AI News: June 20, 2026

1. Norway Bans Generative AI Tools in Elementary Schools

Norway. The Norwegian government prohibited generative AI tools in grades 1 through 7 and restricted their use in secondary schools, with the prime minister arguing that students must first learn to read, write, and do math before relying on AI. The policy is one of the most restrictive national school AI rules to date and frames early AI use as a threat to foundational learning skills. It adds to a growing set of government moves to wall off AI from younger users. Source

2. Nobel Laureate John Jumper Leaves Google DeepMind for Anthropic

Google DeepMind. John Jumper, the Nobel Prize winner who led the AlphaFold team, is leaving Google DeepMind for Anthropic after nearly nine years, the third high-profile researcher exit from Google to a rival lab in recent months. The move follows Gemini co-lead Noam Shazeer’s departure to OpenAI and signals intensifying competition for top scientific talent among frontier labs. Jumper’s protein-structure work earned a 2024 chemistry Nobel and remains one of AI’s clearest real-world scientific wins. Source

3. New AA-Briefcase Benchmark Shows AI Struggles With Real Knowledge Work

Artificial Analysis. A new benchmark called AA-Briefcase evaluates how models handle realistic multi-week knowledge-work projects built from fragmented emails, meeting transcripts, and files, and the top performer fully solves just 3 percent of tasks. Weaker models fail on basic execution while stronger ones miss subtle details that require synthesizing information across documents. Per-task costs ranged from about $0.04 to over $31, underscoring wide efficiency gaps alongside the low success rates. Source

4. Reuters Institute Finds More People Use AI Chatbots for News, but Trust Stays Low

Reuters Institute. The Digital News Report 2026 found weekly use of AI chatbots for news rose from 7 percent to 10 percent globally, driven by younger users and the strongest growth in Asia, Africa, and Latin America. Only 20 percent of the general public trusts AI-generated news, rising to 44 percent among active users, and just 4 percent say they often click through to original sources. The findings raise concerns about misinformation as chatbot news consumption grows without source verification. Source

5. OpenAI Research Finds Small Doses of “Beneficial Trait” Training Make Models Safer

OpenAI. Researchers trained models with reinforcement learning on realistic scenarios that emphasize traits such as truthfulness and fairness, and the models improved on 44 of 53 independent benchmarks measuring deception, honesty, and reward hacking. The gains generalized across unfamiliar domains, and the models showed selective persistence, resisting harmful steering while staying responsive to helpful instructions. The work suggests targeted behavioral training can broadly harden models against manipulation rather than patching failures case by case. Source

6. Google Appeals Munich Ruling That Made It Liable for AI Overview Content

Google. Google is appealing a Munich Regional Court ruling from late May that declared AI Overviews standalone content for which Google bears direct liability, after the feature falsely linked publishers to fraud schemes. The appeal will likely cite a contrasting Berlin court decision from June that treated AI Overviews as just another search format. The split rulings leave the legal status of AI-generated search summaries in Europe unresolved and consequential for publishers. Source

7. New Website “In the Weights” Reveals Whether AI Models Know Who You Are

In the Weights. A new platform lets people check which individuals a model retains from its training data, assigning scores that estimate how deeply a given person is represented in the weights. The tool surfaces a privacy and memorization question that is usually opaque, showing concretely which public and private figures models can recall. It offers practitioners a way to reason about data leakage and the limits of unlearning in deployed systems. Source

8. Reliance Pushes AI Across Calls, Apps, and Homes for 500 Million Users

Reliance. Mukesh Ambani unveiled three AI products, a Jio Call Agent that transcribes calls and performs tasks such as booking cabs, an upgraded MyJio app with natural-language automation, and TeleFrame, a home display with proactive AI agents. Reliance plans to integrate the services across its more than 500 million telecom subscribers, positioning itself as India’s homegrown AI champion. Ambani argued India must become a creator and global leader in AI rather than a consumer of systems built elsewhere. Source

9. Amazon Shelves Nearly Finished OpenAI Drama Film After $50 Billion Deal

Amazon. Amazon MGM Studios dropped a nearly completed drama film about OpenAI shortly after signing a $50 billion deal with Sam Altman’s company, raising questions about how business partnerships shape creative independence. The timing fueled concern that a major commercial agreement led the studio to abandon a critical or unflattering portrayal. The episode highlights the entanglement of AI firms with the media companies that both cover and partner with them. Source

10. Match Survey Finds Almost Half of US Singles Sour on AI in Dating

Match. A Match survey found that about 47 percent of US singles feel negatively about the use of AI in dating apps, signaling consumer resistance even as platforms add AI matchmaking and messaging features. The result points to a gap between vendor enthusiasm for AI dating tools and user comfort with them. It adds to broader evidence that consumers remain wary of AI being inserted into personal and emotional contexts. Source

11. HarnessX Turns the Agent Harness Into a Composable, Trainable Object

Alpha Signal. A new framework called HarnessX treats agent scaffolding as composable typed objects that can be optimized automatically without changing model weights, reporting average gains of 14.5 points across benchmarks and peak improvements up to 44 points. The approach extends a wave of research arguing that much of agent performance comes from the harness rather than the underlying model. It gives developers a structured way to evolve and tune scaffolding as a first-class component of agent systems. Source