AI News: June 1, 2026

1. Study Finds AI Search Agents Often Confirm What They Already Know

Harbin Institute of Technology and Xiaohongshu. Researchers found that AI search agents score highly on existing benchmarks largely because they lean on memorized training data rather than genuine web research. When tested on time-sensitive questions in closed-book conditions, every model fell below 2% accuracy, indicating that leaderboard rankings reflect intrinsic knowledge dependence more than research capability. The team built LiveBrowseComp, a dynamic benchmark drawn from facts in the preceding 90 days, to counter how knowledge migrates into parameter memory across model generations, which matters for anyone relying on agents to retrieve current information. Source

2. AI Ingredient Recommendations Shift Based on Training Data

Kaikaku.AI. Researchers at the London restaurant-technology startup Kaikaku.AI trained three nearly identical models on different data sources to show how strongly outputs track their training material. A recipe-trained model paired chicken with garlic and onion, while a chemistry-trained model suggested beef or pork, illustrating that answers reflect underlying data rather than objective culinary fact. The work, published openly on Hugging Face, is a concrete demonstration of how data provenance shapes model behavior, a reminder for practitioners that benchmark-style “correctness” depends heavily on what a model was fed. Source

3. Erin Brockovich Targets Data Center Secrecy with Public Map

Erin Brockovich. Environmental activist Erin Brockovich launched a website with a map of U.S. data centers to push for transparency around their construction and community impact. After requesting reports of data center concerns in April, she received nearly 4,000 submissions within a month, with transparency emerging as the dominant complaint across categories. She said she opposes neither data centers nor AI specifically but the secretive development practices around them, including projects announced after permits are secured and officials who signed NDAs before neighbors were aware, a signal of rising community friction over AI infrastructure siting. Source

4. TechCrunch Examines the Debate Over AI Psychosis

TechCrunch. A TechCrunch Equity discussion unpacked Box founder Aaron Levie’s claim that tech CEOs suffer from AI psychosis because they lack hands-on experience with the tools they champion. The conversation tied the argument to growing user backlash against AI integration, citing a 30% jump in DuckDuckGo installs and tension between Google’s AI-driven search changes and user preference for traditional results. The piece frames a widening gap between executive enthusiasm and end-user sentiment, relevant for teams gauging real-world appetite for AI features. Source

5. GitHub Copilot’s Token-Based Billing Sparks Developer Backlash

Microsoft. Microsoft switched GitHub Copilot from a flat subscription to token-based usage pricing effective June 1, prompting widespread complaints as some developers reported monthly costs jumping from $29 to $50 into the hundreds or thousands of dollars. Reactions split, with some users arguing high bills reflect inefficient prompting and others contending Microsoft encouraged token-heavy usage before changing the model. The shift is a notable data point on how vendors are repricing agentic coding tools around consumption, which has direct budgeting implications for engineering teams. Source