Blog

Does ChatGPT Make You a Worse Writer? What the Research Actually Says

October 27, 2026 14 min read

AI can improve a finished draft while weakening the writer’s involvement in producing it. It can also teach by example and improve later writing without AI. The growing body of research is not contradictory once we ask what the tool did, when it entered the process, and what outcome was measured.

The fear that ChatGPT is making people worse writers returned with force in 2026. One widely discussed study found that brief access to an AI assistant was followed by weaker independent performance and less persistence when the assistant disappeared. That followed an MIT Media Lab experiment in which people using an LLM for essays showed lower neural engagement, weaker recall, and less ownership of their work than people who wrote without tools.

Those findings sound alarming. They are also only part of the evidence. A separate series of large, preregistered experiments found that people who practiced professional writing with an AI tool later wrote better cover letters without AI than people who practiced alone, used Google, or received personalized feedback from experienced editors.

So does ChatGPT make you a worse writer? The honest answer is conditional.

The practical answer: AI is most likely to weaken writing skill when it replaces idea formation, argument, recall, and revision. It can improve skill when it provides high-quality examples, targeted feedback, or useful opposition after the writer has begun thinking. The same tool can be a shortcut, a coach, or both.

First, separate better writing from a better writer

Most debates about AI and writing mix together four different outcomes:

Speed: Did the person finish the task faster?
Output quality: Was the final text clearer, more persuasive, or better structured?
Skill development: Could the person perform a similar task later without assistance?
Ownership and understanding: Could the person explain, defend, remember, and revise the text as their own?

An AI assistant can improve the first two while doing little for the last two. A person may submit a stronger email in half the time yet learn almost nothing about why the email works. Conversely, a person may use an AI-generated example, notice a better structure, and transfer that lesson to future writing.

This distinction explains why “the AI made the document better” does not prove that the AI made the writer better. It also explains why reduced effort is not automatically harmful. Less effort can mean productive support, or it can mean that the important cognitive work never happened.

What the MIT essay study found

The study most responsible for the “ChatGPT harms your brain” headlines was Your Brain on ChatGPT, published as an arXiv preprint by researchers associated with the MIT Media Lab.

The researchers placed 54 participants into three groups. One group wrote essays with an LLM, one used a search engine, and one wrote without external tools. The participants completed three sessions in the same condition. Eighteen returned for a fourth session in which some LLM users wrote without the model and some unaided writers gained access to it.

Electroencephalography, or EEG, was used to measure patterns of brain activity during the task. The unaided group showed the strongest and most distributed connectivity, the search group showed a middle level, and the LLM group showed the weakest connectivity. Participants in the LLM group also reported less ownership of their essays and had more difficulty accurately quoting what they had written.

The crossover result was particularly interesting. People who had first written without tools and then gained access to an LLM showed stronger recall and engagement than people who began with the LLM and later lost it. That suggests sequence may matter. Starting with your own thinking and adding AI later may produce a different experience from asking AI to establish the argument from the beginning.

What the study did not prove

The study does not show that ChatGPT causes brain damage, permanent cognitive decline, or a general loss of intelligence. It was a preprint with a small sample, a narrow essay task, and only 18 people in the crossover session. EEG connectivity is a measure of activity during a task. It is not a direct score of wisdom, learning, creativity, or brain health.

The result is best treated as an early warning about cognitive engagement, memory, and ownership in one type of writing workflow. It is not a final verdict on every writer, every model, or every use of AI.

A larger 2026 experiment found a persistence cost

A newer study, AI Assistance Reduces Persistence and Hurts Independent Performance, tested 1,222 participants across three randomized experiments involving fraction problems and reading comprehension. It was not a writing study, but it examined a mechanism that matters to writers: what happens when a tool removes the struggle involved in building a skill.

Participants with AI access performed better while the assistant was available. When it was removed, however, they performed worse and were more likely to stop engaging than people who had completed the tasks without AI. The effect appeared after sessions lasting roughly 10 to 15 minutes.

The most revealing detail was how people used the system. In one experiment, 61% of AI users said they mainly asked for direct answers. That group later had the lowest performance and the highest skip rate. Participants who used the AI for hints did not show the same pattern to the same degree.

The researchers are careful about the limits. The tests measured immediate performance after brief exposure. They did not establish whether the effect lasts for hours, days, or years. They also used a maximally helpful assistant that supplied complete answers on request. A system designed to ask questions, give partial hints, or delay the solution might produce different results.

For writing, the implication is straightforward. A model that immediately writes the thesis, selects the evidence, builds the structure, and supplies the conclusion may improve the document while removing the exact practice through which a writer learns to reason and persist.

The evidence also shows that AI can improve writing skill

The strongest counterweight comes from Coach not crutch, a series of preregistered experiments on professional writing. Participants learned principles for effective cover letters and then practiced with an AI writing tool or under comparison conditions.

