NIPPON TELEGRAPH AND TELEPHONE CORPORATION: G7 Hiroshima AI Process (HAIP) Transparency Report

At the NTT Group,we identify risks using a common AI risk checklist, where we classify target use cases and risk levels in our group-wide AI risk management policy. We classify risks into three levels: unacceptable (prohibited), high, and limited.

We have established a comprehensive risk management approach to identify and evaluate various AI-related risks throughout the lifecycle of AI projects. We have established a common AI risk definition and developed an AI risk management flow, which guides our risk assessment processes. Each of our service providers creates and applies specific rules based on this framework to address vulnerabilities, incidents, emerging risks, and potential misuse in their operations. This collective effort contributes to our robust governance framework, ensuring consistent and effective risk management across all projects.

For LLMs (Large Language Models), used by the general public, we seek third-party red team testing to assess AI vulnerabilities. For RAG (Retrieval-Augmented Generation) systems within specific organizations, internal teams conduct functional and operational checks to ensure deployment readiness.

Yes

As we primarily operate in a B2B context, we focus on continuous improvements per project, guided by incident and malfunction reports from customers. We apply both quantitative and qualitative risk evaluation metrics in this process. For large-scale B2C services, while we recognize the potential effectiveness of incentive programs for vulnerability detection, they are not currently implemented due to our focus. Our group has established a common AI risk definition and management flow, supporting stakeholder engagement through specific service provider rules to achieve a comprehensive governance framework. We believe it is important to address not only legal compliance but also concerns of ethics and social acceptability. This view is incorporated into our common policy within the NTT Group to ensure it also guides our actions.

We leverage external independent expertise by revising our AI risk definitions and management flow based on insights from international discussions, industry groups, and academic exchanges.

We will review the necessity of ongoing third-party vulnerability detection after the implementation of AI projects in the future.

In addition to participating in the activities of ISO/IEC JTC1 SC42, we plan to expand our scope to include activities that are emerging in other organizations, mainly around JTC1, widening our understanding of developments (e.g., ITU-T, etc.).

For AI projects, we conduct risk assessments and, if systemic risks are identified, involve AI risk management and legal specialists. This collaboration ensures that we effectively work with stakeholders to implement adequate risk reduction measures.

Our AI risk management office establishes guidelines for assessing risks in advance on a per-use-case basis, allowing us to implement appropriate measures throughout the AI lifecycle.

We use testing measures to identify risks such as hallucinations in AI systems. To mitigate these risks effectively, we implement preventative actions, including advanced warnings to users about AI system behaviors are restricted based on the findings of tests.

For RAG systems incorporating LLMs used within specific organizations, basic functional tests and verifications are conducted. Particular attention is given to the search accuracy for retrieving context information aligned with input prompts, and verifying the factual basis of the system's responses to the context based on how accurately it can answer aligned with the context.

For the LLMs we provide, we ensure data quality by using blacklists and manual verification to exclude information from sites associated with piracy or social issues. Additionally, we confirm that outputs are not biased when input variables like gender, age, or race are altered, ensuring fair and unbiased results.

Regarding RAG systems that incorporate LLMs, based on the use case, the terms of use specify precautionary measures to prevent generating instructions that rely on copyrighted content, ensuring intellectual property protection.

We have established operational rules to identify and mitigate privacy risks, ensuring that any input information that could violate privacy is not collected or used without permission beyond its intended purpose. In our RAG systems incorporating LLMs, we have guidelines specifying that data used for RAG extensions must not include personal information that constitutes unauthorized use or confidential information belonging to third parties outside the customer's scope.

i. Assessing Cybersecurity Risks and Implementing Policies:

• In cases requiring measures against misuse, inappropriate handling of personal information, and prompt injection, we evaluate widely available products with guardrail features, integrating them into our systems if deemed effective. For LLMs (Large Language Models), we use tools to assess vulnerabilities and security risks. In RAG systems, we implement user restrictions and limit access to reference documents.

ii. Protecting Valuable IP and Trade Secrets:

• Our systems integrate with existing foundation models and RAG systems. Since we don't currently perform additional training or fine-tuning on these models, the necessity of addressing related security concerns will be discussed should it become relevant in the future. We manage versions of trained models and datasets, applying access restrictions to ensure secure management..

iii. Vulnerability Management Process:

• For RAG systems, we perform incident-driven monitoring and response post-deployment. We will discuss the need for proactive periodic assessments in future reviews. We identify social reputation risks and take action based on risk levels.

iv. Review of Security Measures:

• We conduct risk assessments and review security measures whenever a vulnerability report is generated, and at least annually, in line with NTT Group security policies.

v. Insider Threat Detection Program:

• "We have an insider threat detection program to safeguard against internal security risks."

For RAG systems incorporating LLMs, we plan to establish a help desk for each project to receive notifications about malfunctions and unexpected behaviors. Additionally, we intend to monitor input-output logs and manage inappropriate inputs and outputs as needed, ensuring ongoing oversight and responsive action.

For RAG (Retrieval-Augmented Generation) systems incorporating LLMs (Large Language Models), we disclose and report test results to customers regarding the impact on the business areas targeted by the RAG system.

Our AI risk management office takes the lead in facilitating exchanges of opinions within the group's AI governance liaison meetings, as well as with private sector organizations, the media industry, and academia.

