The Paris Call on AI Safety: From Voluntary Commitments to Contractual Risk

Complex law. Clear action.

Reviewed by Oleksandr Sobovyi, Founder & CEO of CORVUS AI — editorial responsibility statement below.

On 9 November 2026, ten developers of advanced AI models, including OpenAI, Anthropic, Microsoft and Mistral AI, endorsed commitments to introduce fail-safe mechanisms capable of restricting or shutting down model operations, including during training. The Paris Call for AI Safety and Security is not legally binding. Yet for businesses integrating these models into their products and operations, its significance may extend well beyond a political declaration: it could influence supplier contracts, operational risk management and the allocation of liability.

What Was Signed — and by Whom

The Paris Call for AI Safety and Security was published by the Élysée Palace on 9 November 2026.

As of 11 November, the declaration had been endorsed by 16 countries: Austria, Bahrain, Canada, Croatia, Cyprus, Estonia, Finland, France, Ireland, Kenya, Latvia, Liechtenstein, Luxembourg, the Netherlands, Norway and Portugal.

The corporate signatories are Anthropic, Black Forest Labs, Cohere, Domyn, Hugging Face, Microsoft, Mistral AI, OpenAI, Sakana AI and Synthesia. Seven research and expert organisations have also joined, including LawZero.

The list of those who have not signed is equally revealing.

The United States, the United Kingdom, China and Japan are absent from the published list of signatories. Google, Meta and xAI are not among the participating companies. Neither Germany nor the European Commission has signed, although both supported the September Call for Control of Frontier AI Models, which informed the Paris initiative.

The absence of a signature does not necessarily indicate disagreement with the declaration's principles.

However, the European Commission and the AI Office have not undertaken commitments under this document. The Paris Call should therefore not be presented as an official EU position or a binding instrument for interpreting the AI Act.

The list of signatories remains open.

Three Strategic Commitments, Five Practical Elements

The Paris Call is structured around three main commitments: risk assessment and management, accountability across the AI ecosystem, and information sharing and early warning.

Five practical elements emerge from these commitments.

1. A common framework for assessing and managing frontier-model risks. Particular attention is given to cyber and chemical, biological, radiological and nuclear (CBRN) threats, increasing model autonomy, misalignment and loss of control.

2. Publication of evaluation results. Participants intend to publish non-sensitive findings in a standardised format and establish a shared, publicly accessible repository. This could include red-teaming protocols, benchmarking tools, risk assessments and stress-test results.

3. Developer accountability and fail-safe mechanisms. Developers commit to disclosing model capabilities, limitations and known risks; supporting independent evaluation where appropriate and proportionate; and implementing mechanisms to restrict, pause or shut down model operations within their control, including during training and internal use. Policymakers also recognise the need to clarify legal responsibility for large-scale harm, including cross-border damage.

4. Incident reporting and early warning. The declaration envisages more consistent incident classifications, timely information sharing concerning serious cyber incidents and AI control incidents, and stronger arrangements for identifying emerging risks.

5. Whistleblower protection. Participating governments commit to supporting secure and confidential reporting channels through which employees and other relevant parties can raise concerns about systemic risks without fear of legal or professional retaliation.

The declaration does not establish specific implementation deadlines, its own penalties or a separate enforcement authority. It does not expressly refer to the EU AI Act.

What Is Already Binding Under EU Law?

Several issues addressed by the Paris Call are already covered by binding EU legislation. However, the applicable obligations depend on the type of AI model, the nature of the risk and the role of the relevant operator.

Since 2 August 2025, providers of general-purpose AI models with systemic risk have been subject to additional obligations under Article 55 of Regulation (EU) 2024/1689 — the AI Act.

These include:

• conducting model evaluations, including adversarial testing;

• assessing and mitigating systemic risks;

• documenting serious incidents and reporting relevant information without undue delay to the AI Office and, where applicable, national competent authorities;

• ensuring an adequate level of cybersecurity for the model and its supporting infrastructure.

The penalty regime under Article 101 has applied since 2 August 2026. For relevant infringements committed intentionally or negligently, the Commission may impose fines of up to 3% of the provider's worldwide annual turnover in the preceding financial year or €15 million, whichever is higher.

A transitional provision is important here. Under Article 111(3) of the AI Act, providers of GPAI models placed on the market before 2 August 2025 have until 2 August 2027 to bring those models into compliance.

Whistleblower protection is also addressed by existing EU law.

Article 87 of the AI Act extends the framework established by Directive (EU) 2019/1937 to reports concerning infringements of the Regulation.

The Paris Call goes further in one respect: it addresses reporting of systemic risks that may not yet constitute an established breach of law.

An important distinction concerns shutdown mechanisms.

The AI Act does not impose a general requirement on providers of GPAI models to stop training or internal operations in the manner contemplated by the Paris Call.

Article 14(4)(e) of the AI Act requires high-risk AI systems to enable human intervention or safe interruption of their operation.

That is a different obligation. It concerns human oversight of high-risk AI systems, rather than shutdown procedures for the internal operations of frontier-model developers.

The relevant requirements for high-risk AI systems are scheduled to apply from 2 December 2027 or 2 August 2028, depending on the category of system.

The Paris Call therefore neither replaces the AI Act nor automatically extends its legal obligations. Instead, it provides an additional reference point for assessing risk-management practices and developer accountability.

The CORVUS AI Perspective: Where Expectations Meet Commercial Reality

Non-binding does not mean commercially irrelevant.

The Paris Call does not itself create new contractual obligations for customers of its corporate signatories.

However, the commitments made by developers may influence model terms of use, service-level agreements and incident-response procedures.

