the human’s “free and creative choices” dominate its final form so that the AI functions as a mere tool – exercisable before (prompting), during (iterative refinement) or after (post-processing) generation (see 4.1 Precedent-Setting Judicial Decisions and 8. Generative AI).

Trade Secrets

Where copyright is unavailable or unreliable, Section 2 No 1 GeschGehG (transposing the Trade Secrets Directive EU 2016/943) offers an alternative layer of protection, provided the technology is kept confidential, holds commercial value and is subject to appropriate non-disclosure measures, and the owner has a legitimate interest in non-disclosure.

Only natural persons qualify under German law (Section 7 UrhG; Federal Patent Court, 11 W (pat) 5/21). No jurisdiction has yet accepted an AI as inventor or author.

Case Law

In DABUS (X ZB 5/22, June 2024), the Federal Court of Justice held that AI-generated inventions remain patentable provided a natural person is named – it suffices to name a human who influenced the AI through problem definition, supervision or evaluation. WIPO’s taxonomy (AI models, AI-assisted, AI-based, AI-generated) places only the first three within reach of inventorship.

Moral Rights

The inventor’s right to be named (Section 63 PatG) is inalienable and attaches only to natural persons. The same holds for copyright moral rights under Sections 12–14 UrhG (attribution, integrity, first publication).

Reform Debates

Active threads include:

• statutory clarification of the AI-assisted versus AI-generated line along WIPO categories;
• a sui generis neighbouring right for AI outputs without human authorship (critics warn of public-domain erosion);
• disclosure duties on AI involvement, driven partly by the AI Act’s transparency architecture; and
• potential IP refinements via the proposed EU AI Digital Omnibus.

Germany’s central flashpoint for generative AI is the scope of the TDM exceptions. The TDM exemption may justify use of copyrighted works for AI training where the data is lawfully accessible and no effective opt-out exists. The Higher Regional Court of Hamburg (Kneschke v LAION, 5 U 104/24) confirmed coverage of generative AI training; the battleground has shifted to opt-out form, with natural-language reservations in Terms of Service not automatically meeting the “machine-readable” standard (see 4.1 Precedent-Setting Judicial Decisions). The regulatory layer mirrors this:

• Article 53(1)(c) AI Act requires GPAI providers to maintain a copyright-compliance policy respecting Article 4(3) CDSM opt-outs, concretised by the 2025 GPAI Code of Practice; and
• Article 53(1)(d) adds a duty to publish a detailed training-content summary on the AI Office template.

On outputs, the exception does not necessarily shield providers where protected works resurface: the District Court of Munich treats memorisation – verbatim or near-verbatim reproduction – as a copyright-relevant act outside Section 44b UrhG (see 4.1 Precedent-Setting Judicial Decisions and 8. Generative AI).

General Principle – German Case Law and Literature Converge

AI output is not protected where input is thin (eg, a short prompt) and the AI retains significant creative leeway. Under the CJEU’s Mio/Konektra standard, protection requires the human’s “free and creative choices” to dominate the result – exercisable before, during or after generation (see 4.1 Precedent-Setting Judicial Decisions). Where it qualifies, copyright vests in the natural person making the decisive contribution.

Moral Rights

These are reserved to natural-person authors (see 16.2 AI as Inventor/Author).

Practical Guidance

Businesses should:

• document the creative process (prompting, refinement, post-processing) for key assets;
• calibrate AI use to protection needs, confining AI to assistive roles where copyright is commercially essential; and
• adapt contracts, as full-transfer clauses may be structurally inadequate for AI-assisted deliverables.

Licensing Models

Foundation models are distributed across a spectrum – eg:

• proprietary API access (model never leaves the vendor);
• open-weights with restrictions (such as Meta Llama, with thresholds for commercial scale); and
• permissive open-source (Apache 2.0, MIT).

API Versus Self-Hosted

API users hold only a contractual licence to use outputs, with provider retention and audit rights, and inputs leave their infrastructure. Self-hosting offers confidentiality and independence but shifts compliance and security duties to the user.

Fine-Tuning

Adapting a base model to narrower tasks generally triggers the base licence – including licence inheritance, attribution (eg, “Built with Llama”) and bans on training competing models.

Enforceability

Open-source licences may operate through copyright: breach of a condition forfeits the permission and exposes the user to infringement claims.

Provider Terms

Key clauses to watch are acceptable-use policies and thresholds above which a separate commercial licence is required.

