Does Using AI for Essays Make You a Worse Writer Over Time?

Sami Ullah Khan

October 6, 2026

Does Using AI for Essays Make You a Worse Writer Over Time

Yes, using AI for essays can make you a worse writer over time if you repeatedly let it do the parts of writing that train judgement, memory, argument and revision — but current evidence does not show that all AI-assisted writing causes lasting skill loss. The sharper question is not whether AI touches the essay. It is which cognitive work you hand over. A student who asks a model to produce a thesis, organise the argument, choose evidence and draft the prose is practising far less writing than a student who writes first and uses AI to challenge weak logic or explain why a paragraph is unclear.

That distinction matters because AI-assisted writing is now mainstream. HEPI’s 2026 survey of 1,054 UK undergraduates found that 95% had used AI in at least one way and 94% had used generative AI to help with assessed work; 12% said they directly included AI-generated text in assessed work. The research landscape, however, is more nuanced than the headlines. A widely discussed MIT Media Lab study found lower brain connectivity and weaker recall in an LLM-assisted essay condition, yet the researchers also stress that the experiment was specific to educational essay writing and does not establish a universal or permanent decline in intelligence or writing ability.

This article separates immediate performance from learning, surface-level polish from transferable skill, and productive assistance from cognitive substitution. It also examines what 2025–2026 research says about critical thinking, ownership, memory, EFL writing, academic literacy and student behaviour, then turns that evidence into a practical way to use AI without surrendering the craft you are trying to build.

The Core Distinction: Better Essay, Better Writer, or Both?

For a broader practical framework, the magazine’s AI academic writing guide uses the same principle: AI can accelerate a stage without becoming the unexamined bridge between a claim and its evidence. The measurable question is therefore not “Was AI used?” but “Which decisions still belonged to the writer?”

A polished essay is an output. Writing ability is a capacity. Those are not the same thing, and much of the public debate confuses them. Generative AI can improve an output immediately by supplying cleaner syntax, a more orderly structure or a stronger transition. That does not mean the user learned how to produce those improvements independently.

Writing skill grows through repeated decisions: deciding what you believe, selecting evidence, arranging ideas, anticipating objections, choosing precise language and revising after discovering that your first version was weak. When AI performs those decisions invisibly, the final document can improve while the writer gets fewer repetitions of the underlying skill. That is the central risk of cognitive offloading in essay writing.

The opposite can also happen. An AI system can create more useful practice by exposing weaknesses the writer would not have noticed alone. A student can ask for three objections to a thesis, compare those objections with the evidence, reject two, then rewrite the argument. In that workflow the model increases friction in the right place instead of removing it.

What ImprovesWhat It MeasuresCan AI Boost It Quickly?Does That Prove Skill Growth?
Grammar and sentence fluencySurface qualityOftenNo
OrganisationDocument structureOftenNo
Argument qualityReasoning in the submitted textSometimesNot by itself
Recall of the argumentOwnership and encodingNot reliablyMore relevant
Unaided rewrite or transfer taskPortable writing skillOnly indirectlyYes, more strongly

What the MIT Essay Study Actually Found — and What It Did Not

The study that drives many “AI is making us stupid” headlines is Kosmyna and colleagues’ 2025 preprint, Your Brain on ChatGPT. Fifty-four participants completed the first three essay-writing sessions in one of three conditions: LLM, search engine or brain-only. Eighteen completed a fourth crossover session. EEG was used to measure functional connectivity while the researchers also analysed essay language, recall and self-reported ownership.

The LLM group showed the weakest connectivity patterns during the task, while the brain-only group showed the strongest. LLM users also had more difficulty accurately quoting their own essays and reported lower ownership. In the revised preprint abstract, the authors describe the LLM group as underperforming at neural, linguistic and behavioural levels across the four-month study window.

Those findings are important, but they are not evidence that ChatGPT causes permanent writing decline. The MIT project page explicitly says the findings are context-dependent, focused on essay writing in an educational setting and may not generalise across tasks. It also calls for future longitudinal work on memory retention, creativity and writing fluency. EEG connectivity is not a simple “intelligence meter”, and lower activation during a tool-assisted task can sometimes reflect efficiency rather than damage; interpreting the pattern requires the behavioural evidence around it.

