The question arrives with a nervous edge because it feels less like a technology question than a human one: Is artificial intelligence making us smarter, or is it making us mentally lazier?
The most honest answer is that it is doing both — depending on how we use it.
Artificial intelligence is already making people faster. It helps programmers write code, students summarize readings, workers draft memos, analysts scan documents, and ordinary people translate, plan, search, brainstorm and troubleshoot. In one controlled experiment, developers using GitHub Copilot completed a programming task 55.8 percent faster than those who did not use it. In a large study involving hundreds of Boston Consulting Group consultants, generative AI improved performance on many tasks that fell within the technology’s strengths.
But speed is not the same as wisdom. Output is not the same as understanding. And convenience is not the same as intelligence.
That distinction is becoming one of the central questions of the AI age. The emerging evidence suggests that AI is best understood not as a simple brain booster or brain drain, but as a cognitive amplifier. It magnifies habits we already have. Used actively, it can sharpen thinking. Used passively, it can replace thinking.
The difference matters.
For a skilled writer, AI can serve as a sparring partner: offering alternative structures, identifying weak arguments, suggesting counterpoints and forcing a clearer draft. For a novice writer, the same tool can become a shortcut around the struggle that teaches writing in the first place. For an experienced programmer, AI can automate routine code and free attention for architecture and debugging. For a beginner, it can produce answers that work just well enough to conceal what the user does not understand.
This is the paradox: AI can make us more capable while making some of our underlying abilities more fragile.
The pattern is not entirely new. Humans have always outsourced parts of cognition. Writing reduced the need to memorize oral traditions. Calculators changed arithmetic. GPS changed navigation. Search engines changed recall. A well-known 2011 study on the “Google effect” found that people were less likely to remember information itself when they believed they could retrieve it later, and more likely to remember where to find it.
That was not necessarily a catastrophe. External memory can be useful. Nobody should have to memorize every phone number, tax rule or historical date. The danger comes when offloading becomes substitution: when the tool does not merely store information but performs the judgment, synthesis and expression that used to require internal effort.
Generative AI goes further than search. Search gives links. AI gives answers. Search requires users to compare, inspect and assemble. AI often arrives as a finished paragraph, a confident explanation or a polished conclusion. The mental labor shifts from “figure this out” to “decide whether this sounds right.” That is easier. It is also riskier.
A 2025 Microsoft Research and Carnegie Mellon study surveyed 319 knowledge workers and found that confidence in generative AI was associated with less perceived critical-thinking effort, while greater confidence in one’s own ability was associated with more critical engagement with AI outputs. The finding does not prove that AI is making workers less intelligent. But it points to a plausible mechanism: the more capable the machine seems, the easier it becomes to stop checking it.
Education may be the clearest test case. Students have always looked for shortcuts, but AI changes the scale and quality of the shortcut. A chatbot can write the essay, solve the problem, generate the outline, explain the reading and prepare the discussion post. That can help a student who uses it as a tutor. It can hurt a student who uses it as a substitute.
A study published in Proceedings of the National Academy of Sciences found that generative AI without guardrails could harm learning. Students using GPT-4 during math practice performed better while using the tool but later performed worse when working on their own. A more constrained tutor version, designed to guide rather than simply answer, reduced those harms.
That distinction is crucial. The problem is not that AI helps. The problem is that it can help too much, too early, and too invisibly.
Learning often requires productive difficulty. A student has to sit with confusion, try a method, make an error, revise an assumption and slowly build a mental model. AI can shorten that process in useful ways, but it can also remove the very friction that makes learning durable. When the answer arrives instantly, fluently and painlessly, the student may feel understanding without having earned it.
This is not just a school problem. Many adults now use AI to draft emails, summarize articles, evaluate arguments and make decisions. That can be efficient. But the same question follows them from the classroom to the workplace: Are they using AI to think better, or to think less?
The answer often depends on whether the user remains intellectually responsible for the final result.
One of the most important findings from the BCG research was that AI’s benefits were uneven. On tasks well suited to the tool, consultants improved. But on tasks outside the technology’s competence, AI could degrade performance, in part because users trusted it when they should not have. Researchers described this uneven capability as a “jagged technology frontier”: AI is impressive in some areas, brittle in others and not always obvious about which is which.
That jaggedness is what makes AI different from simpler tools. A calculator reliably adds numbers. A GPS may misroute you, but usually within a constrained map. A generative AI system can produce a brilliant explanation in one moment and a fabricated citation in the next. Its fluency can mask uncertainty. Its confidence can make error feel authoritative.
The danger, then, is not merely laziness. It is miscalibrated trust.
Researchers have long studied automation bias, the tendency to over-rely on automated systems even when they are wrong or when warning signs are present. Recent work on human-AI collaboration shows that overreliance remains a serious challenge as AI tools enter fields like health care, law, government and business.
Still, the bleakest version of the argument goes too far. There is not yet strong evidence that AI is causing a broad decline in human intelligence. Much of the research is early, task-specific or focused on short-term performance. Studies can show that people learn less in a particular setting, rely too much in a particular experiment or report reduced effort in a particular workflow. That is meaningful evidence, but it is not proof that civilization is becoming stupid.