German DPAs have not relaxed their interpretation of the GDPR to facilitate AI training. The EDPB’s December 2024 Opinion 28/2024 provided some clarity but did not materially ease the compliance burden.

Lawful Basis

The EDPB Opinion confirmed that legitimate interests under Article 6(1)(f) GDPR can serve as a valid legal basis for AI model training and deployment. However, controllers must conduct a rigorous three-step assessment:

• identifying a specific and real interest;
• demonstrating necessity; and
• balancing against data subjects’ rights.

Consent is theoretically available but practically infeasible for large-scale training in most cases.

Purpose Limitation and Data Minimisation

Training data must be compatible with its original collection purpose; where data is repurposed, a compatibility assessment is required. German regulators acknowledge the tension between strict data minimisation and the volume of data needed for effective AI, but have not formally relaxed the principle.

Anonymisation and Data Subject Rights

The EDPB set a high bar for anonymity: a model is only anonymous if personal data cannot be identified from it or extracted through queries. This must be assessed case by case. The practical difficulty of responding to deletion and correction requests within trained models remains largely unresolved; output filtering is increasingly preferred over full retraining, though German DPAs have not confirmed its acceptability.

Special Category Data and Unlawful Training

Article 9 GDPR applies where training data contains or can infer sensitive data, regardless of whether that data was publicly available. Where a model was developed using unlawfully processed personal data, this can impact the lawfulness of its subsequent deployment unless the model has been duly anonymised. Deployers must therefore verify the data provenance of third-party models as part of their procurement due diligence.

Lawful Basis and Transparency

Controllers deploying AI systems must identify a valid legal basis for all processing. German DPAs require controllers to ensure that a legal basis is in place, referring to the Baden-Württemberg DPA’s paper on legal bases for AI as practical guidance. Articles 13 and 14 GDPR require clear information about automated processing logic in accessible language.

Automated Decision-Making (Article 22 GDPR)

German DPAs interpret Article 22 strictly: AI-driven decisions must not be solely automated, and effective human oversight must be ensured. The CJEU’s SCHUFA judgment (C-634/21) confirmed that Article 22 applies where an AI output decisively influences a final decision even if a human formally takes it. Controllers must enable data subjects to contest decisions and obtain human review.

Data Subject Rights

Fulfilling erasure and rectification rights within deployed AI systems is technically complex. The German DPA guidance recognises machine unlearning and input/output filters as possible interim solutions, but notes that full compliance may require model retraining or redevelopment.

Special Category Data and Children

Where AI systems process or infer special category data, Article 9 GDPR requires an explicit legal basis and elevated technical safeguards. For children’s data, the age of digital consent in Germany is 16 under Article 8 GDPR; AI systems accessible to minors require age verification and enhanced protections.

Accountability

German DPA’s updated TOM guidance structures accountability across four AI life cycle phases – design, development, implementation and operation – aligned with data protection goals including transparency, data minimisation, intervenability and unlinkability. DPIAs are required for high-risk AI processing and should be embedded in procurement workflows from the outset.

DPIAs

German DPAs generally expect a DPIA to be carried out before deploying AI applications, given the high risk to data subjects that AI processing typically entails.

Controller/Processor Roles

Where a company deploys a third-party AI application, the provider typically acts as processor; where an AI is trained on multiple datasets or developed jointly, organisations may qualify as joint controllers requiring an Article 26 GDPR agreement. The AI Act’s “provider” and “deployer” concepts do not map directly onto GDPR roles – a separate GDPR role analysis is always required.

Cross-Border Transfers

SCCs remain the primary transfer mechanism for AI-related personal data flows to non-EEA countries. Transfers to US AI and cloud providers are generally currently covered by the EU-US Data Privacy Framework adequacy decision, though this remains susceptible to legal challenge. German DPAs expect transfer impact assessments to be conducted and documented, and take a cautious approach to US cloud transfers generally.

Merger Control and AI Partnerships

German and EU regulators increasingly scrutinise not only traditional acquisitions but also “quasi-mergers” in the foundation model market – such as exclusive compute-for-equity partnerships or large-scale acquihires. In Germany, transactions that fall below the ordinary turnover thresholds may still be notifiable if the transaction value exceeds EUR400 million and the target has substantial activities in Germany. Separately, following a sector inquiry, the Bundeskartellamt may impose a notification obligation under Section 32f(2) GWB for future acquisitions in the affected sector.