In a May 2026 MIT Media Lab post, lead researcher Nataliya Kosmyna wrote that AI tools have “undoubtedly, an impact on our cognition”. That is a defensible claim. The stronger claim — that using AI for essays inevitably makes people worse writers — goes beyond the study.

The practical takeaway is narrower and more useful: if the model is producing the essay while the student mostly accepts and lightly edits it, the student is likely doing less of the cognitive work normally associated with composing and remembering an argument.

The Critical-Thinking Evidence Points to Trust, Not Just Tool Use

Writing is not only language production; it is a structured form of thinking. Microsoft Research and Carnegie Mellon researchers surveyed 319 knowledge workers who supplied 936 real examples of generative-AI use. Their CHI 2025 study found that higher confidence in GenAI was associated with less reported critical-thinking effort, while higher confidence in one’s own ability was associated with more.

This shifts the debate from a simple exposure model to a reliance model. Two people can use the same chatbot for the same assignment and experience different cognitive effects. The person who treats AI as an authority is more likely to accept framing, facts and reasoning without interrogation. The person who treats AI as a fallible collaborator may spend more effort verifying, integrating and supervising the result.

That pattern helps explain why a well-designed AI workflow can still be educational. Critical thinking does not disappear; it moves. Instead of generating every sentence from scratch, the writer may spend more time checking sources, comparing alternatives, spotting hidden assumptions and deciding what to reject. But that benefit only appears if the user is capable of and motivated to perform those checks.

Students who want to strengthen that habit can pair AI use with critical thinking exercises that force claim testing, counterargument, source evaluation and reflection before accepting a generated answer.

The risk, then, is not automation in the abstract. It is unearned confidence. Once the writer assumes that fluent output is probably correct, the activity changes from composing and judging to selecting and submitting.

Short-Term Fluency Gains Can Hide Deeper Learning Losses

A 2026 quasi-experimental study by Hien Thi Mai followed 42 Vietnamese high-school EFL learners for six weeks. Students using ChatGPT in a supervised lab initially outperformed the teacher-instruction control group on fluency and structure. Yet the study also reported more formulaic writing and reduced critical engagement where students over-relied on the tool. The control group improved more gradually in originality, argumentation and cognitive ownership.

That result captures the central paradox. AI can make a developing writer look more advanced before the underlying capability has caught up. For a learner who struggles with English, that assistance may be genuinely valuable: seeing a clearer structure or receiving immediate feedback can reduce language friction and make revision possible. But if the student repeatedly submits the model’s solution, the improved paper can mask fewer opportunities to learn.

Another 2026 study of Indonesian EFL university students found no significant post-test difference in essay quality between the ChatGPT-assisted and conventional groups, despite positive student perceptions around usefulness, feedback and confidence. That is a useful counterweight to alarmist claims. AI did not automatically degrade writing performance, but neither did access alone create a clear advantage over conventional instruction.

For students, the practical decision is task-specific. When the learning objective is argument development, the thesis and evidence chain should remain human-led. When the learning objective is editing, using AI to identify redundancy or grammatical problems can be appropriate. The magazine’s

The distinction becomes clearer in the ChatGPT student study guide, which separates legitimate learning support from handing over assessed thinking.

AI RoleLikely Effect on Immediate EssayPractice Preserved?Long-Term Risk
GhostwriterLarge improvement possibleLowHigh if habitual
Outliner after no independent attemptFaster structurePartialModerate
Tutor asking questionsSlower but often deeperHighLow
Counterargument generatorCan strengthen reasoningHigh if writer evaluatesLow
Copy editor after human draftCleaner proseMostlyLow
Source verifierCan improve accuracyHigh if sources are checkedLow

Why Repeated Outsourcing Can Erode a Writer’s Voice

Voice is not a decorative style layer added after the argument. It emerges from repeated choices about what to notice, what to omit, how strongly to state a claim, where to slow down, which examples feel convincing and what kinds of uncertainty deserve acknowledgement. A model can imitate a style, but if it chooses most of those decisions, the user gets fewer chances to discover their own habits.