The more grounded concern is narrower and more practical: skills weaken when they are not practiced. If students stop writing, writing will suffer. If analysts stop checking, judgment will suffer. If readers stop reading deeply, comprehension will suffer. If professionals accept AI-generated work without inspection, expertise will become thinner even as output becomes more polished.
At the same time, refusing AI altogether is not a serious answer. The technology is too useful, too widespread and too integrated into modern work. The better question is not whether to use AI, but how to use it without surrendering the human capacities that make it valuable in the first place.
That means designing AI use around effort rather than escape.
A student should ask for hints before answers. A writer should ask for criticism before prose. A manager should ask for counterarguments before recommendations. A programmer should ask AI to explain code, not merely generate it. A reader should use summaries as previews, not replacements. A professional should treat every AI output as a draft, not a verdict.
The healthiest use of AI may be adversarial: not “Do this for me,” but “Challenge me.” Not “Give me the answer,” but “Show me what I am missing.” Not “Write my argument,” but “Find the weakest part of my argument.” Used that way, AI can increase the amount of thinking a person does, not reduce it.
The future of intelligence may therefore depend less on the technology itself than on the habits we build around it. AI will make some people sharper, faster and more creative. It will make others more dependent, more credulous and less practiced at sustained thought. The same tool can produce both outcomes.
So is AI making us smarter or mentally lazier?
The evidence suggests a more uncomfortable answer: AI is making it easier to choose either path.
Evidence & Source Transparency
Evidence First shows its work. The article ends above; this section is included so readers can inspect the main sources behind the factual claims.
The list below does not source every sentence. It focuses on the factual claims most important to the argument.
1. AI and coding productivity
Claim or topic:
Generative AI tools can improve short-term task performance, including programming speed.
Source:
The Impact of AI on Developer Productivity: Evidence from GitHub Copilot
Source type:
Academic research
What it supports:
This study found that developers using GitHub Copilot completed a coding task substantially faster than those who did not use the tool.
Important caveat:
The study measured performance on a specific coding task. Faster completion does not automatically prove deeper understanding or long-term skill improvement.
2. AI and professional knowledge work
Claim or topic:
Generative AI can improve performance on some professional tasks but can also hurt performance when used outside its strengths.
Source:
Boston Consulting Group: How People Create and Destroy Value with GenAI
Source type:
Analysis
What it supports:
The source discusses research involving consultants using AI and describes how AI improved performance on many tasks while creating risks when users relied on it for tasks it was poorly suited to handle.
Important caveat:
The findings are task-dependent. They do not show that AI improves all types of professional judgment or decision-making.
3. The “jagged frontier” of AI capability
Claim or topic:
AI systems are highly capable in some areas and unreliable in others, and users may not always know which is which.
Source:
Harvard Business School faculty page: Navigating the Jagged Technological Frontier
Source type:
Academic research
What it supports:
This research supports the article’s point that AI’s usefulness is uneven: it can help on tasks within its competence but degrade performance when users trust it on tasks outside that frontier.
Important caveat:
The “frontier” changes as models improve, so the exact boundaries of AI competence are not fixed.
4. Cognitive offloading and the “Google effect”
Claim or topic:
People may remember less information when they expect to be able to retrieve it later.
Source:
Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips
Source type:
Academic research
What it supports:
This study supports the article’s broader argument that humans often outsource memory to external tools and that access to information can change what people remember.
Important caveat:
The study concerns search and memory, not modern generative AI directly. It is relevant as background evidence, not proof of AI-specific effects.
5. AI, critical thinking, and overreliance
Claim or topic:
Higher confidence in AI tools may be associated with lower critical-thinking effort, while users with higher self-confidence may engage more critically.
Source:
Microsoft Research: The Impact of Generative AI on Critical Thinking
Source type:
Academic research
What it supports:
This source supports the article’s claim that AI can reduce perceived cognitive effort and that overreliance is a real concern in knowledge work.
Important caveat:
The study relies on self-reported behavior from knowledge workers. It does not prove a population-wide decline in intelligence.
6. AI in education and learning outcomes
Claim or topic:
AI can help students during practice but may harm learning when it gives answers too easily or removes productive struggle.
Source:
Proceedings of the National Academy of Sciences: Generative AI Can Harm Learning
Source type:
Academic research
What it supports:
This study supports the article’s distinction between AI used as a tutor and AI used as a substitute for learning. Students using unrestricted AI support performed better during practice but worse later when working without it.
Important caveat:
The finding applies to the study’s educational setting and design. It should not be read as evidence that all AI use in education is harmful.
7. Automation bias and human-AI decision-making
Claim or topic:
People can over-rely on automated systems, including AI, even when those systems are wrong.
Source:
AI & Society: Automation Bias and Human-AI Decision-Making
Source type:
Academic research
What it supports:
This source supports the article’s concern that AI’s fluency and apparent confidence can lead users to accept outputs without enough scrutiny.
Important caveat:
Automation bias is a broad human-factors issue. The exact level of risk varies by task, user expertise, interface design, and consequences of error.
How to read this evidence
This article is the author’s analysis. The sources above are provided so readers can see where the factual claims come from and judge the evidence for themselves. Some sources support direct facts, while others provide context, estimates, or background evidence.
Corrections and updates
If a factual error is identified, this post will be corrected in the web version with a dated note explaining the change. Because email versions cannot be edited after sending, the web version should be treated as the current version.