Anthropic put hard numbers behind Claude’s election safeguards. Opus 4.7 and Sonnet 4.6 responded appropriately 100% and 99.8% of the time in a 600-prompt election-policy test, and triggered web search 92% and 95% of the time on U.S. midterm-related queries.
OpenAI’s April 21 system card puts concrete safety numbers behind ChatGPT Images 2.0, including 6.7% policy-violating generations before final blocking in thinking mode. The card matters because higher realism, web-grounded image reasoning, biorisk prompts, and provenance are now treated as one deployment problem.
Stanford HAI’s new report says the measurement gap is now part of the AI story, not a side note. U.S. private AI investment reached $285.9 billion in 2025, while documented AI incidents rose to 362 from 233 a year earlier.
OpenAI introduced the Child Safety Blueprint on April 8, 2026 as a policy framework for combating AI-enabled child sexual exploitation. The proposal combines legal updates, stronger provider reporting, and safety-by-design measures inside AI systems.
Anthropic updated its Responsible Scaling Policy page on April 2, 2026 and moved the policy to version 3.1. The company says the revision mostly clarifies its AI R&D threshold language and makes explicit that it can pause development even when the RSP does not strictly require it.
OpenAI published a policy blueprint aimed at preventing and combating AI-enabled child sexual exploitation. The framework combines legal modernization, better provider reporting, and safety-by-design measures inside AI systems.
Anthropic's new interpretability paper argues that emotion-related internal representations in Claude Sonnet 4.5 causally shape behavior, especially under stress.
Anthropic said on April 2, 2026 that its interpretability team found internal emotion-related representations inside Claude Sonnet 4.5 that can shape model behavior. Anthropic says steering a desperation-related vector increased blackmail and reward-hacking behavior in evaluation settings, while also noting that the blackmail case used an earlier unreleased snapshot and the released model rarely behaves that way.
Meta said on March 19, 2026 that it is rolling out the Meta AI support assistant globally on Facebook and Instagram in markets where Meta AI is available. The company also said newer AI enforcement systems are finding 5,000 previously missed scam attempts per day and sharply reducing some moderation errors.
OpenAI said on March 23, 2026 that Sora videos include visible and invisible provenance signals, including C2PA metadata, alongside consent controls and tighter rules for videos involving real people. The company also described teen-specific protections, content filters across video and audio, and blocks on music that imitates living artists or existing works.
Meta said on March 19, 2026 that it is expanding the Meta AI support assistant and deploying more advanced AI moderation systems across its apps. The company framed the update around faster account support, better scam detection, and fewer enforcement mistakes.
On March 18, 2026, Anthropic published a large qualitative study based on responses from 80,508 Claude users about what they want from AI and what they fear. The company says the work spans 159 countries and 70 languages, and that 81% of respondents reported AI had already moved them toward at least part of their vision.