This is one reason AI-heavy essays can converge toward generic structures: balanced three-part thesis, tidy signposting, symmetrical paragraphs and frictionless transitions. Those features are not inherently bad. The problem is that the student may learn to recognise polished structure without learning to build or deliberately break it.

The 2025 MIT study also found within-group linguistic homogeneity, supporting concern that AI assistance can compress variation. That does not prove that every AI-assisted writer loses originality, but it makes a practical warning visible: if your draft starts sounding interchangeable with the model’s default prose, you may be practising acceptance rather than authorship.

Jeanne Beatrix Law, a Kennesaw State University professor who teaches AI-integrated writing, described the opposite design goal in a 2026 interview: “The goal is always to cultivate engagement with the writing process, not just the product.” Her classroom GPTs are intentionally configured not to write the paper for students.

That process-first approach is also why a tool-specific workflow such as the magazine’s Gemini essay writing guide keeps thesis, evidence selection and final wording under human control instead of treating generation as the default.

Academic Literacy Is More Than Producing Correct Sentences

Long-term writing ability includes behaviours that rarely appear in a finished essay: finding a source, deciding whether it deserves trust, distinguishing a quotation from an interpretation, noticing when evidence does not support the claim, tracking uncertainty and revising after feedback. These are forms of academic literacy, and AI can either scaffold them or bypass them.

A 2026 longitudinal qualitative study in System followed two advanced ESL students across an academic year using monthly interviews, surveys, ChatGPT logs and supporting artefacts. Its small sample prevents broad causal claims, but it is valuable because it shows sustained AI use does not produce one uniform trajectory. Students can retain agency when they use the system strategically inside a larger literacy practice rather than allowing it to become the author.

A 2026 systematic review of 64 empirical studies in computer-science education reached a similarly mixed conclusion. Generative AI could support adaptive learning and problem solving, but hallucinated or misleading outputs increased cognitive load and over-reliance in some settings. The review emphasised the role of metacognition: learners who monitor, verify and reflect on AI output are in a different position from learners who consume it passively.

This is why “AI makes writing easier” is not sufficient evidence of educational value. The right question is whether the easier step is one the student has already mastered or one the assignment is specifically designed to teach.

For students dealing with disclosure rules as well as learning goals, the magazine’s guide on telling your teacher about AI offers a contribution-based approach rather than treating every AI interaction as equivalent.

The Five Writing Muscles Most at Risk

If habitual AI use does weaken writing, it is unlikely to do so as one single ability. The more realistic model is selective disuse: skills that are repeatedly outsourced receive fewer repetitions. Five are especially vulnerable.

1. Thesis Formation

Turning a broad topic into a defensible claim requires narrowing, prioritising and accepting trade-offs. If AI supplies the thesis every time, the writer may become better at choosing from options without becoming better at producing a position from uncertainty.

2. Evidence Selection

Strong writers do not merely collect facts; they judge which evidence actually changes the argument. Models can suggest citations or examples, but the learning occurs when the student reads, verifies and decides what the evidence can legitimately support.

3. Structural Reasoning

An outline is a theory of how one idea leads to the next. Automatic outlining can remove the productive struggle of deciding whether a counterargument belongs before or after a piece of evidence, or whether two apparently separate points are actually the same point.

4. Sentence-Level Retrieval

Writing fluency partly depends on retrieving vocabulary, syntax and rhetorical patterns without external help. Constantly asking AI to phrase ordinary ideas can reduce those retrieval attempts, particularly for learners still acquiring the language.

5. Revision Judgement

Revision is where writers learn to diagnose their own work. Accepting one-click rewrites may improve the paragraph while weakening the habit of asking what is wrong, why it is wrong and which change best preserves the intended meaning.

A Safer Human-First Workflow for AI-Assisted Essays

The safest way to use AI is not to ban it from the workflow. It is to place it after the moments where learning depends on independent retrieval and judgement. A simple sequence keeps the difficult intellectual decisions visible.

