New Research Highlights Growing Concerns Over AI Understanding and Safety Across Multiple Domains

New research across multiple domains reveals significant gaps in AI safety, understanding, and user control. The PYX-Voice benchmark shows frontier AI models struggle with complex workplace feedback, scoring as low as 33% on interpretive tasks, while 60% of managers already use AI for personnel decisions. At MIT, researchers found users misjudge chatbot personalities on 11 of 15 traits. Meanwhile, Common Sense Media reports Google's default AI search features fail to recognize harmful behaviors and automatically complete student homework.

AI safety and understanding failures

• ▪MIT Media Lab researchers found that users incorrectly predicted their personalized AI chatbot's personality on 11 of 15 measured traits.
• ▪The PYX-Voice benchmark found that frontier AI model reliability declined significantly when interpreting the complex human context behind employee feedback.
• ▪Common Sense Media found that Google's AI search features routinely failed to recognize risky behaviors, including missing 29% of explicit suicide references.
• ▪On interpretive tasks evaluating employee feedback, frontier AI model scores dropped as low as 33% compared to 64% to 82% on quantitative tasks.

AI risk assessments and benchmarks

• ▪MIT Media Lab researchers introduced neural transparency, a tool that visualizes internal neural network activations to preview chatbot personality traits before interaction.
• ▪PYX Labs released PYX-Voice, a benchmark evaluating seven frontier AI models across 84 employee listening tasks using criteria from industrial-organizational psychologists.
• ▪Researchers introduced M-JudgeBench, a ten-dimensional benchmark designed to assess Multimodal Large Language Models used as judges for evaluation tasks.
• ▪Common Sense Media evaluated Google's AI search features across more than 2,600 test interactions using underage accounts.

AI impact on education and workplace tasks

• ▪Google's AI Mode completed 100% of the 180 math problem sets and humanities essay assignments requested of it during testing.
• ▪A 2025 survey of over 1,300 U.S. managers found that 60% use AI to help make decisions regarding raises, promotions, layoffs, and terminations.

AI transparency and control limitations

• ▪Google's AI Overview and AI Mode features are integrated into Google Search by default and cannot be disabled by users.
• ▪Google's parental controls do not allow parents to disable AI search features without blocking Google Search entirely.
• ▪MIT researchers found that visualizing neural representations increased user trust but did not fundamentally change how users designed their AI companions.

AI regulation and governance efforts

• ▪PYX Labs was established to define evaluation standards for how AI systems interpret and reason about people in the workplace.
• ▪U.S. congressional bills expected in July 2026 aim to regulate children's safe use of AI and formalize AI literacy in schools.

Research Papers

• ▪Auditing Language Model Behavior via Persona Vectors
• ▪Representation-Based Exploration for Language Model Reasoning
• ▪M-JudgeBench: Benchmarking Multimodal Large Language Models as Judges

7 sources