{
  "id": 2577604,
  "title": "New York 8th grader tests AI for stress; basic model beats ChatGPT-4o",
  "url": "https://urgent.news/2026/08/22/new-york-8th-grader-tests-ai-for-stress-basic-model-beats-chatgpt-4o",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-22T12:24:03.000Z",
  "source": {
    "name": "Times of India",
    "slug": "times-of-india",
    "url": "https://timesofindia.indiatimes.com/world/us/meet-zeynep-demirbas-the-new-york-eighth-grader-who-tested-whether-ai-can-recognise-stress-a-basic-machine-learning-model-beat-chatgpt-4o/articleshow/133399352.cms"
  },
  "original_language": "en",
  "account": "Fifteen-year-old Zeynep Demirbas from East Amherst, New York, has discovered that a stress-detection AI model called MentalBERT produces more accurate results than ChatGPT-4o and other large language models (LLMs). For her project \"Evaluating the reliability of Large Language Models for stress detection,\" Zeynep analyzed more than 3,500 Reddit posts tagged by humans as stress-related or not. She tested four models: MentalBERT, a specialized version of BERT; BERT; Random Forest, a basic machine learning algorithm; and ChatGPT-4o. MentalBERT scored 82% in identifying stressed posts, followed by BERT at 79%. ChatGPT-4o scored 74%, and Random Forest, as a simpler baseline, scored the lowest. The results surprised Zeynep, leading her to question the reliability of LLMs for mental health assessment, suggesting they should not replace human therapists.",
  "summary": "Fourteen-year-old Zeynep Demirbas tested four AI models on 3,553 Reddit posts to see how accurately they could detect stress. MentalBERT performed best at about 82%, while ChatGPT-4o scored about 74%. Her findings suggest general-purpose LLMs may not be reliable enough for mental health assessment.",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}