Are We Becoming Too Dependent on AI?

The core argument

Artificial intelligence has changed how teenagers learn, communicate, and get through their day. That change is not automatically good or bad, but this paper makes one specific claim: the real danger of teen AI dependence is cognitive offloading, letting AI do the mental work, and skipping the productive struggle that actually builds thinking, judgement, and resilience. The evidence below suggests this is already measurable. And because the pull toward AI is built into how convenient and how graded our lives are, the fix is not willpower or another awareness poster. It is institutional: schools have to change how they assess learning so that effort, not output, is what gets rewarded.

In short: AI is not making teenagers stupid. It is offering them a way to opt out of the difficulty that makes them smart, and right now, nothing in the system stops them from taking it.

Causes of AI dependence

1. Accessibility and convenience. AI tools are free, instant, and everywhere, a chatbot for homework, a virtual assistant for questions, a generator for essays, all one click away. When the easy path is always open, reliance is the natural result: it removes the effort and time a task would otherwise demand. Convenience is not a character flaw; it is a design that rewards taking the shortcut.

2. Academic pressure. Under pressure to keep grades up and finish work on time, many teens reach for AI as a shortcut. The scale of this shift is now documented. According to Pew Research Center, the share of U.S. teens who use ChatGPT for schoolwork doubled from 13% in 2023 to 26% in 2024 (Pew Research Center, 2025). By late 2025, a broader Pew survey found that 54% of teens had used an AI chatbot for help with schoolwork and 57% to search for information (Pew Research Center, 2025). Some of this use is legitimate learning support, but when it becomes a way to hand in work you did not do the thinking for, it is a form of cheating, and it is unfair to students who did the work themselves.

3. Substitution for human interaction. AI chatbots and "companion" apps offer instant, frictionless replies and never argue back, so some teens start to prefer them to real relationships, which are slower and harder. This is now common: a nationally representative Common Sense Media survey found that 72% of U.S. teens have used an AI companion at least once, and about half use one regularly (Common Sense Media, 2025). Importantly, most teens keep perspective, two-thirds said conversations with AI companions were less satisfying than with real people, so the risk is not that AI replaces friends outright, but that it becomes the default when human connection takes effort, weakening the social muscles teens are supposed to be building.

The impact of AI dependence

1. Weaker critical thinking and creativity. This is the heart of the concern. When you let AI do the reasoning, you skip the mental effort, "cognitive offloading", that builds the skill in the first place. A 2025 study of 666 people published in the journal Societies found a significant negative correlation between frequent AI-tool use and critical-thinking ability, and it was cognitive offloading that explained the link (Gerlich, 2025). Younger participants were both the most dependent on AI and the lowest-scoring on critical thinking, though higher education seemed to buffer the effect, which suggests the habit can be trained the other way. A 2025 MIT Media Lab experiment adds a striking mechanism: using EEG to watch students' brains as they wrote essays, researchers found the group using ChatGPT showed the weakest brain connectivity, and over several months these users "got lazier," increasingly resorting to copy-and-paste (Kosmyna et al., 2025). (This second study is a small preprint of 54 students and not yet peer-reviewed, so it should be read as suggestive, not settled.) Together they point the same way: outsource the thinking often enough and the capacity for it appears to fade.

2. Privacy and ethical concerns. As AI systems collect and analyse huge amounts of personal data, people often have no idea how their information is being used, and that information can be misused through identity theft, hacking, or blackmail. The public knows it feels powerless here: Pew Research Center found that 73% of Americans feel they have very little or no control over the data companies collect about them, 81% are concerned about how that data is used, and 70% of those who have heard of AI do not trust companies to make responsible decisions about it (Pew Research Center, 2023). The more teenagers pour personal questions, drafts, and worries into AI tools, the more of themselves they hand over to systems they neither see nor control.

3. Mental-health effects. Over-reliance on AI can affect well-being. AI companions can offer a kind of emotional support, but they can also substitute for human connection and empathy, deepening isolation for the teens who lean on them most. Common Sense Media, which ran the companion survey above, concluded the risks are serious enough that it recommends no one under 18 use AI companions until stronger safeguards are in place (Common Sense Media, 2025). The evidence on whether this causes anxiety or depression is still emerging and should not be overstated, but the direction of concern is clear: when someone feels unable to cope, decide, or connect without a tool, that is dependency, and dependency erodes resilience.

Recommendations

Awareness campaigns and "use AI responsibly" reminders do not work, because the incentive to offload is stronger than the reminder not to. The pull has to be met with structure. The most effective lever is the one that already shapes teen behaviour: assessment. Three concrete, adoptable mechanisms:

1. A mandatory AI-use disclosure on every assignment. Instead of a blanket ban (unenforceable) or silence (which rewards hiding it), require every submission to carry a one-line declaration of how AI was used, using a simple published tool: the AI Assessment Scale (Perkins, Furze et al., 2024), a five-level scale running from "No AI" through "AI for planning/editing" to "full AI collaboration." The teacher sets the permitted level for each task; the student states what they actually did. This makes honesty the norm, turns AI use into something teachers can teach and grade, and removes the grey area that makes cheating easy.

2. Randomized oral defenses of written work. For a share of essays, students are picked at random to spend a few minutes explaining and defending their own submission aloud, why they chose an argument, where a source came from, what a sentence means. A portion of the grade rides on the defense. Because you cannot defend reasoning you never did, this directly authenticates authorship and rewards genuine understanding over polished output. It is the single most practical answer to the copy-paste problem the MIT study describes.

3. Periodic "AI-off" assessment blocks. Keep some assessments supervised and unassisted, handwritten or offline, in class, so that unaided skill is still practised and measured. This is the "brain-only" condition that showed the strongest cognition in the MIT study, rebuilt into the school calendar. It guarantees that students keep exercising the thinking AI would otherwise do for them.

Parents still matter, setting limits and modelling AI as a tool that supports effort rather than replaces it, and genuine digital literacy (knowing that AI can be confidently wrong, and checking its claims against a real source) belongs in the curriculum. But the decisive move is institutional. If schools change what they reward, teenagers will change what they do. Dependence is not inevitable; it is what happens when the system quietly rewards it.

Sources

1. Share of U.S. teens using ChatGPT for schoolwork doubled from 13% (2023) to 26% (2024).
2. By late 2025, 54% of U.S. teens had used an AI chatbot for help with schoolwork and 57% to search for information (n=1,458).
3. Gerlich (2025), Societies 15(1):6, n=666: significant negative correlation between frequent AI-tool use and critical thinking, mediated by cognitive offloading; younger users most dependent and lowest-scoring; education buffers the effect.
4. Kosmyna et al. (2025), MIT Media Lab, "Your Brain on ChatGPT" (n=54, EEG): the ChatGPT group showed the weakest brain connectivity and, over months, "got lazier," increasingly copy-pasting. (Preprint; not yet peer-reviewed; small sample, noted in text.)
5. Pew (2023, n=5,101): 73% feel very little/no control over data companies collect; 81% concerned about how companies use it; 70% of those aware of AI distrust companies to make responsible AI decisions.
6. Common Sense Media / NORC (2025, n=1,060 teens): 72% have used an AI companion; two-thirds find AI conversations less satisfying than human ones; recommends no under-18 use until stronger safeguards exist.
7. Perkins, Furze et al. (2024), "The AI Assessment Scale Revisited": a published five-level framework (No AI → full AI) for declaring permitted AI use on assignments; also aiassessmentscale.com.
8. Ahmad et al. (2023), the paper's original "Dr Ahmed" source, marked RETRACTED by the editors (used only to justify removing it).