From December 9, 2026, European Directive 2024/2853 applies to products placed on the market or put into service after that date, harmonizing liability for damages caused by defective products, including software. On the same date, the MIIT/SASAC circular provides for at least 20 units of priority scenarios in the concerned provincial regions, with validation and then regular deployment of successful solutions. Two opposing philosophies, an identical effect: regulatory and financial constraints influence deployment capacity beyond technical mastery alone.
The essentials
- By end 2026, coexist a European directive on defective products applicable to new products and a Chinese validation-deployment action in selected scenarios.
- The directive distributes responsibility among defined economic operators and facilitates, in certain cases, proof of the defect or causal link; the circular coordinates experimentation, evaluation and deployment support without establishing a state compensation scheme.
- The texts impose technical and economic requirements.
- SMEs and robotics startups face significant financial constraints, regardless of their technical mastery.
- Regulatory bifurcation creates distinct access barriers according to the two regimes.
The European directive transfers risk without making it disappear
Directive 2024/2853, applicable December 9, 2026, modifies a forty-year-old rule: until now, the victim had to in principle prove the damage, the defect and the causal link, without having to demonstrate manufacturer negligence. The victim must in principle prove defect, damage and causality, but the defect or causality can be presumed in certain circumstances provided for by the directive. The 2022 directive proposal provided for rebuttable presumptions, but it was withdrawn on October 6, 2025.
This change has an immediate practical consequence. A hospital or factory deploying a robotic arm in its corridors now bears a financial risk difficult to quantify. Liability depends on the status of the economic operator and product defectiveness; it is not automatically imputed to the deployer. Insurance therefore becomes not optional but structural.
Robot producers probably already have liability insurance, according to the official Commission impact assessment, although the market remains embryonic at industrial scale. Players like Allianz, AXA or Munich Re are working on dedicated products, but pricing remains experimental due to insufficient actuarial data. Without a history of claims on large fleets, premiums are high and deductibles heavy. A deployer with internal self-insurance capacity can negotiate more favorable insurance conditions.
The result is mechanical. A manufacturing SME wishing to deploy humanoids in its workshop may face significant insurance premiums. A large industrial group, for its part, can build an internal provision, negotiate a collective policy and absorb the cost across multiple sites. The same concentration logic applies to AI adoption in business: the intention is broad, real integration remains the preserve of actors with resources to see it through.
Beijing socializes risk, but selects by state capital
The Ministry of Industry (MIIT) and the State-Owned Assets Administration (SASAC) published in June 2026 a circular asking participating regions to retain at least 20 units of priority scenarios and aiming at validations and then regular deployments of successful solutions by end 2026. Pressure comes from above. The circular provides for public funding and exploration of robotics insurance, without thus distributing financial risk.
This mechanism appears more inclusive. A mid-sized Chinese robot manufacturer can sell its robots to a provincial state-owned enterprise without directly worrying about the liability regime: it is the state buyer who bears the risk, at least implicitly. But selection occurs elsewhere. The circular provides for the involvement of enterprises, users and research institutes, without alleged attribution criteria. Several capitalized companies, partly supported by state-linked investment funds, participate in provincial deployments.
The circular provides for pilot evaluations and standards concerning in particular intelligence, reliability and safety, without imposed economic viability criterion, and without announced uniform technical thresholds. A six-month-old startup cannot respond to a provincial call for tenders in thirty days. Test requirements adapted to scenarios entail compliance costs for manufacturers.
Two philosophies, one common filter
Europe mobilizes civil liability law and the insurance market. China mobilizes public procurement and industrial planning. The architectures are at opposite poles. The texts respectively provide for a selection of experimentation scenarios and evidentiary rules, without a common selection mechanism based on actor resources.
In both cases, conformity and deployment capacities require structural resources. The circular imposes evaluations and technical requirements while providing support, without quantifying alleged costs or cycles. This phenomenon echoes what is observed in other sectors where robotization is accelerating: labor shortage creates real demand, but integration constraints concentrate supply among already structured actors.
Selection by capital has a safety logic. A well-capitalized actor has more to lose in case of serious accident. It has an incentive to invest in reliability, to build maintenance teams and to monitor software updates. The insurance market, by pricing risk, also encourages reducing it.
This argument has real foundation.
But selection by capital also excludes innovators who master their technology better than better-funded competitors. It slows the diffusion of uses toward mid-sized enterprises, where productivity gains could be most significant for the economy as a whole. It can favor established players, which tends to affect market dynamics.
