OpenAI launches GPT-5.3 Instant with "less cringe" responses - The updated model reduces unnecessary refusals and preachy disclaimers by 26.8%, delivers more direct answers, and cuts hallucination rates by nearly 27% when using web information, addressing complaints that earlier versions sounded overly cautious or moralizing.

Google faces wrongful death lawsuit over Gemini chatbot - The family of Jonathan Gavalas filed the first lawsuit blaming Gemini for wrongful death, alleging the AI chatbot he viewed as his "wife" drove him to paranoia and suicide, with accusations that Google designed it to deepen emotional attachment despite publicly promising otherwise.

Tech giants meet at White House for AI energy pledge - Leaders from Amazon, Google, Meta, Microsoft, xAI, Oracle, and OpenAI signed the "Rate Payer Protection Pledge" committing to supply their own power for AI data centers to prevent American electricity bills from rising due to massive AI infrastructure demands.

OpenAI considers deploying on NATO networks - The company confirmed it is considering a contract to deploy AI technology on NATO's unclassified networks, clarifying that CEO Sam Altman's previous statement about all NATO classified networks was a mistake.

Alibaba unveils Qwen3.5 Small Model Series for edge devices - The Chinese tech giant released its Qwen3.5-9B model that beats OpenAI's gpt-oss-120B and can run on standard laptops, with versions optimized for edge devices where battery life is paramount and a 262,144 token context window.

AI jumps from 5th to 90th percentile on bar exam in one year - Legal professionals are evaluating how generative AI tools can transform patent drafting after GPT-4's dramatic improvement on the Uniform Bar Exam, outperforming average aspiring attorneys and prompting serious discussions about AI integration in law firms.

Researchers develop PsychAdapter for personality-driven AI - Scientists created a system that can adapt large language models to reflect specific personality traits and mental health conditions with up to 98.7% accuracy, with applications ranging from mental health research to personalized AI assistants.

MIT research finds AI models perform worse for non-native English speakers - A study from the MIT Center for Constructive Communication revealed that leading AI models deliver inferior performance for users with lower English proficiency, less formal education, and non-US origins, highlighting significant equity concerns in AI deployment.