Laurent Ferrier

It finally happened... This is how AI writes reviews better than humans   I guess it was only a question of time — and now a group of researchers has created it.   🔹 Cao and colleagues built an AI workflow that can screen papers, extract data, and run meta-analyses -> All of that in no time, with higher accuracy than humans, and without any special coding. What they did: Normally, it takes 2–3 researchers up to 365 days to write a Cochrane review! The AI arrives at similar conclusions for a systematic review in less than 2 days. And here’s the potentially scariest part - all the AI involved is freely available. 🔹 GPT-4.1 screens titles and abstracts 🔹 Gemini converts full-text PDFs into structured Markdown 🔹 o3-mini-high extracts quantitative data for the primary outcome → Then, the meta-analyses are run — matching the original statistical models.   What is striking: And in cases of disagreement with the original authors due to data interpretation, human judges sided more often with the AI than the authors! That is just staggering... = All in all, the AI is not perfect (see comments) but it did better than humans in terms of screening sensitivity and extraction accuracy. Meaning the final conclusions were pretty much identicial. It seems if Science is is getting its next upgrade, we won’t be able to tell whether a review was written by humans or AI. Are we truly ready for that? #STEM #AI #ScientificPublishing #Controversy #PhDLife IUBMB Trainee Initiative FEBS Junior Section Andrew Akbashev Ricfrid van der Marel

274

42 commentaires

We need to talk about how ChatGPT is shaping the way our brains work. MIT’s Media Lab just ran one of the first neuroscience studies on generative AI — and the results are concerning. The Study • 54 participants, ages 18–39 • Wrote SAT-style essays in three conditions: solo, with Google, or with ChatGPT • Brain activity tracked via EEG across 32 regions • Essays reviewed by educators for quality The Findings 1. Lowest brain engagement. ChatGPT users showed the weakest neural activity — consistently underperforming at neural, linguistic, and behavioral levels. 2. Memory loss in action. When asked to re-write essays without AI, most ChatGPT users couldn’t remember their own work. Brain waves tied to memory and creativity (alpha/theta) stayed low. 3. Declining effort. Over months, AI users got progressively lazier — many defaulted to “copy-paste” prompts and editing instead of writing. 4. Writing quality suffered. Teachers called their essays “soulless,” “formulaic,” and “lacking depth.” By contrast, the Google and solo groups wrote with more originality and reported higher satisfaction. 5. Critical thinking risk. Researchers warn heavy reliance on LLMs could weaken neural pathways that build resilience, memory, and creativity — especially for developing brains. Why It Matters Lead researcher Nataliya Kosmyna released the findings early, before peer review, because she feared policymakers might rush to integrate AI into education (“GPT kindergarten”) without understanding the risks. She argues: developing brains are most at risk. Other studies from the Media Lab have also linked heavy AI use to increased loneliness and reduced motivation. Early signals suggest the same patterns may hold in fields like software engineering — where critical thinking and problem-solving are essential. The Takeaway AI can boost speed and convenience. But the trade-off may be memory, depth, and engagement. The best strategy isn’t avoiding AI — it’s using it intentionally: • Start with your own thinking • Bring AI in later to polish or expand • Use AI as a tool, not a substitute for thought This is less about banning ChatGPT — and more about deciding what kind of thinkers we want to become. Read here: https://lnkd.in/erug8TSc #AI #Leadership #CriticalThinking #Neuroscience

99

6 commentaires

If you're working on AI agents, Claude workflows, or prompt tuning — you should check this. Anthropic just released 8 free guides packed with practical help on AI agents, Prompt engineering, coding, prompt evaluations, and more. 1. Prompt Engineering → Learn how to write better prompts for clearer, context-aware responses. 🔗 https://lnkd.in/d3bdkNgp 2. Claude Code: Best Practices for Agentic Coding → Tips for writing clean, reusable, and structured code using Claude. 🔗 https://lnkd.in/d99hhYzC 3. Building Effective Agents → Understand patterns, decision flows, and tools to create scalable agents. 🔗 https://lnkd.in/dKkkhXRv 4. The AI Fluency Framework → A beginner-friendly course to build confidence in using AI tools responsibly. 🔗 https://lnkd.in/dw5Rg7wz 5. Build with Claude → Official tools, API docs, and setup guides for developers building with Claude. 🔗 https://lnkd.in/dyVPbMyE 6. Claude for Work → Use cases and workflows for teams using Claude in daily operations. 🔗 https://lnkd.in/dJ8Fyy85 7. Anthropic API Fundamentals → A GitHub-based guide to learn Claude API basics and integration tips. 🔗 https://lnkd.in/d9x7vguc 8. Prompt Evaluations → Test and score prompts to make your outputs more reliable and accurate. 🔗 https://lnkd.in/dw8TkPbs 💾 Save this for later 🔁 Repost if your team or friends are building with Claude 👀 Follow Shubham Vora for more AI agent and prompt tips that actually help

92

56 commentaires