In one study with 2,238 participants, people who practiced with AI wrote better cover letters in a later no-AI test than people who practiced without the tool or did no practice. Their letters were also more likely to receive a hypothetical interview offer. The AI group spent less time, logged fewer keystrokes during practice, and reported less effort, yet still performed better later.

A follow-up comparison found that AI-supported practice produced better later writing than searching Google for examples and tips, and better writing scores than personalized feedback from experienced human editors. Another experiment showed why: people who merely saw a strong AI-generated example improved about as much as people who actively practiced with the tool.

That finding matters because it rejects a simple equation between effort and learning. Writers often learn by observing good examples. If an AI system can produce a relevant example at the moment it is needed, it can make an abstract principle concrete. The benefit may come from exposure to quality, not from outsourcing the task.

The authors also warn against overgeneralizing. Their task was a short, structured form of professional writing. The participants interacted with the AI for a limited period. The results do not prove that repeated dependence improves every kind of writing, or that generating an entire novel, essay, or report teaches the same lessons.

AI already makes many people faster and more effective

The performance benefits are not limited to cover letters. In a widely cited experiment published in Science, Shakked Noy and Whitney Zhang assigned professional writing tasks to 453 college-educated participants. Access to ChatGPT reduced completion time by about 40% and increased assessed output quality by about 18%.

These gains are real, and they explain why AI writing tools have spread so quickly. For routine emails, summaries, reports, and first drafts, a model can remove blank-page friction, improve structure, and help less experienced writers reach a competent baseline.

But productivity and skill are not interchangeable. A calculator can improve an answer without improving mental arithmetic. GPS can improve a journey without improving a person’s internal map. The relevant question is not whether offloading is bad. It is which parts of writing you want to retain as human capabilities.

Why the studies are not actually in conflict

The research points toward a more useful model than “AI good” or “AI bad.” Outcomes change according to four design choices.

1. Timing

AI used before a writer has formed any view can anchor the entire piece. The model supplies the first framing, the obvious examples, and the shape of the conclusion. AI used after the writer has made notes or written a rough draft has less power to replace the underlying thought.

2. Type of help

A direct answer ends the search. A hint, critique, counterargument, example, or question can extend it. The 2026 persistence study found the clearest harm among users who requested direct solutions, while the cover-letter experiments found benefits from seeing strong examples.

3. Goal

A writer optimizing for immediate output will use AI differently from a writer optimizing for learning. Both goals can be legitimate. A busy professional may reasonably want a faster routine email. A student practising argument should not treat completion speed as the only measure of success.

4. Human investment

Ownership grows when a writer chooses the claim, selects evidence, decides what matters, and makes meaningful revisions. It shrinks when the person approves fluent sentences they did not have to understand. Longer prompts alone do not guarantee investment. The important issue is whether the writer made substantive decisions.

A coach workflow for AI-assisted writing

The safest way to use AI without surrendering the writing process is to give the model a supporting role and preserve the decisions that build skill.

Think before you prompt. Write a rough position, a few claims, the intended reader, and what you want the reader to do or understand.
Ask for questions before answers. Have the model identify gaps, ambiguities, counterarguments, or missing evidence before it generates replacement prose.
Use examples as study material. Ask for two or three alternative openings or structures, then explain what each one does well before writing your own.
Draft the hard part yourself. Write the thesis, analysis, or central story before asking AI to improve clarity. These are the parts most closely tied to thought and voice.
Request critique with criteria. Tell the model to evaluate logic, evidence, tone, repetition, and reader assumptions. A targeted review is more educational than “make this better.”
Verify every external claim. Open the cited sources, check quotations, confirm dates, and remove facts the model cannot support.
Rewrite rather than accept. Read the suggestion, close it, and express the useful idea in your own words. This tests whether you understood it.
Finish without the model. Conduct the final pass yourself. Make sure you can explain every sentence and defend every conclusion.

This workflow may feel slower than asking for a complete draft. It is still faster than beginning without support, and it preserves more of the cognitive work that makes future writing easier.

The three-pass method

For everyday work, a simple three-pass process is easier to remember.

Pass one: Human position. Write the purpose, audience, core claim, evidence, and rough structure without AI.
Pass two: AI challenge. Ask the model to find weak logic, missing context, alternative interpretations, and unclear language.
Pass three: Human responsibility. Decide what to accept, verify the facts, restore your voice, and approve the final text.

The order is important. Beginning with human intent reduces anchoring. Ending with human responsibility prevents the tool from becoming the final authority.

Signs that AI is becoming a crutch

Dependence is not measured by how often you open a chatbot. It is better measured by what happens when the tool is unavailable.

