AI Flattens Thinking by Decoupling Words From Ideas

The real threat of AI isn't that it writes better than us--it's that it makes us think worse. While AI-generated essays now score higher on creativity metrics and have even won literary prizes, Georgetown neuroscientist Adam Green’s analysis of over 370,000 college admissions essays reveals a disturbing pattern: AI use correlates with sharper prose but blander ideas. The tools we’re adopting to enhance our output are quietly eroding the diversity of human thought. This isn’t just a concern for writers or educators--it’s a warning for anyone whose job involves generating ideas. The advantage goes to those who recognize that creativity isn't about the final product, but the process of thinking itself. Those who outsource the act of creation may win short-term efficiency but lose long-term cognitive resilience and originality. The real competition isn’t between human and AI writing--it’s between those who let AI do the thinking and those who use it without surrendering their minds.

Why Polished Prose Masks Cognitive Atrophy

Here’s the unsettling twist: AI is making student writing look more creative while making it less original. In a study spanning hundreds of thousands of college admissions essays before and after ChatGPT’s 2022 release, external experts rated AI-assisted essays as more creative. The language was richer, the vocabulary more varied, the sentences more sophisticated. By every conventional measure, the writing had improved. But when researchers analyzed the underlying ideas, they found the opposite: the essays had become more similar to each other. The ideas were narrower, more predictable, and clustered around a shrinking set of themes.

This is the core contradiction--AI decouples words from ideas. For centuries, linguistic variety was a reliable proxy for conceptual originality. When a writer used unusual words or complex syntax, it often meant they were thinking in novel ways. But AI breaks that link. It can generate linguistically diverse text without engaging in divergent thinking. The machinery behind large language models doesn’t “think” in associative networks built from lived experience. Instead, it predicts the next word based on statistical patterns in training data--data that overwhelmingly reflects dominant cultural voices. So while the surface-level output appears creative, the conceptual engine is convergent, pulling toward the average.

"The most diverse words expressed the least diverse ideas--the least original ideas."

-- Adam Green

That sentence alone should give anyone who values originality pause. It’s not that AI is bad at writing. It’s that its version of “good” writing is homogenizing. And because that homogenization is masked by stylistic polish, it’s invisible to traditional evaluation methods--even to experts. In Green’s study, 22 creativity experts were fooled. AI detectors were fooled. The system wasn’t just evading detection; it was gaming the very metrics we use to define creativity.

This has immediate implications for education, hiring, and creative industries. If we continue to assess work based on linguistic sophistication alone, we’ll reward outputs that are technically proficient but conceptually sterile. Worse, we’ll incentivize students and professionals to outsource their thinking to tools that, by design, optimize for palatability over originality.

The Hidden Cost of Convergent Thinking at Scale

The danger isn’t just that AI produces similar ideas--it’s that it rewards sameness. Large language models are trained on data that reflects what’s already been published, shared, and validated. They’re further refined through reinforcement learning from human feedback (RLHF), where annotators rate outputs based on clarity, coherence, and safety. The result? A feedback loop that favors writing that’s inoffensive, well-structured, and familiar--what Green calls “the over-lit street.” It’s writing that leaves no potholes, no shadows, no ambiguity. It’s clear, but it’s also bloodless.

And because this style is reinforced across millions of training examples and thousands of human raters, it becomes the dominant mode. When students use AI to write essays, they’re not just getting help--they’re being subtly steered toward a narrow band of acceptable expression. The tool doesn’t just respond to prompts; it shapes them. Over time, this doesn’t just affect individual pieces of writing. It reshapes how people think.

The real kicker? The ideas being squeezed out aren’t bad ones. In fact, Green’s team found that the most conceptually diverse student essays--those that fell “outside the bots”--were associated with higher GPAs and better performance on standardized tests. These weren’t the ramblings of underperformers. They were the outputs of students who thought differently. And they’re precisely the kind of thinkers AI tools discourage.

