Introduction

Recently we have experienced a dramatic increase in requests and consultations concerning AI development agreements.
We have been consulted from the perspective of both users seeking to outsource AI development to a vendor as well as from AI vendors seeking to perform contracted AI development for users using data provided by such users. The details of such consultations vary and include the following:

• How should an AI vendor respond to a strong request from a user for certain warranties concerning accuracy of the output using a trained model?
• What kind of provisions concerning inspection standards and defect liability should be included in an AI development agreement?
• What points concerning development results and intellectual property rights should be focused on in negotiations and what type of contract terms should be included?
• Who is liable if a user or third party incurs damage when the system incorporating AI developed by the vendor using data provided by the user malfunctions?
• In anticipation of an AI malfunction, what kind of provisions should be in an AI development agreement?

If we classify the consultations about these AI development agreements, they roughly fall into the following three areas:

1. What provisions should be included in an AI development agreement with respect to performance assurance, acceptance inspections, and defect liability(performance assurance, acceptance inspections, and defect liability)?
2. Who has what rights with respects to generated training datasets, trained models, and trained parameters (rights/intellectual property)?
3. What provisions should be included in an AI development agreement to address damages that may occur when developing or using the AI (liability)?

Simply put, these on-site problems derive from the difference between AI software development and the conventional rule-based software development, in particular with diametrically opposed claims of both the vendor and user with respect to the second area of rights/intellectual property mentioned above.
Consequently, the negotiation for execution of an AI development agreement takes an extremely long time, resulting in cases where the negotiation ends up being cancelled or a company loses its competitiveness by falling behind other companies.
In other words, the unsuccessful negotiation of the three areas mentioned above in the AI development agreement is directly linked to the detriment of the individual companies, thereby creating a large obstacle for Japan’s AI development.

Therefore, having certain guidelines when the AI development agreement is executed is necessary; the Contract Guidance on Utilization of AI and Data formulated by the Ministry of Economy, Trade and Industry on June 15, 2018 will be very helpful on this point.

Reference:
Contract Guidance on Utilization of AI and Data

Of the two sections into which this Guidance is divided (the “Data Section” and the “AI Section”), I served as a member of the AI Section Guidance Committee and Subcommittee. Naturally, since the AI Section of the Guidelines alone was more than 180 pages, many voiced a desire for an explanation of the Guidelines that could be easily understood.
As a result, based on my personal opinion, I have tried to summarize the “8 Points about AI Development Agreements” that can be learned from the “Contract Guidance on Utilization of AI and Data” based on the “Contract Guidance on Utilization of AI and Data (AI Section)” (hereafter, the “AI Guidelines”).
Further, please note that this article is based on my personal opinion about the AI Guidelines and does not constitute an “overview of the Guidelines” or “the essence of the Guidelines”, let alone an official interpretation of the Guidelines. I urge you to refer to the text of the Guidelines for the actual contents of the Guidelines.
An overall picture of the “8 Points about AI Development Agreements” that can be learned from the “Contract Guidance on Utilization of AI and Data” is presented below.

The “8 Points about AI Development Agreements” that can be learned from the “Contract Guidance on Utilization of AI and Data”

1. Performance assurance, acceptance inspections, and defect liability
   (1) Mutual understanding of characteristics of AI and its limitations
   (2) Dividing the process from the contract
   (3) Devising the contents of a development agreement
2. Rights/Intellectual Property (1) Among the materials, interim deliverables, and deliverables, know which are or are not covered by intellectual property rights (2) With respect to (1) above, know who has what rights under the default rules (i.e., a legal rule) (3) Know how to craft contract provisions that benefit your own company (without being particular about the “ownership of intellectual property rights”, prioritize the “terms of use”) (4) Know the limitations of the contract
3. Liability Know the types of liability in AI development and control them in the contract

What type of provisions concerning performance assurance, acceptance inspections, and defect liability should be included in an AI development agreement? (performance assurance, acceptance inspections, defect liability)

Why is this a problem?

