Posts

Showing posts with the label mental safety

OpenAI's New Under-18 Principles Enhance AI Ethics and Teen Safety in ChatGPT

Image
On December 18, 2025, OpenAI updated its Model Spec —the written set of behavioral expectations that guides how ChatGPT should respond—by adding a new section: Under-18 (U18) Principles . The goal is straightforward: teens (ages 13–17) have different developmental needs than adults, and a “one-size-fits-all” safety posture can create gaps in higher-risk situations. At a high level, the update clarifies how existing safety rules apply in teen conversations and adds age-appropriate guidance where needed. The principles emphasize prevention, clearer boundaries, and stronger encouragement toward real-world support when risks show up. This article explains what the U18 Principles are, why they matter, and what “safe, age-appropriate behavior” looks like in practice—without turning teen safety into vague slogans. If you’re interested in related context on teen safety work, you may also want to read: OpenAI’s Teen Safety Blueprint . TL;DR What changed: OpenAI added ...

OpenAI Launches $2 Million Grant Program to Advance AI and Mental Health Research

Image
OpenAI has launched a grant program offering up to $2 million to support research on the relationship between artificial intelligence (AI) and mental health. The initiative focuses on exploring both potential risks and benefits of AI in practical mental health settings. TL;DR The text says OpenAI's grant program funds projects examining AI's impact on mental health safety and care. The article reports that research should address real-world AI applications and their ethical implications. The text notes the program aims to guide responsible AI use in mental health through rigorous study. FAQ: Tap a question to expand. ▶ What is the main goal of OpenAI's grant program? The program aims to support research that investigates how AI affects mental health, focusing on safety, benefits, and risks. ▶ Which types of research projects are eligible for funding? Projects studying AI's role in mental health diagnosis, treatment, ...

Evaluating Safety Measures in Advanced AI: The Case of GPT-4o

Image
Artificial intelligence models like GPT-4o present both opportunities and challenges. This article reviews the safety measures applied before GPT-4o’s release, focusing on understanding risks to human cognition and behavior and approaches to mitigate these risks. AI safety is important to minimize potential harm to users and society. TL;DR External red teaming involves experts probing GPT-4o for safety vulnerabilities and harmful behaviors. Frontier risk evaluations use frameworks to assess serious AI risks and societal preparedness. Mitigations are designed and tested to reduce risks related to misinformation and negative human impact. External Red Teaming as a Safety Experiment External red teaming is a method where independent experts test GPT-4o for potential weaknesses or risks. These tests simulate various scenarios to identify if the AI might produce harmful outputs or misinformation. This experimental approach helps reveal limitations and ...

Assessing AI Risks: Hugging Face Joins French Data Protection Agency’s Enhanced Support Program

Image
This analysis is based on the regulatory landscape of the European Union and the French CNIL's action plan as of May 2023. As AI governance frameworks are currently under intense negotiation within the European Parliament, the interpretations of data protection law regarding Large Language Models (LLMs) are subject to immediate and significant changes. This content does not constitute legal advice and may not reflect later domestic or international legislative updates. The rapid growth of artificial intelligence (AI) technologies raises urgent questions about knowledge reliability, privacy, and accountability. As foundation models and their “tool ecosystems” move into everyday products, data protection concerns increasingly sit alongside traditional safety concerns: how data is collected , how outputs are generated , and how individuals can exercise their rights when automated systems shape information and decisions. TL;DR Hugging Face has been selected ...