You cannot state the argument without reopening the generated draft.
You remember the wording but not where the evidence came from.
You accept a paragraph because it sounds complete, even though you cannot explain its logic.
Your writing voice changes dramatically from task to task.
You stop making outlines because the model can generate one instantly.
You avoid difficult revisions by asking for a fresh draft.
You feel unable to begin, continue, or finish without the assistant.

None of these signs proves permanent skill loss. They do indicate that the tool may be replacing practice rather than supporting it. The remedy is not necessarily to stop using AI. It is to reclaim one or two high-value stages of the process and complete them independently.

A detector cannot tell whether you learned anything

AI detectors and watermarks address provenance, not cognition. A detector may estimate that a model influenced the language. A watermark may indicate that text passed through a particular system. Neither can reveal whether the person formed the ideas, understood the sources, learned from an example, or accepted the output without reading it.

This is why schools and employers should be cautious about treating an AI score as a verdict on effort or authorship. As AIGator’s analysis of Claude’s invisible text watermark explains, a positive signal may show that Claude processed the text, not that Claude supplied every idea or that the human contributed nothing.

A better evaluation looks at process evidence: notes, drafts, version history, source records, revision decisions, and the writer’s ability to explain the work. Those signals are less convenient than a percentage, but they are much closer to the question institutions actually care about.

What writers, editors, and teachers should do

For writers

Decide which skills you want to keep. You may be happy to automate routine formatting or a standard email. You may want to protect your ability to build an argument, tell a story, synthesize sources, or develop a distinctive voice. Use AI most heavily where the capability is not central, and more cautiously where practice matters.

For editors

Ask authors to identify the human contribution, not merely whether AI was used. Require links to primary sources, inspect substantive revisions, and make sure a named person accepts responsibility for the final claims. Judge the article on evidence and usefulness, then use provenance tools as supporting information rather than proof.

For teachers

Design assignments that expose thinking. Staged submissions, research notes, short oral explanations, source annotations, and reflective memos make the process visible. They also let students use AI in ways that can be discussed and evaluated instead of driving use underground.

The real risk is passive fluency

ChatGPT’s greatest strength is that it can make unfinished thought look finished. It supplies transitions, balance, confidence, and closure before the writer has necessarily earned them. That smoothness can be useful. It can also conceal gaps that productive struggle would have exposed.

The goal is not to romanticize difficulty. Writers do not become better by suffering through every repetitive task. The goal is to keep the forms of effort that produce understanding while automating the forms that merely consume time.

Current evidence does not support the claim that all AI use makes people worse writers. It does support a narrower and more actionable warning: when the model consistently performs the thinking, the writer gets less practice thinking through language. When the model provides examples, feedback, and challenge, it can accelerate learning.

The question to ask after using AI is therefore not only, “Is the draft better?” Ask, “What can I now do that I could not do before?” If the answer is nothing, the tool improved the document. If the answer is that you see structure more clearly, recognize a weak argument faster, or can write the next version unaided, it may also have improved the writer.

Frequently asked questions

Does ChatGPT reduce writing ability?

It can reduce practice and engagement when it replaces planning, argument, drafting, and revision. Research also shows that AI examples and feedback can improve later unaided performance in some structured writing tasks. The result depends on the workflow and the skill being measured.

Does using ChatGPT damage the brain?

The available evidence does not establish brain damage. The MIT essay study found different EEG activity, recall, and ownership patterns across conditions, but it was a small preprint and cannot support broad claims about permanent neurological harm.

Can AI make someone a better writer?

Yes, in some contexts. Large preregistered experiments found that participants exposed to AI-assisted cover-letter practice or strong AI-generated examples performed better on later writing tests without AI. Whether that transfers to other genres and long-term repeated use remains an open question.

Is it better to draft first and use AI later?

For preserving ownership and independent thought, that is a sensible default. Writing notes or a rough draft first gives you a position against which to evaluate the model’s suggestions, rather than allowing its first answer to define the piece.

What is the best prompt for learning to write?

Ask for critique, questions, examples, and explanations instead of a complete replacement draft. For example: “Identify the three weakest reasoning steps, explain why they are weak, and ask me questions that would help me repair them.”

Can an AI detector measure how much help a writer received?

No. Detector scores are probabilistic estimates about language patterns. They do not measure idea ownership, effort, understanding, source verification, or learning. They should not be treated as a complete record of the writing process.

Sources and research caveat: This article was prepared on October 27, 2026, using the MIT Media Lab’s summary of Your Brain on ChatGPT, the 2026 randomized study AI Assistance Reduces Persistence and Hurts Independent Performance, the preregistered writing experiments reported in Coach not crutch, and Noy and Zhang’s experimental study of generative AI and professional writing productivity. Several findings discussed here come from preprints or controlled tasks. They should not be generalized into claims about permanent cognitive decline or every form of writing without further longitudinal evidence.