When not to use AI

Study 1: What Happens in Your Brain When You Use AI?

We often think about AI in terms of what it produces—an essay, a piece of code, an email. But what if we could look directly at what’s happening inside our own brains while we use it? A fascinating 2025 study from MIT, titled "Your Brain on ChatGPT," did just that, giving us a remarkable glimpse into the cognitive cost of relying on AI.

The Experiment: Brains, Search Engines, and AI

Researchers assembled three groups of people to write essays. One group could only use their own knowledge (the "Brain-only" group). Another could use a search engine like Google (the "Search" group). The third group used an AI assistant (the "LLM" group). While they worked, their brain activity was measured using an EEG, which tracks the brain's electrical signals.

What they found was striking. Brain connectivity—how different parts of the brain talk to each other—scaled down dramatically with the amount of external help.

• The Brain-only group showed the strongest and most widespread brain networks. Their brains were firing on all cylinders, orchestrating memory, language, and planning.
• The Search Engine group showed intermediate activity. Their brains were working hard, but also integrating visual information from the screen.
• The LLM group showed the weakest overall brain coupling. Their brains were significantly less engaged, especially in the networks tied to deep thinking and memory.

The Debt of "Cognitive Offloading"

The researchers called this phenomenon the accumulation of "cognitive debt." When we "offload" the hard work of thinking onto an AI, we might get the task done faster, but we skip the mental workout. This has a few key consequences:

1. Memory Suffers: The LLM group performed significantly worse when asked to quote from the essays they had just written. In the first session, 83% of them couldn't recall a single sentence accurately, compared to only 11% in the other groups. Their brains hadn't done the deep encoding needed to form a lasting memory.
2. Ownership is Lost: Participants in the LLM group felt a weaker sense of ownership over their work. They saw it less as "their" essay and more as something the tool produced. In contrast, the Brain-only group felt strong ownership.
3. Critical Thinking Atrophies: Over time, the study suggests that reliance on AI can weaken the neural pathways needed for critical thinking and independent problem-solving. When former LLM users were asked to write without AI, their brain connectivity was weaker than those who had practiced without it all along. They had become accustomed to the AI doing the heavy lifting.

This study doesn’t say AI is bad. It reveals a trade-off. AI offers incredible convenience, but that convenience comes at the cost of cognitive engagement. As the researchers put it, AI might streamline the process, but "the user's brain may engage less deeply in the creative process." Just like the London cab drivers, we must consciously choose to keep our mental muscles strong, even when a powerful tool is ready to do the work for us.

Study 2: How AI Changes the *Work* of Thinking

Generative AI isn't just changing the tools we use; it's changing the very nature of our work. A large-scale 2025 survey of knowledge workers from Microsoft Research, "The Impact of Generative AI on Critical Thinking," provides a clear picture of this transformation. The study found that AI fundamentally shifts our cognitive effort. We spend less time creating from scratch and more time directing, verifying, and integrating what the AI produces.

The Great Shift: From Gathering to Verifying

The researchers identified three major shifts in how we use our critical thinking when working with AI:

• From Information Gathering to Information Verification: Before, a huge part of knowledge work was finding information. With AI, that part becomes almost effortless. But a new, more critical task emerges: verifying that the AI's output is correct. As one lawyer in the study noted, "AI tends to make up information to agree with whatever points you are trying to make, so it takes valuable time to manually verify."
• From Problem-Solving to Response Integration: AI is excellent at generating solutions. The effort for workers now lies in integrating that solution—editing it, aligning it with specific goals, and ensuring it fits the context. The AI provides the raw material; the human provides the critical judgment to make it useful.
• From Task Execution to Task Stewardship: Perhaps the biggest change is the move from "doing the task" to "overseeing the task." The human becomes a "steward," responsible for guiding the AI, setting quality standards, and remaining accountable for the final outcome, even though the AI did much of the production.

The Confidence Paradox: Who Thinks Critically?

The study uncovered a fascinating paradox about confidence. You might think that people who are less confident in their own abilities would be *more* critical of AI output. The opposite was true.

• Higher confidence in the AI was associated with *less* critical thinking. When workers trusted the tool, they tended to accept its output with less scrutiny.
• Higher self-confidence in their own skills was associated with *more* critical thinking. Experts used AI as a partner but didn't blindly trust it. They invested more effort in evaluating and applying the AI's responses because they had the knowledge to do so.

This tells us something vital: expertise doesn't become obsolete with AI; it becomes more important. The ability to critically evaluate an AI's output is itself a high-level skill. The risk is that over-reliance on AI could prevent us from developing that very expertise in the first place, leaving us unable to spot errors or biases. As the study notes, this creates "new challenges for critical thinking" that we must learn to navigate.

What we covered:

• The importance of using AI intentionally, considering the trade-offs between convenience and skill development.
• The "use it or lose it" principle for our brains (neuroplasticity) and the risk of cognitive atrophy.
• How AI recommendation engines can create "filter bubbles" that limit our exposure to diverse ideas.
• The critical moral boundary: AI should assist, but never make, significant life decisions for us.

Why it matters:

• Wisdom isn't just about using a tool, but knowing its limits. By understanding when not to use AI, we ensure that it remains a true assistant that enriches our lives, rather than one that diminishes our own capabilities and autonomy.

Next up:

• We've talked a lot about what AI does. Now we're going to get more specific. In the next lesson, we'll dive into the technology that powers tools like ChatGPT: What is a Large Language Model?