Humanoids on Construction Sites and Legal Discipline: Reflections on the Challenges Identified in the McKinsey Report (2025)

0 Introduction

The landscape of construction sites is about to change. A report published by McKinsey & Company in October 2025 suggests that humanoid robots have stepped out of the laboratory and into the practical application phase. Note 1 This is not merely a story of technological innovation. It is a survival strategy for a construction industry struggling with labor shortages, and at the same time, it signifies the emergence of questions that shake the existing frameworks of labor law, tort law, and safety management.

Construction is an industry with a high concentration of law to begin with. This is because there are many dangers, and because there are many dangers, contracts and regulations become heavy. Into this environment enters a machine that has a general-purpose body and operates on the premise of the irregularity of the site. What is being questioned, before whether the robot works, is whether we can explain what the robot did.

This time, using the data and forecasts presented in the report as a guide, I will examine the legal issues raised by the introduction of humanoids in the construction industry.

1 Location of the Problem

Why humanoids now? The answer lies in the numbers. According to the report, from 2000 to 2022, while the global economy's overall productivity grew at an average annual rate of 2.0 percent and manufacturing at 3.0 percent, construction remained at 0.4 percent. Note 1 While other industries are increasing efficiency through digitalization and automation, only the construction industry is being left behind.

The background to this stagnation is the non-routine nature of the site. Unlike a factory, it is not a well-organized environment; conditions change daily due to weather and construction processes. Arrangements change, heavy machinery and people move, and there is constant danger. Communication is also unstable. There is a premise that even single-function robots are difficult to deploy on a large scale. Note 1

In addition, the aging of skilled workers and the decline in young entrants are overlapping, making the decline in labor supply capacity serious. The report estimates that demand for housing and infrastructure will increase, and supply will fall short by approximately $40 trillion. Note 1 The conventional solution of increasing manpower no longer works.

The technological turning point lies in the evolution of embodied AI. In particular, with the advent of Vision-Language-Action foundation models, robots are becoming able to interpret complex visual information on site and move according to verbal instructions in natural language. If a site supervisor instructs, 'Carry that red toolbox over there to the third floor,' the robot autonomously recognizes the object, climbs the stairs while avoiding obstacles, and carries out the task. This is different from conventional industrial robots that only move their arms according to pre-programmed coordinates.

According to the report's estimates, the general-purpose robot market will reach approximately $370 billion by 2040. Note 1 However, large-scale deployment in the construction sector is still considered to be on a ten-year horizon. Note 1

2 Organizing Legal Issues

The report points to the lack of legal and regulatory guidelines as a barrier to large-scale deployment, alongside technology and costs. Note 1 From a legal perspective, the introduction of humanoids to the site raises three points of contention.

First is the tension with occupational health and safety legislation. Conventional industrial robots have generally been operated in isolated spaces enclosed by fences. Note 2 Safety has been guaranteed by physically separating the spaces of humans and robots. However, the true value of humanoids lies in operation without fences, where they work mixed with humans. The report also cites fenceless operations as the next step. Note 1 While safety standards for collaborative robots are being formulated in international standards such as ISO 10218 and ISO/TS 15066, Note 4 it is unclear to what extent current legislation can tolerate the coexistence of autonomously moving heavy objects and humans in a high-risk environment like a construction site.

Second is the relationship between product liability and learning AI. Usually, accidents caused by machine defects are subject to product liability law. Note 3 However, humanoids equipped with VLA models may perform additional learning through data collection on site. If a robot that was safe at the time of shipment begins to behave unexpectedly as a result of learning specific data from a particular site, can this be questioned as a design defect by the manufacturer? The manufacturer will likely claim immunity as a transformation due to post-shipment learning, and the user will likely claim it is a defect in the learning algorithm itself. Whether the cause of the accident can be identified will determine the outcome of the litigation.

Third is the issue of data privacy. Since humanoids see and understand their surroundings, they continue to capture video, audio, and location information. The faces and voices of workers are included. This means that the behavior, remarks, and work efficiency of every person working on site are constantly recorded as digital data. The Act on the Protection of Personal Information regulates the acquisition, purpose of use, and safety management measures of personal information, Note 5 but data collection on site is linked to safety management itself. There is a tension where if you prioritize safety, surveillance becomes stronger, and if you weaken surveillance, safety drops. Information necessary for collision avoidance is different from information that can be used to evaluate workers. When the latter is mixed in, labor management issues are introduced.

3 The Causal Chain of Responsibility Becomes Longer

'Responsibility dissolving' does not mean that it becomes unclear who is at fault. It means that the chain of causes becomes longer, and parts of the chain are rewritten daily by software updates.

It is enough to imagine a scene where a humanoid causes an accident on a construction site. During the morning setup, the humanoid carries materials, a worker gives instructions by voice, and another worker continues to work nearby. The floor may be wet, and the scaffolding is temporary. When a collision or fall occurs here, the cause is unlikely to be singular. Blind spots in visual sensors, errors in language understanding, habits in path planning, changes in layout on the site side, ambiguity in instructions, communication delays, and the presence or absence of remote control intervention can all be part of the causal chain.

With conventional machines, the dichotomy of defect versus negligence functioned to a certain extent. Product liability law stipulates the liability of manufacturers, etc., for damages caused by defective products, and the Civil Code provides for tort liability due to negligence. Note 3 and Note 6 However, a humanoid is a product, and at the same time, its character changes depending on how it is operated. This is because its behavior can change every time an update is applied. The boundary between defects at the manufacturing stage and management failures at the operational stage is easily blurred.

