{
  "slug": "knowing-isnt-understanding",
  "title": "Knowing Isn’t Understanding: AI Is Handing Out Answers Faster Than People Can Make Sense of Them",
  "url": "https://emberandoakinstitute.com/blog/knowing-isnt-understanding.html",
  "date": "2026-10-25",
  "category": "AI news",
  "tags": [
    "artificial intelligence",
    "education",
    "learning",
    "AI in science",
    "protein design",
    "mathematics",
    "game design",
    "AI news"
  ],
  "summary": "A CEPR working paper on 26,811 Chinese secondary students found AI raised homework scores but lowered closed-book and entrance-exam scores; Anthropic's Claude generated 318,484 scorpion-toxin binder designs, of which 17 passed lab binding tests and none worked against multiple toxins; Steven Strogatz and the Association for Human Mathematics show mathematicians weighing AI-solved problems against human understanding; Epic's AI MegaJam rewards dev logs that show how a game was built.",
  "key_facts": [
    {
      "fact": "Strömberg, Lei and Wu tracked 26,811 students in grades 7-12 in one county in central China for 30 months across nine subjects; about 80% had adopted AI (mostly Doubao and DeepSeek) by June 2025.",
      "sources": [
        "https://cepr.org/publications/dp21577",
        "https://cepr.org/voxeu/columns/generative-ai-learning-penalty-secondary-school"
      ]
    },
    {
      "fact": "After AI adoption, homework scores rose 18% and homework time fell from about 64 to 45 minutes (about 30%); monthly closed-book exam scores fell 20% within six months.",
      "sources": [
        "https://cepr.org/voxeu/columns/generative-ai-learning-penalty-secondary-school",
        "https://www.yanhuiwu.com/documents/AI_learning_penalty.pdf"
      ]
    },
    {
      "fact": "High-school and college entrance exam scores fell 24% and 18%, with full effects after about two years; losses were about 27% in social science, 22% STEM, 17% English and 9% Chinese, and largest for younger students, boys and high achievers.",
      "sources": [
        "https://cepr.org/voxeu/columns/generative-ai-learning-penalty-secondary-school",
        "https://www.yanhuiwu.com/documents/AI_learning_penalty.pdf"
      ]
    },
    {
      "fact": "About four in five AI users behaved consistently with outsourcing homework; AI users who spent the same homework time as non-users learned as much. The paper is not peer reviewed and the authors caution against extrapolating effect sizes.",
      "sources": [
        "https://cepr.org/voxeu/columns/generative-ai-learning-penalty-secondary-school",
        "https://cepr.org/publications/dp21577"
      ]
    },
    {
      "fact": "An internal Anthropic research model, working from a prompt plus four follow-ups over 64 hours with 12 design tools, generated 318,484 protein designs against five scorpion neurotoxins; 180 were selected by Claude and 31 added by the team with Claude's help (211 tested), using on the order of 10,000 GPU hours.",
      "sources": [
        "https://www.anthropic.com/research/claude-scorpion-toxin-binder-design",
        "https://www-cdn.anthropic.com/31ac578806cffd8b3d92d8affe6edc4c04d502e1.pdf"
      ]
    },
    {
      "fact": "Adaptyv Bio and Twist Bioscience lab tests found 17 designs met Anthropic's binder criteria, each against one toxin; none of 151 multi-toxin designs met criteria for more than one toxin. Neutralization has not been tested and Anthropic says it does not claim an antivenom.",
      "sources": [
        "https://www.anthropic.com/research/claude-scorpion-toxin-binder-design",
        "https://www-cdn.anthropic.com/31ac578806cffd8b3d92d8affe6edc4c04d502e1.pdf"
      ]
    },
    {
      "fact": "Scorpions sting 1.2 million or more people a year, causing more than 3,250 deaths, mostly children in places with limited medical care (per Anthropic).",
      "sources": [
        "https://www.anthropic.com/research/claude-scorpion-toxin-binder-design"
      ]
    },
    {
      "fact": "Outside experts said binding is no longer the main bottleneck but that translating a binder into a medicine is the hard part.",
      "sources": [
        "https://www.anthropic.com/research/claude-scorpion-toxin-binder-design",
        "https://www.theverge.com/ai-artificial-intelligence/1016443/anthropic-claude-citrini-vibecoding-scorpion-antivenom"
      ]
    },
    {
      "fact": "Steven Strogatz wrote in The Atlantic (Oct. 25, 2026) that math is 'in crisis' after OpenAI's Oct. 6 report of close to 400 solved open problems; Cornell logician Justin Moore will open a cognac saved since 2010 only once trusted peers agree the AI proof is correct and understand it.",
      "sources": [
        "https://www.theatlantic.com/ideas/2026/10/ai-math-human-future/689133/"
      ]
    },
    {
      "fact": "The Association for Human Mathematics lists 1,008 members pledging not to work with commercial AI companies or publish AI-generated math; 654 are in an AI-free caucus that does not use AI in research.",
      "sources": [
        "https://www.ahmath.org/members",
        "https://www.newscientist.com/article/2593621-a-thousand-human-mathematicians-pledge-to-shun-ai-in-their-research/"
      ]
    },
    {
      "fact": "Epic's 2026 AI MegaJam (Nov. 12-15, open to developers 18+, AI optional) includes a Glass Box category for the best dev log showing how a game was built, and advises entrants to keep notes as they go.",
      "sources": [
        "https://www.unrealengine.com/news/2026-epic-ai-megajam",
        "https://itch.io/jam/2026-epic-ai-megajam"
      ]
    }
  ]
}