Data-Driven Market Power and Tying

A core concern is the vertical integration of cloud infrastructure with proprietary AI models. Regulators are actively monitoring dominant tech firms for tying, exclusive dealing and self-preferencing within their digital ecosystems. Germany addresses this aggressively via Section 19a GWB, a tool enabling the Bundeskartellamt to prohibit anti-competitive practices by digital platforms designated as having “paramount significance across markets” (eg, Google, Microsoft).

Algorithmic Collusion

While no AI-specific cartel legislation exists, traditional antitrust rules apply fully to AI-driven price-fixing. Companies cannot evade liability by delegating pricing decisions to autonomous algorithms. Authorities treat algorithmic co-ordination – including tacit collusion via shared third-party pricing models – as an unlawful concerted practice.

The intersection of AI and cybersecurity is now strictly regulated. In December 2025, Germany’s NIS-2 Implementation Act (NIS2UmsCG) formally entered into force. It vastly expands the number of entities obligated to implement comprehensive cyber-risk management. Crucially, NIS-2 mandates stringent supply chain security and secure software development life cycles, directly impacting the procurement and deployment of AI systems. In the event of significant cyber-incidents, organisations face strict reporting obligations, requiring notification to the Federal Office for Information Security (BSI) within 24 hours.

Simultaneously, the EU AI Act demands that high-risk AI systems ensure robustness against adversarial attacks, such as data poisoning. The BSI continuously highlights the dual nature of AI: while generative AI accelerates sophisticated cyber-attacks (eg, malware and automated phishing), deploying AI-driven defensive solutions is strongly recommended to identify and mitigate these threats effectively.

In Germany, the Corporate Sustainability Reporting Directive (CSRD) mandates strict ESG-related disclosures, and companies are increasingly using AI tools to manage compliance. To curb AI’s environmental footprint, the Energy Efficiency Act (EnEfG) sets demanding PUE and waste-heat-reuse targets for data centres starting from 1 July 2026. The EU AI Act complements this by requiring GPAI model providers to document the known or estimated energy consumption of their models.

For social and governance aspects, some firms are adopting internal AI-governance boards and bias-mitigation practices as best practice, while the Corporate Sustainability Due Diligence Directive (CSDDD) and German Supply Chain Act (LkSG) require comprehensive ESG due diligence across supply chains, including when procuring AI systems and services.

AI governance in Germany has transitioned from voluntary best practice to mandatory compliance under the EU AI Act. Practical implementation requires integrating AI oversight into existing corporate governance and risk management structures rather than creating parallel silos.

Governance Structure and Inventory

A foundation of any compliant governance framework is a comprehensive AI inventory covering internal builds, vendor-embedded AI, and informal deployments. Most enterprises still lack a complete inventory of AI systems across business units. Accurate inventory enables risk classification under the EU AI Act and ensures that third-party systems undergo rigorous vendor due diligence before procurement. Roles must be clearly mapped – provider, modifier or deployer – as this determines the applicable obligations and compliance deadlines.

Risk Management and Impact Assessments

Organisations must implement continuous risk management systems addressing data quality, model drift and bias throughout the AI life cycle. For high-risk deployments – particularly in the public sector and for private entities conducting credit scoring, insurance risk assessment or employment decisions – the AI Act mandates Fundamental Rights Impact Assessments (FRIAs). DPIAs under the GDPR apply in parallel wherever personal data is processed. These assessments should be integrated into existing compliance workflows rather than conducted separately. 

Third-Party AI Governance

Vendor governance is a critical pressure point: enterprises must ensure that suppliers can produce technical documentation and evidence of conformity assessment. Contracts must include flow-down of AI Act obligations, change notification requirements, and audit rights. Substantial modification of a third-party system by the deployer can trigger a shift to full provider obligations.

Incident Response and Monitoring

Organisations must establish rapid incident response protocols for AI failures, including serious incident-reporting obligations to authorities under the AI Act for high-risk systems. Post-market monitoring – with defined thresholds for performance drift, bias signals and safety incidents – must be operational before deployment, not retrofitted afterwards.

Proportionate Governance and Implementation Challenges

Governance should be proportionate to risk: lightweight documentation for minimal-risk tools, full quality management systems for high-risk deployments. Standardised frameworks such as ISO/IEC 42001 provide a practical structure. Key implementation challenges include:

• establishing cross-functional accountability across legal, technical and business teams;
• building AI literacy across the workforce; and
• managing the pace of AI deployment against the availability of harmonised technical standards, which remain incomplete.