AI Ethics News — 2026-10-28 (Wed)

Overview

Questions of consent and disclosure dominated the past 48 hours. Rights holders sued over AI training data, California moved on safeguards for children using AI companion chatbots, and experts warned about AI-driven disinformation ahead of the US midterms. Fairness measurement and transparency about generative AI data were also raised, including in Japan. 【Bias, discrimination and fairness】 - A report argues that AI evaluations can look unbiased without being fair, with the headline "AI Evaluations Look Unbiased Without Being Fair" (Technology Org). - An ensemble AI model that flags and rewrites media bias in news text was reported (Bioengineer.org). - The headlines do not say who is disadvantaged, nor what measurement method or sample size was used.

Privacy and surveillance

- Several pages promote adult chatbot, dating and sex app services, and their titles mention privacy controls, consent and age verification (NYC Health + Hospitals, Shodoshima Town). - The headlines do not say which data is collected, how consent is obtained or how the data is used. 【Transparency and accountability】 - CPA Practice Advisor published a piece on where AI belongs in the audit process and where it does not (CPA Practice Advisor). The headline does not say what is disclosed or withheld. 【Creators, copyright and consent】 - Elvis Presley's music rights holders are suing Anthropic for $48.6M over AI training. The issue is whether music was used for AI training without the rights holders' consent (Türkiye Today).

Disinformation and deepfakes

- Confirmed: Gizmodo describes ChatGPT's election-related anti-deepfake measures as "intriguing, but incomplete" (Gizmodo). - Claim: experts warn that AI-driven disinformation is running rampant on social media before the US midterms (The Guardian). The headline gives no figures on the scale of harm. 【Protecting children and vulnerable people】 - Children: California is moving to add safeguards for kids using AI companion chatbots, so a public authority is acting (KBAK). - Anthropic says children have used its Claude AI, and it flags unknowns about mental health effects (EEW Magazine). This is a disclosure by the company itself. 【Ethical principles and guidelines】 - Bioengineer.org reports work on turning AI ethics principles into checkable clinical controls. This appears to be a proposal (Bioengineer.org). - The ILO is examining labor standards for ethical AI. The headline describes an examination, not a decision (Digital Watch Observatory). - The World Economic Forum discusses how the creator economy is redefining trust in the age of AI (World Economic Forum).

Japan's AI ethics

- Google and four other companies responded to the Japanese government on measures to ensure transparency of generative AI data (Asahi Shimbun). The headline does not give the content of the responses. - Two further pages cover generative AI pornography: an NSFW AI chat service with customizable role-play and privacy settings, and a review of AI porn creation tools covering free credits, features and ethics (NYC Health + Hospitals).

Statements by key figures

- Yoshua Bengio (AI safety researcher) is to join Berlin Science Week and address trust and responsibility (idw - Informationsdienst Wissenschaft). - No developments were found for Timnit Gebru or Fei-Fei Li. 【Integrated analysis: how the ethical questions connect】 - Lack of consent is a shared structure. The Elvis Presley rights holders' suit over training data, the child-safety measures for companion chatbots, and the adult chatbot pages that stress consent and age verification all turn on who agreed to what. (Türkiye Today, KBAK, NYC Health + Hospitals) - Lack of disclosure is the second shared structure. Training data transparency in Japan, Anthropic's statement about mental health unknowns, and the question of where AI belongs in audits all concern what is revealed and what is not. (Asahi Shimbun, EEW Magazine, CPA Practice Advisor) - Measurement is a third theme. The point that evaluations can look unbiased without being fair, and the description of ChatGPT's anti-deepfake measures as incomplete, both suggest that having a safeguard does not show it works. (Technology Org, Gizmodo) - Inference: AI developers may need to document where training data came from. Operators of companion chatbots may need to weigh age verification and child safeguards. Organizations that rely on a single bias score may need additional fairness checks.

What to watch

Whether the suit against Anthropic over AI training leads to wider disputes about consent for training data (likely to be tracked through court filings). 2. How California's safeguards for children using AI companion chatbots are shaped, and whether other operators follow Anthropic in disclosing child use. 3. Whether the responses from Google and the other companies to the Japanese government lead to concrete transparency measures for generative AI data.