Science Vs

Is AI Making Us Stupid?

December 18, 2025

Key Takeaways Copied to clipboard!

  • Research suggests that learning guided by LLMs like ChatGPT results in sparser, more generic output and a lower perceived sense of learning compared to traditional web searches, transforming learning into a more passive process. 
  • The use of AI tools may lead to 'de-skilling,' where reliance on the technology causes a measurable decline in users' independent abilities, as evidenced by a study on doctors using AI for colonoscopy analysis. 
  • AI can enhance productivity by automating rote tasks, potentially freeing up time for deeper analysis, but studies indicate a significant portion of this saved time is often wasted rather than reinvested in more valuable work. 

Segments

Initial AI Fears and Usage Stats
Copied to clipboard!
(00:00:00)
  • Key Takeaway: Widespread adult interaction with AI (62% several times a week) contrasts with high student usage (around 80% for schoolwork), fueling panic about cognitive decline.
  • Summary: The episode opens by framing the central conflict: claims that ChatGPT makes people stupider versus potential benefits. Statistics show that 62% of US adults interact with AI weekly, and approximately 80% of high school and college students use it for schoolwork. This high adoption rate drives significant public concern regarding brain function and education.
LLM Inaccuracy Concerns
Copied to clipboard!
(00:05:06)
  • Key Takeaway: Early LLMs like ChatGPT performed poorly, sometimes only being correct about half the time, meaning users risk being fed incorrect information.
  • Summary: A key initial problem identified is the inaccuracy of LLMs; early research indicated error rates around 50%. While newer versions show improvement, they are not 100% accurate, making it difficult for users to discern truth from falsehood. This constant potential for misinformation is cited as one way AI could be making users ‘stupid.’
LLMs vs. Web Search Learning Study
Copied to clipboard!
(00:06:23)
  • Key Takeaway: Advice generated after researching with ChatGPT was significantly sparser, more generic, and cited fewer facts than advice generated after using a traditional Google search.
  • Summary: A study by Dr. Shiri Melumad compared learning outcomes when using ChatGPT versus Google search for research tasks. Participants using LLMs produced advice that was less informative and less engaging to external reviewers. Furthermore, those using ChatGPT felt they learned less deeply than those who navigated search links themselves.
Passive Learning and Brain Connectivity
Copied to clipboard!
(00:12:31)
  • Key Takeaway: The synthesis provided by LLMs transforms learning from an active process to a passive one, which correlates with weaker brain connectivity during task completion.
  • Summary: The act of manually processing links and interpreting information is crucial for deep learning, a step skipped when using LLMs. EEG measurements in a small study suggested that brain connectivity was weakest when participants used ChatGPT to write an essay compared to using Google or their own brains. This reduced engagement may also negatively impact memory retention of the generated work.
De-skilling and Professional Impact
Copied to clipboard!
(00:17:51)
  • Key Takeaway: Reliance on AI can cause de-skilling, where professionals become worse at tasks they previously knew how to perform once the AI tool is removed.
  • Summary: The concept of de-skilling was illustrated by a study on doctors performing colonoscopies; after using AI to spot polyps, the doctors performed worse when the AI was taken away than they did before they ever used it. This suggests that outsourcing cognitive tasks can lead to the atrophy of necessary skills, even in critical fields.
AI as Crutch vs. Enhancer
Copied to clipboard!
(00:22:45)
  • Key Takeaway: The ultimate effect of AI depends on user intent: it can serve as a ‘crutch’ leading to stagnation or an ’enhancer’ allowing users to tackle more complex work.
  • Summary: Expert Aaron French posits that AI usage is bifurcated: some users will rely on it to avoid learning, while others will leverage it to handle rote tasks. Studies show AI can save significant time (up to 80% in some cases) for professionals like teachers, but there is no guarantee this saved time is reinvested productively.
Historical Parallel: The Calculator
Copied to clipboard!
(00:27:45)
  • Key Takeaway: Historical data on calculators shows that fears of basic skill loss were largely unfounded; instead, the tool enabled students to engage with and improve at more complex mathematics.
  • Summary: The introduction of calculators in the 1970s sparked similar controversies, with fears of creating ‘calcuholics’ who couldn’t perform basic arithmetic. A meta-analysis of over 50 studies found that basic math skills did not worsen, and problem-solving skills actually improved when calculators were used as a learning tool.
AI Potential in Education and Science
Copied to clipboard!
(00:31:05)
  • Key Takeaway: AI shows potential in education by acting as a personalized tutor or providing writing feedback, and broadly in science for tasks like protein folding and data analysis.
  • Summary: Research indicates that customized chatbots can assist students with learning disabilities, and using ChatGPT as a math tutor can be as effective as using a textbook. Furthermore, broader AI applications, like machine learning in biochemistry and physics, are becoming essential for solving problems beyond human capacity, with scientists expecting increased importance over the next decade.