"Those students that were thinking outside the bots were going on to achieve higher GPAs in college and actually scoring higher on tests."

-- Adam Green

This flips the script on the supposed efficiency of AI. If we define success narrowly--get a good essay, win a prize, pass a class--then AI wins. But if we define success as long-term cognitive development, adaptability, and the capacity to solve novel problems, then outsourcing thinking to AI is a losing strategy. It’s a Faustian bargain: sharper sentences today, duller minds tomorrow.

And the consequences compound. In creative fields, homogenization means fewer breakthroughs. In science, it means more incremental research that rehashes existing ideas rather than challenging them. In society, it means a shrinking pool of truly original perspectives--especially from those whose experiences don’t align with the cultural mainstream embedded in training data.

Green points out that AI models are especially likely to homogenize the voices of people who don’t “look like me and you”--a quiet erasure of diverse ways of thinking masked as linguistic improvement. The tool doesn’t just flatten ideas; it privileges certain kinds of people while making others sound more like the norm.

The 18-Month Payoff Nobody Wants to Wait For

Here’s where competitive advantage hides: in the willingness to endure short-term discomfort for long-term cognitive resilience. Most people use AI to get better output faster. But the real edge goes to those who use it without surrendering the thinking process.

Green’s research suggests that the act of writing isn’t just about producing text--it’s about developing the ability to think. When you write, you’re navigating your own semantic network, making connections between ideas based on your unique experiences. That process builds cognitive flexibility, which pays off in ways that no AI-generated paragraph can replicate.

The problem is, this payoff is delayed. You won’t see it in your next essay. You might not see it in your first job. But over 12--18 months, the gap widens. The person who practiced thinking through writing develops a mental agility that allows them to spot opportunities, solve problems, and innovate in ways that those who outsourced their cognition simply can’t.

This is where conventional wisdom fails. We assume that using AI to improve output is always beneficial. But Green’s work shows that when the process is outsourced, the long-term cost is atrophy. It’s like using a forklift to carry every object so you never build muscle. In the moment, you’re more efficient. Over time, you become weaker.

The system responds. Students who rely on AI for essays may get better grades initially. But when they hit coursework that demands original thought--research, problem-solving, synthesis--they’re at a disadvantage. The tool that helped them early on has hollowed out the very capacity they now need.

And here’s the irony: the people best positioned to benefit from AI are those who resist its most seductive promise. They use it for editing, brainstorming, or fact-checking--but they keep the core act of thinking for themselves. They accept that their first drafts may be clunkier, their language less polished, but their ideas more original. Over time, that’s what matters.

Key Action Items

• For students: Use AI as a collaborator, not a substitute. Over the next quarter, experiment with writing first, then using AI to refine. Keep a log of how often you generate ideas independently versus prompting AI. The goal isn’t to avoid AI--it’s to ensure you’re adding idea value that couldn’t come from the model alone.

• For educators: Stop trying to detect AI use. Over the next semester, redesign assignments to assess conceptual distinctness, not just linguistic quality. Ask students to explain their thinking process, trace their idea development, or compare early drafts with final versions. The metric shifts from “Is this AI?” to “Did you think for yourself?”

• For professionals: In the next 6 months, conduct a personal audit of your AI use. How much of your writing, research, and synthesis is outsourced? Identify one high-leverage task (e.g., strategy memos, creative pitches) where you’ll commit to doing the thinking first, even if it’s slower. This pays off in 12--18 months as your cognitive edge sharpens.

• For teams: Over the next quarter, establish a “thinking-first” norm. Require that drafts begin as unpolished, idea-rich documents before any AI refinement. Reward originality and risk-taking, not just clarity and polish. This creates a culture where AI enhances human creativity instead of replacing it.

• For anyone: Recognize that writing is thinking. If you’re not uncomfortable sometimes--if your ideas always flow smoothly, your sentences always balanced--ask whether you’re truly engaging your own mind. Discomfort now is the price of cognitive resilience later. Pay it.