I asked the essay writer on this site a stock question - should homework be banned in primary schools - and read the result twice. The structure was sound: a title, a thesis at the end of the introduction, three body paragraphs each making one point, a conclusion. It invented no statistics, no quotations and no fake journal articles. Then I reached the end, where it had added a references section listing the American Psychological Association and Harvard beside titles that do not exist, each politely marked "Hypothetical citation - imagine a real article". That is the shape of the problem in one page: the thing is a competent scaffold with a trap at the bottom, and the trap looks like homework already done.
TL;DR: A small language model writes a well-structured first draft and no usable evidence. Measured over two runs of this one: word counts landed within 8% of target, zero invented statistics or quotations, but four placeholder citations naming real institutions beside imaginary titles. Use the draft for its skeleton, replace every sentence of its content with your own, and check your institution's policy before any of it goes near an assignment. Also know that AI detectors are not a reliable check on any of this, in either direction.
What does it genuinely do well?
Structure, and that is not nothing.
The hard part of a first draft for most people is not the sentences, it is the shape: where the thesis goes, how many points the body needs, what a paragraph is for. A model trained on a great deal of expository prose has that shape memorised. Both drafts I generated put the thesis as the final sentence of the introduction, gave each body paragraph a single claim, and closed without introducing new material. That is the standard taught in every composition class, executed correctly on the first attempt.
Length control was better than I expected. Asked for 500 words, the two runs produced 538 and 526. That matters because padding to a word count is the other thing students lose hours to, and seeing a correctly-sized draft tells you how much room three points take.
What it is doing is pattern completion over essay-shaped text. That makes it good at the parts of an essay that are conventional, and bad at exactly the parts that are supposed to be yours.
Does it make up sources?
Not quite, and the distinction matters more than a yes or no.
The prompt behind this tool instructs the model not to invent quotations, statistics or sources, and to describe the kind of source you should cite instead. Measured across both runs, that held: no percentages, no "studies show", no fabricated DOIs, no invented author names with plausible years.
What appeared instead was a references section of this form:
2. ... Research Publication, [Year]. (Hypothetical citation - imagine a real article)
3. American Psychological Association, "Homework and Academic Achievement", [Year].
(Hypothetical Citation - imagine a real report)
Read carefully, that is honest. It is labelled. It tells you to find the real thing. But it names a real organisation next to a title that organisation never published, and it is sitting in a numbered list under a heading, which is the exact visual form of a finished bibliography. A tired student at 1am deletes the words in brackets and submits it. That is not a hypothetical failure mode, it is the obvious one, and it is worth naming because the tool's own caution is what creates it.
The rule that follows: treat anything resembling a citation in an AI draft as a note to yourself about what to go and find. Never as a reference.
Why do small models hallucinate at all?
Because they are not retrieving anything. A language model predicts likely continuations, and a plausible-looking citation is a likely continuation of a sentence that needs evidence. There is no lookup, no database, no moment where the model checks whether the APA published that report.
Size makes this worse. This tool runs Gemma 3 1B, a one-billion-parameter model, chosen because it fits on a CPU-only server and can be offered free. Smaller models have less memorised detail and lean harder on pattern, so they are more prone to confident invention than the large commercial ones. The structural competence survives the small size; factual reliability does not.
This is also why it knows nothing recent. There is no browsing step, so anything that happened after its training data was assembled does not exist as far as the draft is concerned.
Will a teacher's AI detector catch it?
That is the wrong question, and the honest answer is uncomfortable in both directions: detectors catch work that was not AI-written, and miss work that was.
OpenAI, which had every commercial reason to make detection work, withdrew its own AI Text Classifier on 20 July 2023 because of its low accuracy. Its own published evaluation before that: the classifier correctly identified 26% of AI-written text, while flagging 9% of human-written text as AI-written. A tool that catches a quarter of what it is looking for and falsely accuses nearly one in ten innocent writers is not a tool, and its makers said so by retiring it.
The false positives are not evenly distributed either. Liang and colleagues at Stanford tested seven detectors on writing by native and non-native English speakers and found they consistently misclassified non-native writing as AI-generated, published in Patterns in 2023. The mechanism is perplexity: detectors flag text whose word choices are predictable, and a competent writer working in a second language produces exactly that. The authors' own recommendation is that these detectors should not be used in evaluative or educational settings.
So: do not rely on a detector to vindicate you, and do not assume one will catch you. Neither is a reason to submit AI work where it is not allowed. The reason not to is that it is not allowed, and the people marking your work can usually tell by other means.
What is the honest way to use a draft like this?
Three uses that survive scrutiny, and one that does not.
As a shape to argue with. Generate the draft, read its three points, and notice which one you disagree with. The disagreement is your essay. This is the most useful thing it does, because reacting to a position is far easier than inventing one from a blank page.
As a check on your own structure. Write your essay first, then generate one and compare the skeletons. If the model found a point you missed, go and research it properly.
As a vocabulary and register sample. The academic-level setting changes how formal the prose is. Seeing the difference between the high school and graduate versions of the same argument teaches something about register that is hard to explain abstractly.