To prevent inappropriate use of personal information and privacy violations, we outline these risks in our AI governance policy. These are included in a risk checklist to ensure thorough checks through established operational rules.

While we maintain training data in a state that allows us to manage its quality and accuracy, the information itself is not disclosed publicly.

We have not implemented a transparency report but it is scheduled to be released in the future.

Within the NTT Group's overall risk management framework, AI risk is considered significant. We have established an AI risk check flow based on a common AI risk definition, allowing each service provider to formulate specific rules and achieve a governance framework. We continuously review and update our policy to align with the EU AI Act, other national and international AI guidelines, and in response to significant AI incidents.

As the NTT Group, we are putting in place regulations and, through AI governance liaison meetings and briefings, are communicating about governance policies and risk management practices while addressing issues through the establishment of help desks. We are working towards rolling out group wide educational materials.

Our AI initiatives and risk management policies are published on the company's website: About AI at NTT. [1]

Additionally, we actively participate in AI governance-related events and interviews to communicate with the general public.

[1] https://group.ntt/jp/group/ai/

We maintain internal documentation of reported incidents, which includes details of the reported risks and analyses of the risk characteristics and root causes. These documents are used to organize and implement measures to prevent recurrence.

We share incident cases and risk detection information through the NTT Group AI governance policies and internal group AI Governance committee meetings.

We have established a common AI governance policy for the NTT Group and are actively engaging in information-sharing activities. We plan to share incident and risk detection cases through liaison meetings within the group.

For generative AI, which presents various societal risks, we have established and shared a set of generative AI usage guidelines within the organization. These guidelines help identify and address risks for stakeholders like model developers, service providers, and AI users, providing management approaches.

We align our AI risk definitions and governance processes with standards such as the HAIP code of conduct, the EU AI Act, and the Japanese government's AI business operator guidelines.

Across the organization, we extensively implement guidelines and measures for AI generation. For RAG systems specifically, we ensure users are informed by presenting guidelines and conducting explanatory sessions.

From the development stage of an AI project, we conduct training after implementing contracts and reaching agreements with data providers about intended use. We consider this to be best practice because the providers are operating internationally as a business.

Currently, the services we offer using LLMs are limited to text generation and do not include images. In the case of text generation, incorporating messages that indicate the content's source or that it is AI-generated may affect the user experience. Therefore, to enable flexibility in the user experience and our product design we handle the inclusion of such messages through user service agreements. Should we add support for image generation in the future, we will evaluate the availability of existing technical tools and proceed accordingly.

Were we are developing or providing LLM services, as part of our product development we utilize publicly available evaluation tools to appraise them in areas of quality, ethics and security. Our research also includes:

1. Reducing Hallucinations and Controlling Inappropriate Speech:

• Efforts are aimed at minimizing hallucinations in AI outputs and managing inappropriate speech.

1. Composite AI System for Cybersecurity:

• This system automatically tracks and adapts to the evolution of cybersecurity attack techniques, enhancing interpretability and transparency in detecting and blocking attacks.

1. Technologies and Practices for Addressing Latest AI Threats:

• Inspection Technologies: Tools for assessing AI systems against emerging threats.
• Protection Technologies: Mechanisms to safeguard AI systems from identified threats.
• Trust Enhancement Practices: Developing methods to strengthen trust, especially regarding threats within the AI supply chain."

 Utilizing the outlined technologies to develop methods that strengthen trust, particularly in addressing threats within the AI supply chain.

Currently, the services we offer using LLMs are limited to text generation and do not include images. For the text generation we provide, incorporating messages that indicate the content's source or that it is AI-generated may affect the user experience. Therefore, we handle the inclusion of such messages through use service agreements. Should we add support for image generation in the future, we will evaluate the availability of existing technical tools and proceed accordingly.

We are conducting research and investing in mechanisms to manage inappropriate AI-generated speech, aligning AI systems with democratic principles and human rights. We are also pursuing research on composite AI systems that reduce hallucinations and autonomously adapt to evolving cyberattack techniques such as those aiming to deceive humans, enhancing the interpretability and transparency of threat detection and blocking mechanisms.

We are conducting research to develop models that achieve high performance in Japanese language processing using extremely lightweight models compared to large parameter LLMs like GPT-4. This research aims to provide more efficient language models, reducing both training and inference(operational) costs and minimizing environmental impact.

We are conducting research to develop models that achieve high performance in Japanese language processing using extremely lightweight models compared to large parameter LLMs like GPT. This research aims to provide more efficient language models, reducing both training and inference(operational) costs and minimizing environmental impact.

We address social risks from Generative AI by summarizing risks and management approaches into generative AI usage guidelines for model developers, service providers, and AI users. Additionally, we are creating educational content to enhance AI risk management literacy among all employees, covering objectives, structure, risks, management, and case studies, including comprehension questions.

In our AI governance Charter, we define 'human-centered AI' as a key element and prioritize AI projects that focus on human-centric values.

We engage in initiatives such as regional revitalization through AI-operated buses, AI imaging diagnostics for tuberculosis for 100,000 people in India, and the development of AI-powered predictive failure detection technology for storage batteries, contributing to stable infrastructure for a decarbonized society. For more details, see our sustainability report [2].

[2] https://group.ntt/jp/csr/data/pdf/sustainability_report_2022_databook_all_20230228.pdf