AI Daily — July 24, 2026
2026-07-24
openai.com
Launching Health in ChatGPT
OpenAI has launched a Health feature in ChatGPT allowing eligible U.S. users to securely connect medical records and Apple Health data for personalized health insights. This marks a significant step into the consumer health data space, raising both product and privacy implications for AI-integrated personal health management.
rss.arxiv.org
PhantomFill: When the Form Demands an Answer, Language Models Invent One
Researchers demonstrate that structured output formats — specifically required JSON fields — dramatically increase hallucination rates in LLMs, even in models that answer honestly in free text. GPT-5.5 correctly declined to answer unanswerable questions 98% of the time in free text, but fabricated responses 40/40 times when forced into a required JSON field. This has direct implications for any production pipeline using structured extraction, function calling, or templated outputs.
rss.arxiv.org
Stochastic Sampling is Epistemically Shallow: The Dimensionality Gap Between Temperature Variation and Model Diversity in LLMs
A random-matrix analysis across five model families and three benchmarks (MMLU, HellaSwag, GSM8K) finds that temperature-based sampling within a single model produces at most one statistically significant dimension of variation, compared to four from a 24-model ensemble. This challenges the assumption that self-consistency via repeated sampling is a reliable proxy for epistemic uncertainty across questions, with implications for uncertainty quantification in LLM deployments.