Frequently asked questions

• How should an AI vendor respond to a strong request from a user for certain warranties concerning accuracy of the output using a trained model?
• What kind of provisions concerning inspection standards and defect liability should be included in an AI development agreement?

Performance assurance, acceptance inspections, and defect liability are extremely contentious areas in negotiating an AI development agreement.

A conventional system development agreement not using AI technology usually contains provisions in which the vendor warrants a certain performance of the deliverables, conducts an acceptance inspection of the deliverables based on certain acceptance inspection standards, thereafter deciding whether the deliverables pass such inspection, and stipulates defect liability for cases where a defect is found in the deliverables. For that reason, even in AI development agreements, users often strongly request the vendor to incorporate provisions with certain performance assurance, acceptance inspection, and defect liability.
However, based on its technical characteristics, AI software has certain unique features:the difficulty, in principle, of performance assurances for unknown data since the inclusion of statistical biases in the training dataset cannot be avoided; the difficulty of analyzing flaws that may occur due to the existence of multiple causes (such as the quality of data, setting hyper parameters, and bugs in the source codes); and when the acceptance inspection of deliverables is conducted, an independent dataset not using learning is necessary and, obviously, when a test of unknown data is not possible.

Consequently, the vendor often rejects the user’s requests above because the vendor is unable to comply with them, thereby making compromise between the vendor and the user practically impossible in many cases.
The reason for the opposing views of the conventional system development and the AI software development with respect to performance assurance, acceptance inspections, and defect liability stems from the fact that the development approach for conventional system development is deductive while that of AI software is inductive.

Please see this article for more details about “deductive” and “inductive”.

Reference article:
Deduction and induction in a couple next to me at a Shibuya beef tongue restaurant and AI development

Incidentally, in AI software development cases, initially the business department and technology department of the user and the vendor are very enthusiastic, saying “let’s definitely proceed with this matter”; however quite often this type of talk stops the instant that the legal and intellectual property departments of the user become involved. This is due to the fact that the performance assurance, acceptance inspections, and defect liability [in AI development] are quite different from conventional system development.

So, what should we do?

If this type of difference exists between the conventional system development and AI software development, the problem is how to overcome [this difference] when entering into an AI development agreement. As I earlier wrote briefly about “Deduction and induction in a couple next to me at a Shibuya beef tongue restaurant and AI development”, I think the following 3 points need to be addressed in order to overcome this difference].

1. The user and the vendor’s understanding of the characteristics of AI development
2. Dividing the development process from the contract
3. Crafting the contents of a development agreement

1. The user and the vendor’s understanding of the characteristics of AI development As I wrote earlier, from the very start, there is a difference between the concepts of development techniques for conventional system development and that of AI software development. It is very important for the user and the vendor to have a common understanding of this point.

In fact, the AI Guidelines were intentionally drafted with this point in mind.
That is the reason why the Guidelines spend a considerable amount of space to describe AI technology (Section 2 “Explanation of AI Technology”).
Although there is a large body of printed material and information concerning AI development available for engineers, it seems that until now there have been no straightforward printed materials to enable staff in company’s legal or intellectual property division to understand the technical characteristics of AI software and the differences between AI development and conventional system development.
In the AI Guidelines, Section 2 “Explanation of the AI Technology”, which is divided into four parts (“1. Explanation of Basic Concept”, “2. Target AI Technology”, “3. Assumed Process for Practical Application of AI Technology”, and “4. Characteristics of Software Development Using AI Technology”), describes in a manner as detailed and easy to understand as possible the basic concept of AI development, the characteristics of AI technology, and its advantages and limitations.
If contract negotiations between the user and the vendor reach an impasse or if the user’s technical division wants its internal corporate legal or intellectual property division or top management to understand [the characteristics of AI development], please do not hesitate to use these AI Guidelines. [Advertising]