⚡ Quick Summary
Published by Japan AI Safety Institute, this report investigates the socio-technical influences of AI systems that include foundation models handling multimodal information. It examines effects extending beyond direct end users to surrounding individuals, institutions, social systems and industries. The research combines publicly available literature with interviews intended to provide a detailed understanding of selected issues in Japan. It frames its findings through 12 research topics in four categories: ethics and law, economic activities, information space, and environment.
The report relates these topics to six key elements of AI Safety identified in AISI’s Guide to Evaluation Perspectives on AI Safety: Human-centric, Safety, Fairness, Privacy protection, Ensuring security and Transparency. Its research process collected approximately 50 preliminary sources, selected 20 for detailed research, reorganised them into 12 topics, and interviewed ten organisations across five selected topics. The final chapters identify technical, social and institutional issues, then propose responses at national or regional, corporate and individual layers, including training-data quality, output controls, literacy, work redesign and institutional discussion.
🧩 What’s Covered
The report proceeds from its rationale and method to topic-based findings and cross-cutting considerations.
- Background and scope: Explains why the focus moves beyond direct effects on end users to interactions between AI systems and their social context. It cites the NIST AI Risk Management Framework and the International AI Safety Report 2025 to situate AI as inherently socio-technical.
- Research method: Describes preliminary literature research across domestic and international news, papers and reports; detailed investigation of 20 selected sources; and interviews with ten organisations. It cautions that interview findings reflect the perspectives of each organisation.
- Ethics and law: Covers psychological and physical risks from overreliance, discriminatory output bias, cyberattack exploitation, and generation and distribution of obscene materials. Examples and interviews address dependency, critical thinking, phishing, deepfakes, AI-CSAM, stakeholder responsibilities and gaps in applicable rules.
- Economic activities: Examines intellectual-property, publicity and portrait-rights concerns; labour-market substitution and complementarity; AI-generated spam and review pollution; inequality between workers, firms and countries; and training or leakage of confidential information.
- Information space: Addresses AI-generated misinformation and disinformation, including fake articles, video, audio and documents, as well as echo-chamber effects. It also considers attribute and cultural diversity in outputs, alongside examples where AI can assist accessibility and inclusion.
- Environment: Sets out reported energy use, emissions and local power-grid pressures associated with generative AI and data centres, while noting examples of AI-assisted wind forecasting, electricity-market forecasting and renewable-energy procurement.
- Future considerations: Groups current issues as training-data quality, inappropriate outputs, widening inequality, insufficient individual literacy and handling of AI-generated outputs. It discusses measures spanning public education and institutional design, corporate guardrails and risk management, and individual checking of sources, evidence and related information.
💡 Why it matters?
The report gives governance, security and policy teams a structured view of harms that can arise when generative AI is deployed within social systems rather than assessed only as a technical product. It links particular issues to affected stakeholders, including developers, providers, users, platforms, educational institutions, governments and infrastructure operators. Its matrix also connects the 12 topics to AISI’s six AI Safety elements, helping teams relate matters such as biased outputs, confidential-information disclosure, cyberattack misuse and misinformation to broader safety priorities.
For organisations adopting generative AI, the interview findings make implementation questions concrete: monitoring customer-facing systems, setting input rules, improving AI literacy, maintaining human verification, and adapting threat-information sharing.
❓ What’s Missing
The report states that its topics and the content under each topic do not ensure complete comprehensiveness because technical and social conditions are changing rapidly. Its interviews are a sample of ten organisations, and the document says their findings reflect each interviewee’s perspective; detailed understanding is focused on the situation in Japan for selected topics. The final measures are explicitly described as considerations rather than an exhaustive action list. Although it identifies measures such as guardrails, provenance metadata, literacy education and risk management, it does not provide a single implementation checklist, detailed control specifications, or a standardised method for measuring and comparing the prevalence of every harm discussed.
👥 Best For
AI governance leads, public-policy teams and risk owners seeking a broad issue map for generative AI adoption. It is also suited to security teams considering AI-enabled phishing and deepfakes, education and workforce leaders addressing literacy and work redesign, and organisations mapping social impacts to AISI’s Human-centric, Safety, Fairness, Privacy protection, Ensuring security and Transparency elements.
📄 Source Details
Report on the Specific Influence on AI Safety is an 81-page English report dated October 31, 2025. The cover identifies Japan AI Safety Institute and lists NRI SecureTechnologies, Ltd. as “Research Contractor.” No individual authors, edition number or URL to this report are printed in the supplied document.