StageHuman Must Do FirstUseful AI RoleProof You Still Own the Work
QuestionRestate the assignment and constraintsClarify ambiguous wordingYou can explain the task without the chat
ThesisWrite a provisional claimGenerate objections or alternative framingsYou can defend why your thesis is preferable
EvidenceRead and select sourcesSummarise supplied passages; flag gapsYou have opened and checked every cited source
OutlineArrange the argumentStress-test order and missing stepsYou can recreate the outline from memory
DraftWrite core paragraphsAsk diagnostic questionsMost substantive sentences begin with your own language
RevisionIdentify what feels weakOffer options and explain trade-offsYou choose and can justify each major edit
Final CheckRead aloud and verify claimsProofread mechanicsYou can explain every paragraph without AI

A strong prompt in this workflow does not ask the model to solve the essay. It asks the model to create useful resistance: “What assumption am I making here?”, “Which sentence is least supported?”, “Give me the strongest objection to this paragraph”, or “Do not rewrite this; explain why the logic is hard to follow.”

Tool-specific guides can still be useful when the process is constrained. The magazine’s Grok essay workflow evaluates prompts partly by whether they preserve human decision-making and create an auditable source trail.

How to Tell Whether AI Is Helping or Replacing Your Writing

Students often focus on how much AI text appears in the final draft, but dependence is easier to detect through behaviour. You can use very little generated text and still outsource most of the reasoning; you can also use substantial AI feedback while keeping authorship firmly human.

Three tests are more revealing than percentage estimates. First, the reconstruction test: close the chatbot and reproduce the thesis, major evidence and paragraph order from memory. Second, the defence test: explain why each source belongs and what would falsify your claim. Third, the transfer test: write a short new paragraph on a related question without AI. If performance collapses, the original workflow may have produced output without building transferable skill.

Warning signs include opening the chatbot before thinking about the assignment, asking for a full outline before reading sources, accepting phrases you would not normally use, being unable to explain a citation, repeatedly requesting rewrites instead of diagnosing a paragraph, and feeling unable to start without AI.

None of these signs proves harm. They indicate that the tool is occupying a role that would otherwise create practice. The remedy is not necessarily abstinence; it is to move AI later in the process or change the prompt from generation to critique.

This is also where academic-integrity boundaries matter. The magazine’s explainer on AI homework and cheating makes the useful distinction between support and substitution: the ethical boundary often overlaps with the learning boundary.

When AI Can Make You a Better Writer

AI can improve writing ability when it expands feedback, creates deliberate practice and exposes the writer to decisions they must still make. This is especially useful where human feedback is scarce or delayed.

A model can act as a Socratic tutor by asking the writer to define a vague term, justify a causal claim or distinguish correlation from causation. It can compare two versions of a paragraph and explain the trade-off without choosing. It can role-play a sceptical reader. It can identify where the writer assumes background knowledge. It can generate practice sentences that target a recurring grammar problem, then withhold the answer until the student attempts a correction.

Law’s 2026 classroom work offers a concrete version of this idea. Her custom GPTs are grounded in curated course materials, prompt students to revise and are configured not to write papers for them. She argues that AI can “empower students, not replace them” when process, reflection and agency remain central.

HEPI’s 2026 findings also show why the same tool can produce opposite experiences. One survey respondent said AI freed time for deeper analysis; another said, “I’m not using my brain at all.” That contrast is not a contradiction. It is evidence that usage pattern matters.

For writers choosing among platforms, the magazine’s overview of AI writing tools in 2026 is most useful when read as a workflow comparison, not as a reason to outsource authorship.

What We Still Do Not Know About Long-Term Skill Erosion

The phrase “over time” is the hardest part of the keyword because the technology is young and genuinely longitudinal evidence is still limited. The MIT study followed a specific essay task across several sessions and a four-month period, but it was not a multi-year test of writing development. The 2026 System study followed two ESL students for an academic year, providing rich qualitative evidence but not a population-level effect. Six-week classroom experiments can reveal short-term learning patterns, not lifetime outcomes.