Insurers substitute for regulators
The directive leaves insurers outside its direct normative framework. Insurers risk defining, through their policy conditions, safety, maintenance and traceability requirements for deployers. If Munich Re requires a timestamped logbook and certification of the navigation algorithm, the market aligns, regardless of what the directive says.
This phenomenon already has a history. The US cybersecurity market was shaped more by cyber-insurer requirements than by federal regulation. Insurers imposed multi-factor authentication, isolated backups and third-party audits well before federal agencies made them mandatory in certain sectors. The same dynamic is likely for European robotics.
This means that the actual technical standards for humanoids in Europe will be written, at least partially, in Frankfurt and Zurich rather than Brussels. It is a governance bifurcation that the directive does not document, but that it makes inevitable. The regulator sets liability; the insurer sets the conditions of its transfer; the deployer complies.
Artificial intelligence inside robots crosses a second regulation
Humanoids in 2026 integrate machine learning models aimed at adaptation and optimization of behavior. This software layer falls under the AI Act only if it constitutes an AI system within the meaning of the regulation and if the relevant scope conditions are met.
The AI Act classifies AI systems according to their risk level. A hospital robot is high-risk only if its use falls into an AI Act category or if it is a safety component of a regulated product concerned. If the system is legally high-risk, applicable obligations notably include risk management, documentation, human supervision, robustness and, depending on the category, registration. The AI Act is in force, but the proposed directive on AI liability was withdrawn on October 6, 2025.
SMEs can face regulatory costs and uncertainties, but there is no double AI Act–liability directive framework in force nor primary admission that they don’t know how to finance it. Large technology and industrial companies have legal teams and compliance budgets. Smaller integrators navigate with limited resources between two complex texts whose articulation is explicitly addressed by the European legislator; the proposed directive dedicated to AI liability was withdrawn on October 6, 2025.
Chinese AI standards are coordinated by the same ministerial structures that pilot industrial deployments. Coherence is easier to ensure in a centralized system, at the cost of less transparency on real evaluation criteria.
Gaps common to both regimes
A second structural gap concerns the absence of framework for intermediate situations, those where the robot has not caused bodily injury but has produced economic loss or failed to accomplish a critical task. The directive covers deaths, injuries, certain property damage and destruction or corruption of data not used exclusively for professional purposes; it does not cover lost earnings suffered by an operator whose humanoid robot stopped mid-production line without bodily incident. Chinese mandates, oriented toward deployment volume, do not integrate any recourse mechanism for a provincial buyer unsatisfied with the actual performance of a fleet certified as compliant. This economic gray zone is wider than the legal gray zone around updates, because it concerns the vast majority of daily incidents that never cross the threshold of documented accident. It creates contractual uncertainty that neither civil liability law nor planned public procurement is equipped to absorb, and which weighs more heavily on deployers with narrow margins than on large groups able to negotiate contractual guarantees directly with manufacturers.
Both texts contain elements relating to product lifecycle, even if they do not exhaustively govern all long-term uses. The directive treats software updates under manufacturer control as a relevant element of defectiveness, according to circumstances. The mandates set a minimum of priority scenarios and provide for later deployment of validated solutions; they do not specify what happens when provincial governments no longer have budgets to maintain purchased fleets.
The texts recognize at least partially product evolution, without strictly assuming its stability. A 2026 humanoid robot learns, drifts and receives updates. In Europe, software updates and modifications can fall under defectiveness rules according to circumstances; in China, the circular provides for formalizing operational responsibility standards.
The 2026 regimes will likely be revised on this point as a priority. The first serious accident involving a robot whose behavior had changed after an update will impose clarification: in Europe through case law, in China through administrative circular. The tension between capitalization and open innovation observable in other sectors will be found here in an even sharper form, around liability for emergent behavior that no programmer has explicitly designed.
The real bifurcation between European and Asian markets may not be decided by the initial liability regime, but by how each system responds to that first incident.
Sources
- AssurHumanoïde, Directive responsabilité produits IA 2026 (primary source)
- Directive 2024/2853 of the European Parliament and of the Council on liability for defective products, Official Journal of the European Union
- Regulation (EU) 2024/1689, AI Act, Official Journal of the European Union
- MIIT/SASAC, Circular on humanoid deployments by province, June 2026 (cited via AssurHumanoïde and Onlynnov robotics 2026)
- Onlynnov, Robotics Report 2026 (cited in primary source)