What becomes an issue here is the authority for human intervention and audit logs. Intervention authority is necessary to avoid accidents and is also necessary to allocate responsibility. Logs are necessary to identify the cause after an accident occurs and to draw the line of responsibility. Without logs, the cause will fall into a vague conclusion that the AI is somehow at fault or the site is somehow at fault. Without clarity, the contractual risk allocation described later cannot be structured.

4. Contracts and Logs as Preparation for Implementation

A problem that arises in the early stages of considering implementation is the possibility of identifying the cause in the event of an accident. If the cause cannot be identified, the allocation of responsibility cannot be determined. Cause identification must be built into the design and operation. It is not something that investigative agencies can simply figure out after the fact.

To this end, there are things that must be decided by contract. These include who has final operational responsibility, who holds the authority for remote operation, who can trigger an emergency stop and when, and what the procedures are for applying updates. While updates are convenient, they can also become a weapon from the perspective of accident investigation. Accidents that occur immediately after an update is applied tend to blur the line between whether the cause was a design defect or an operational error. It is necessary to incorporate update history and verification procedures into the contract so that it can be explained what changed before and after the update.

The same applies to data. It must be decided what is collected, where it is stored, who has access to it, and when it is deleted. The requirements of the Personal Information Protection Act cannot be retrofitted to suit operational convenience. Note 5: Distinguish between data necessary for ensuring site safety and data that could be diverted for other purposes. If they cannot be distinguished, do not collect them at all.

The report proposes prior verification using digital twins. Note 1: Instead of introducing robots directly into the physical site, simulations are conducted in a virtual space utilizing BIM data. This allows for the verification of task feasibility and additional AI training without incurring physical accident risks. From a legal perspective, identifying risks at this stage is also effective.

5. Practical Implementation and Future Outlook

The report states that it is unclear how insurers will react to humanoids and which risks they will underwrite. Note 1: This sentence is short, but its implications are significant. Whether robots as capital investment are viable depends on whether it is decided who pays in the event of an accident. In implementation contracts, unlike conventional construction machinery rental contracts, it is necessary to specify in detail data governance clauses, the scope of exemption from liability resulting from AI judgment errors, and the allocation of cybersecurity risk burdens.

Economic hurdles also remain. Currently, the unit price of a humanoid is said to be between $150,000 and $500,000. The report states that this needs to drop to between $20,000 and $50,000 during the mass adoption phase. Note 1: From 2022 to 2024, capital inflow into general-purpose robots increased fivefold, exceeding $1 billion annually. Note 1: History shows that the logic of capital accelerates technological innovation, and it is highly likely that the price hurdle will lower over time.

What the report consistently emphasizes is that humanoids are not meant to replace humans, but to support them. Note 1: Robots take on heavy labor and dangerous tasks, while humans focus on higher-level decision-making and precision work. Clarifying this division of labor and promoting implementation in tandem with reskilling for workers will increase the feasibility of adoption.

6. Implications for Japan

The Artificial Intelligence Basic Plan, approved by the Cabinet in December 2025, includes several points related to humanoids at construction sites. Note 7

First, physical AI has been positioned as a pillar of policy. The plan states that it will "strategically and integrally promote research, development, and demonstration of physical AI, including the creation of public demand for AI robots and the introduction of more advanced autonomous driving technology, so that Japan can lead the world." The construction field is also explicitly mentioned as a field for promoting utilization under "infrastructure construction and management." It could be said that the landscape ten years from now, as depicted by the McKinsey report, has come within the scope of Japanese policy documents.

Second, a review of civil liability is being signaled. The plan explicitly states that it will "examine the nature of civil liability and its scope in the event of accidents or damages occurring in the use of AI." The issues discussed in this paper, such as the length of the causal chain of responsibility, the relationship between product liability and learning algorithms, and cause identification through logs, are points that cannot be avoided in future institutional design.

Third is the positioning of data. The plan positions high-quality data in "quasi-public sectors such as construction" as a "winning strategy for Japan" and proposes the construction of a data linkage infrastructure. However, as pointed out in this paper, data collection at construction sites is two sides of the same coin as safety management, and it carries tensions with privacy and labor management. To make it a winning strategy, the design of what to collect and what not to collect must come first.

Fourth, the strengthening of the functions of the AI Safety Institute is proposed. The plan states that it will immediately expand personnel to about twice the current level, using the scale of the UK's AI Safety Institute as a benchmark. The role that the AISI can play in accumulating knowledge on humanoid safety evaluation and accident investigation is significant.

The policy framework is being put in place. What is being questioned is the detailed design that can withstand implementation in the field.

7. Conclusion

Humanoids in the construction sector are easily talked about as a novel vision of the future. However, to actually make them work on-site, details are more necessary than a vision of the future. How far to make them autonomous, who stops them, what to record, and how much to share. The law is the party that demands this detailed design.

As the report indicates, large-scale adoption is still on the horizon. Note 1: However, allocation of responsibility and contract design are too late if left until just before implementation. As preparation for implementation, create logs, contracts, and data governance first. That seems to be the shortest path to bringing humanoids into the field.

(Reference) Physical AI

(Magazine) "AI and Law - Random Thoughts"

*Please refer to the following for the table of contents

note General Terms of Service Article 3, Paragraph 2, First Sentence
3.2 The copyright of digital content created by the creator belongs to the creator.