Providers may, for example, clarify the circumstances in which access can be restricted or suspended, notification arrangements and the allocation of liability.

For businesses whose critical operations depend on a single AI model, this is a reason to review existing supplier contracts and business continuity arrangements.

At present, this is a foreseeable contractual risk rather than an established market practice.

The argument that a risk was unknown requires evidence.

Directive (EU) 2024/2853 on liability for defective products extends the relevant product liability framework to software.

Following the official corrigendum of 7 May 2026, its temporal scope covers products placed on the market or put into service on or after 9 December 2026. Member States must transpose the Directive by that date.

Under Article 11(1)(e), an economic operator may, subject to specific conditions, be exempt from liability if it demonstrates that the objective state of scientific and technical knowledge did not allow the defect to be discovered. Article 18 permits certain national derogations from this defence.

Public commitments by leading AI developers could become relevant evidence of the state of scientific and technical knowledge and of risks known within the industry.

However, acknowledging the risk of losing control of an AI model does not, by itself, establish that a particular product is defective, that a causal link exists or that compensable damage has occurred.

The potential influence of the Paris Call on the application of Articles 7 and 11 of the Directive is CORVUS AI's legal assessment, not an established position in case law.

The new Directive is also not a general mechanism for recovering every commercial loss resulting from the suspension of an AI service.

Contractual liability and liability for defective products must be assessed separately.

The Netherlands is already preparing for the new liability regime.

On 10 November 2026, the Dutch Senate (Eerste Kamer) approved Bill No. 36.906, which implements the revised Product Liability Directive through amendments to the Dutch Civil Code.

As of the date of publication, the Bill has passed both chambers of Parliament. Its entry into force remains subject to completion of the applicable national procedures.

For businesses bringing AI-enabled products to the Dutch market, this provides an additional reason to review the allocation of liability between model developers, system integrators and manufacturers of final products.

The defence exclusion does not override supplier contracts.

Under Article 2(3) of the AI Act, AI systems placed on the market, put into service or used exclusively for military, defence or national security purposes fall outside the Regulation's scope in the circumstances specified by that provision.

However, exclusion from the AI Act does not remove contractual restrictions imposed by a commercial AI-model provider.

For dual-use solutions, the applicability of the Regulation must be assessed separately for the relevant uses, together with the compatibility of supplier terms and obligations owed to the end customer.

This is particularly relevant to EU–Ukraine projects in which commercial AI components may be integrated into products serving both civilian and defence markets.

Limitations of This Analysis

• The list of signatories to the Paris Call remains open. The information presented here is current as of 11 November 2026.

• The declaration does not establish detailed legal definitions of a serious AI incident or an AI control incident. Their relationship with the definition of a serious incident under Article 3(49) of the AI Act requires further assessment.

• The declaration does not specify implementation deadlines or technical criteria for fail-safe shutdown mechanisms.

• Its potential influence on commercial contracts and future case law is a professional assessment, not an established legal consequence.

What Businesses Should Do Now

1. Map your AI dependencies. Identify the models supporting your products and business processes, and determine whether their providers have endorsed the Paris Call.

2. Review supplier contracts. Examine unilateral suspension rights, notification periods, service availability commitments, liability exclusions and termination provisions.

3. Protect business continuity. Assess whether critical operations can be transferred to an alternative model or performed manually, and determine acceptable recovery times.

4. Assess product liability exposure. Where an AI component forms part of a product to be placed on the EU market or put into service on or after 9 December 2026, assess the applicability of the revised Product Liability Directive, documentation of known risks and the allocation of liability across the supply chain.

5. Review defence and dual-use contracts. Where a solution relies on a commercial AI model, compare the model provider's restrictions with the end customer's requirements concerning availability, information security, export controls and operational continuity.

When is a formal legal review necessary?

If suspension of access to an AI model could prevent the performance of a contract with a public authority or defence customer, the entire contractual chain should be reviewed before the next agreement is signed.

Disclaimer

This article has been prepared by CORVUS AI for general informational and educational purposes only. It is intended to make complex legal and regulatory developments easier to understand.

It does not constitute legal advice and does not create a professional adviser–client relationship. The information should not be relied upon as a substitute for advice based on the specific facts, circumstances and applicable law relevant to your organisation or project.

The article reflects our understanding of the law and regulatory framework as of the date of publication. Legislation, case law, regulatory guidance and administrative practice may subsequently change. While reasonable care has been taken in preparing this article, CORVUS AI does not warrant that the information is complete or remains current after the date of publication. We do not undertake to update this content.

To the fullest extent permitted by applicable law, CORVUS AI excludes liability for loss arising from reliance on this article. Nothing in this article constitutes an offer or solicitation to provide regulated legal services in any jurisdiction where doing so would be unlawful.

AI-assisted preparation: This article was prepared with the assistance of AI tools. Its legal analysis, conclusions and final text were subject to human review and editorial control and were reviewed and approved prior to publication by Oleksandr Sobovyi, Founder & CEO of CORVUS AI. CORVUS AI retains editorial responsibility for the published content.

For advice tailored to your organisation, project or specific circumstances, please contact CORVUS AI.

Oleksandr Sobovyi
Founder & CEO, CORVUS AI
www.corvusai.eu

Editorial Transparency Statement

This article was prepared with the assistance of artificial intelligence tools. Its content has undergone substantive human review and editorial oversight. CORVUS AI assumes editorial responsibility for the publication, with Oleksandr Sobovyi, Founder & CEO, as the responsible person.

The use of AI is disclosed voluntarily, taking into account the principles of transparency and Article 50(4) of Regulation (EU) 2024/1689.