Researchers also need better separation between different kinds of AI use. A study that labels everyone “ChatGPT users” can hide crucial differences between full-draft generation, grammar feedback, brainstorming, retrieval practice and Socratic tutoring. Frequency also matters. One AI-assisted essay a month is not the same intervention as having a model generate every paragraph for two years.

Baseline skill is another confounder. Novice writers may gain more from explicit examples and immediate feedback, yet they may also be more vulnerable to accepting inaccurate or formulaic output. Expert writers can offload routine editing while preserving judgement because they already possess the underlying skill. The same action can therefore have different developmental consequences.

The strongest 2026 research direction is not “AI versus no AI” but task decomposition: which cognitive operations are delegated, when, by whom, with what feedback, and what happens when the tool is removed. Until those studies mature, claims of inevitable brain damage or guaranteed skill enhancement should both be treated sceptically.

A Practical Rule: Protect the Struggle That Teaches

Not all difficulty is useful. Spending twenty minutes fixing a citation format may add less learning than spending twenty minutes testing whether evidence actually supports a claim. AI is valuable when it removes low-value friction and dangerous when it removes high-value struggle.

For most essay assignments, the high-value struggle includes interpreting the question, forming a thesis, reading evidence, connecting claims, confronting counterarguments and deciding what the final wording should mean. Lower-value friction may include mechanical formatting, repetitive proofreading or generating practice questions after the student has already learned the material.

Charlotte Armstrong, co-author of HEPI’s 2026 survey, argued that “AI literacy and capability must be embedded across the curriculum.” That literacy should include knowing when not to delegate. A competent AI user is not someone who can get a complete essay fastest; it is someone who can choose which parts of the task should remain deliberately human.

The simplest rule is this: before using AI, identify the skill the assignment is supposed to train. Do that part yourself first. Then use AI around it — to test, challenge, explain, compare or proofread — rather than through it.

Our Editorial Verification Process

This article was built from a live October 2026 search review of the exact keyword and close variants. We examined ten prominent or directly relevant results and research pages spanning consumer explainers, MIT Media Lab reporting, Microsoft Research, peer-reviewed education studies and current university-sector evidence. Recurring SERP structures tended to use one of three frames: “AI makes you dumber”, “AI is just a tool”, or “use AI responsibly”. The main gap was failure to separate immediate essay quality, in-task cognitive engagement and longitudinal writing skill.

Primary evidence was cross-checked against the MIT Media Lab project and 2025 preprint, Microsoft Research’s CHI 2025 study, HEPI’s Student Generative AI Survey 2026, a 2026 Discover Education quasi-experiment, a 2026 Indonesian EFL essay study, a 2026 System longitudinal study and a 2026 systematic review in Computers and Education: Artificial Intelligence. We report sample sizes and study durations where available and distinguish peer-reviewed findings from preprints and qualitative evidence.

The site sitemap.xml endpoint was attempted but was not retrievable through the available browsing layer. Rather than inventing sitemap entries, eight live indexed Perplexity AI Magazine articles were selected by semantic relevance to student AI use, academic writing, critical thinking, essay workflows and AI writing tools. Each internal URL appears once in a body section.

No commercial pricing table is included because this question is about learning and skill development, not a product purchase decision. Current plan caps or API integrations would not change the answer and would add noise rather than information gain.

This article was researched and drafted with AI assistance and reviewed by the Sami Ullah Khan editorial desk at Perplexity AI Magazine. All data, citations, pricing figures, and named quotes have been independently verified against primary sources before publication.

Conclusion

Using AI for essays can make you a worse writer over time, but only as a conditional claim: the risk is greatest when AI repeatedly replaces the cognitive operations that writing practice is supposed to strengthen. Current evidence supports concern about lower engagement, weaker ownership, reduced recall and formulaic output in some AI-heavy workflows. It does not establish that any exposure to AI permanently degrades writing ability.