Not as a thing to submit. Many institutions ban or restrict AI-written work, and the policies differ by course, not just by university. Check yours before you use any of this in an assignment. Where it is permitted, disclosure is usually part of the permission.
Does the draft read as machine-written?
Yes, and measurably so, in ways that are easy to learn to spot.
I ran both drafts through the same checks this site applies to its own writing. Across roughly 530 words each, the output carried five en dashes and, from the list of words that mark generated prose, "foster" three times, plus "holistic" and "crucial". None of those is wrong English. All of them are the words a model reaches for when a human would pick something plainer: encourage rather than foster, whole rather than holistic, important rather than crucial.
The other tells are structural. Every paragraph is about the same length. Every list uses the same shape. The conclusion restates the introduction with the words swapped. A marker who reads forty essays a week notices this pattern long before they reach for a detector, which is the real reason the detector question is beside the point.
If you are rewriting a draft into your own voice, those are the things to attack first: vary the paragraph lengths, cut the dash, replace the abstract noun with the concrete one, and make the conclusion say something the introduction did not.
Why does it take so long?
Because it runs on a shared CPU rather than a GPU, and that trade is what makes it free.
Measured on the live site, a 500-word draft took 170 seconds from request to final word. That is the shortest option. A 1,500-word essay takes proportionally longer. The text streams as it is written, so you can read the opening while the rest arrives, which makes the wait less obvious than the number suggests.
The alternative design would be a commercial API, which answers in a few seconds, costs money per request, and sends your topic to a third party who may retain it. This one sends your topic over HTTPS to a server that builds the prompt, runs the model, returns the text, and keeps nothing: the topic and the draft are used to write the reply and then discarded, not logged, not used for training, not passed to any AI company. For an essay topic that is usually not sensitive. For the draft of a personal statement, it matters.
What should you check before submitting anything?
A short list, in the order that catches the most.
- Your institution's policy, before anything else. Course-level rules often differ from university-level ones.
- Every factual claim, against a real source you have opened. The draft will not have invented a statistic, but it will have asserted things confidently with no evidence behind them.
- Every citation, which in practice means replacing all of them.
- The voice, which is where the word list above earns its keep.
- The argument, last. If you still agree with all three of the model's points after researching them, you have written the model's essay rather than yours.
For the rewriting itself, the paraphrasing tool changes wording but not substance and will not fix any of the above, and running a draft through the text humanizer addresses the tells rather than the thinking. Neither is a substitute for having something to say.
What about the word count and the formatting?
Both are better handled afterwards, by tools that are certain rather than probabilistic.
A model asked for 500 words produces approximately 500 words, which is fine until a brief says "500 words maximum" and the draft lands at 538. Paste the final text into the word counter and know the number rather than estimating it. The same goes for reading level and sentence length, which most markers care about more than they admit.
Formatting is the other thing worth separating. The draft arrives as plain text with markdown-style headings, which is convenient for editing and wrong for submission in most courses. Convert it once at the end, when the words have stopped changing, rather than fighting the formatting through six revisions.
Do the seven essay types change anything?
More than the academic level does, which surprised me.
The tool offers argumentative, persuasive, expository, narrative, descriptive, compare and contrast, and cause and effect. Those are not cosmetic labels on the same draft: they change the job the body paragraphs are doing. An argumentative draft takes a position and defends it against an implied objection. An expository one explains without taking a side, so the paragraphs become categories rather than claims. Compare and contrast reorganises everything around two subjects, which forces either a point-by-point or a block structure, and the model picks one and holds it.
The academic level setting does less. It adjusts vocabulary and how much hedging appears, so the graduate version of an argument says "the evidence suggests" where the high school version says "this shows". The underlying argument is usually the same three points. That is worth knowing if you were hoping the graduate setting would produce graduate-level thinking: it produces graduate-level register over the same reasoning.
The practical consequence is to pick the type carefully and the level casually. A persuasive draft when you needed expository is a structurally wrong essay; a high school draft when you needed undergraduate is twenty minutes of rewording.
Is a large commercial model better for this?
At writing, yes. At the things that get you in trouble, not obviously.
A large model writes more fluently, holds a longer argument together, and is far less likely to produce the placeholder-citation pattern described above. If the only criterion is prose quality, it wins and it is not close.
Three things it does not fix. It still has no lookup step, so it still invents confidently - the failures are rarer and therefore harder to catch, which is not the same as safer. It still produces the structural tells, because those come from the training distribution rather than from model size. And it usually retains your text: most commercial assistants keep conversations by default, and some use them for training unless you opt out, which is a different proposition when the document is a personal statement or an unpublished thesis chapter.
That is the trade this tool makes deliberately: a smaller, less fluent model that runs on a server which keeps nothing, offered free. For a stock essay question, use whatever writes best. For anything you would not paste into a public forum, the retention question is the one to ask first. The same reasoning runs through how we write everything here, where the tells are catalogued in more detail than this page has room for.
