ChatGPT’s divide isn’t IQ, it’s how you use it

The argument that “less intelligent people are more impressed by ChatGPT” is a tempting dunk. It is also lazy. What ChatGPT elicits in us depends less on raw IQ and more on how we calibrate trust, how we choose use cases, and whether we keep our critical faculties engaged. Those are skills. They can be taught.

In one survey of knowledge workers, participants said that for 40% of tasks they used no critical thinking at all when relying on AI assistants (Microsoft/Carnegie Mellon via Live Science).

That finding captures the real divide. Not smart versus gullible, but calibrated versus uncalibrated use. People who treat a language model like an oracle tend to be dazzled by its fluency. People who treat it like a fallible collaborator keep the benefits while containing the damage.

Large language models are startlingly good at producing paragraphs that look and feel right. That is the point. The human brain is tuned to reward coherent language, so we routinely mistake verbal confidence for accuracy. Psychologists have a name for this anthropomorphic reflex: the ELIZA effect.

AI fluency sits on top of other human quirks. The Dunning–Kruger effect encourages novices to overestimate their grasp of a topic; a model that sounds certain can seal the illusion. Add automation bias—our tendency to over-trust computerized recommendations—and you get a powerful recipe for overconfidence in wrong answers.

When the Reddit poster asked ChatGPT to list “all the great extinction events” and relate them to climate variables, the model’s omissions were predictable. The training distribution overemphasizes the well-known “Big Five,” downplays pre-Cambrian crises like the Great Oxygenation Event, and rarely encodes the quantitative nuance to align extinction pulses with geochemical proxies. That is not because the model is stupid; it is because it was not designed to be a vetted database. It is, as critics put it, a very capable word predictor. On topics where truth requires retrieval, citations, and methods, a general model will miss things unless you force it to search, cite, and show its work.

A 2025 study linked frequent AI use with lower critical thinking scores (r = −0.68), arguing that heavy reliance encourages cognitive offloading and shallow scrutiny (Phys.org) and Big Think.

None of this means the wow factor is a mark of dimness. Experts are often the first to be underwhelmed by LLMs on their home turf, because they instantly spot errors. Yet the same experts can be stunned by the tool’s reach in adjacent domains—drafting scaffolds, translating styles, writing test harnesses, summarizing sprawling threads—because there the benchmark shifts from “be right” to “unblock me.”

What separates the impressed from the unimpressed is usually three habits, none of which require Mensa credentials:

• Picking the right jobs. Use it for ideation, scaffolding, rewriting, tutoring, or converting formats. Be cautious with facts and fresh information unless the system retrieves and cites sources you can check.
• Managing trust explicitly. Ask for sources; verify them. Compare outputs against known-good references. Calibrate confidence like a pilot reading instruments, not a passenger enjoying the view. The NIST AI Risk Management Framework exists for this reason.
• Designing the workflow around the flaw. Break complex prompts into steps; constrain the answer space; supply your own data; force the model to reason and then audit the reasoning. Treat it as a junior analyst, not a judge.

When those habits are missing, two things happen at once. First, the model’s surface polish fools us. Second, we get mentally rusty. Multiple lines of research now warn that offloading too much thinking to AI can erode the very muscle we need to evaluate it. The Microsoft/CMU work presented at CHI found a measurable drop in scrutiny as AI reliance rose, and respondents admitted they often did not think critically at all when using assistants (Live Science). Michael Gerlich’s survey of 666 people reported a similar pattern and singled out cognitive offloading as the mechanism (Big Think).

There is a cultural layer too. Younger users are steeped in software that improves like software does. When a newer model posts eye-popping scores on standardized puzzles, it is easy to slide from “useful” to “self-aware.” Reporting this spring highlighted a survey in which one in four Gen Z respondents said AI is already conscious and many treat chatbots like companions (The Outpost). Intelligence is not consciousness, but fluency makes that distinction feel academic.

Fluency persuades. Persuasion without verification is how misinformation slips past us, whether it is a chatbot’s invented book list or state-backed propaganda seeded into models (PCMag).

The solution is not to sneer at the impressed or to swear off the tool. It is to use it where it amplifies you and to defend the parts of your mind that matter.

Try this reframing:

Use ChatGPT as a language engine, not an oracle. It is brilliant at turning messy intent into structured prose, code, and plans. Have it outline a report from your notes, translate your tone for a different audience, or propose ten ways to test a hypothesis you already care about. For facts, push it into a citing mode, demand URLs, and click them.

Keep the hard parts for yourself. Deciding which questions matter, which tradeoffs to accept, which sources to trust—that is your job. If a model helps you see options faster, great. If it is doing your thinking for you, slow down. Some educators are already weaving this into curricula: critique the output, identify its assumptions, rebuild the argument, then answer on your own.

Audit your own reliance. If you notice the tool sliding from assistant to default brain, course-correct. The studies raising red flags about critical thinking erosion are not destiny. They are warnings about uncritical adoption in moments that do not deserve it.

One last point on the Reddit prompt that started this: comparing “sorting my life out” tasks to scientific synthesis is apples to anvils. A model can excel at habit scaffolding, meal planning, or turning an overwhelming goal into bite-sized steps precisely because those are language problems masked as life problems. Scientific synthesis is a literature and methods problem. Expect different outcomes.

So no, being impressed by ChatGPT does not make you less intelligent. Being impressed without guardrails makes you more vulnerable. The difference is not innate ability. It is discipline, literacy, and the humility to ask the machine to show its work—and the courage to do yours.