The better distinction is between substitution and scaffolding. Ghostwriting substitutes. A full outline generated before the student has thought about the problem often substitutes. A model that challenges a thesis, points to an unsupported sentence, quizzes the writer on sources or explains why a paragraph is confusing can scaffold.

The unresolved question is longitudinal: how different patterns of AI use affect independent writing after months or years, especially across skill levels and languages. Until that evidence matures, the safest approach is conservative but not technophobic. Protect the parts of writing that build judgement. Use AI where it increases feedback and reflection. And periodically remove the tool to check whether the skill still exists without it.

Frequently Asked Questions

Does using AI for essays make you a worse writer over time?

Potentially, yes, if AI repeatedly does the thesis-building, evidence selection, organisation and drafting that you need to practise yourself. Current research shows risks around reduced cognitive engagement, ownership and critical thinking in some workflows, but it does not prove that all AI-assisted writing causes permanent decline.

Does ChatGPT make your brain less active when writing?

A 2025 MIT Media Lab preprint found lower EEG functional connectivity in its LLM-assisted essay group than in search-engine and brain-only groups. The researchers caution that the result is context-dependent and does not establish universal cognitive harm.

Can AI improve writing skills?

Yes, especially when used as a tutor, critic or feedback partner. AI can help writers spot weak logic, practise grammar, generate counterarguments and receive rapid feedback. The benefit is strongest when the writer still makes the important decisions and revises actively.

Is using AI to edit an essay bad for learning?

Not necessarily. Editing support is lower risk when you have already written the argument and can judge each change. The learning risk rises when you accept rewrites automatically or depend on AI to express ideas you cannot explain in your own words.

How much AI use is too much for an essay?

There is no reliable percentage threshold. A better test is whether you can reproduce the argument, defend the evidence and write a related paragraph without AI. If you cannot, the workflow probably outsourced too much of the learning.

Does AI make student writing more generic?

It can. Research has reported formulaic writing, lower ownership and linguistic homogeneity in some AI-assisted conditions. But careful prompting and human revision can preserve distinct voice, especially when the model is used for critique rather than first-draft generation.

What is the safest way to use AI for essays?

Write the question interpretation, provisional thesis, source notes and core argument yourself. Then use AI to test assumptions, generate objections, flag unclear reasoning and proofread. Verify every factual claim and citation before submission.

Should teachers ban AI for writing assignments?

Blanket bans may protect some forms of practice but can also remove useful feedback tools and fail to teach AI literacy. A stronger approach is assignment-specific rules that identify which cognitive work must remain human and which forms of AI assistance are permitted.

References

  1. Adejumo, A., Oyelere, S. S., Sanusi, I. T., & Suhonen, J. (2026). A systematic review of the impact of GenAI on learning performance, AI hallucinations, and problem-solving in computer science education. Computers and Education: Artificial Intelligence, 10, 100570.
  2. HEPI. (2026, March 12). Student Generative Artificial Intelligence Survey 2026 (Report 199).
  3. Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your Brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task (v2). arXiv.
  4. Kosmyna, N. (2026, May 6). AI and the Brain. MIT Media Lab.
  5. Law, J. B. (2026, May 21). Rebooting the writing classroom: Custom GPTs, student agency, and ethical AI. OpenAI Education.
  6. Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. Proceedings of CHI 2025.
  7. Mai, H. T. (2026). Evaluating the impact of using ChatGPT on EFL students’ argumentative skills. Discover Education, 5, 189.
  8. Puteri, C. G., Cahyono, B. Y., Widiati, U., & Suryati, N. (2026). Using ChatGPT as a writing assistant: A study on essay quality development among Indonesian EFL university students. Indonesian Journal of Applied Linguistics, 15(2), 332–346.
  9. OpenAI. (2026, May 21). Rebooting the Writing Classroom: Jeanne Beatrix Law on Custom GPTs, Student Agency, and Ethical AI.

Stay Ahead of AI

Get the latest AI news delivered to your inbox.

We don’t spam! Read our privacy policy for more info.

Published

Spotted an error? Read our corrections policy or email editorial@perplexityaimagazine.com.