What should you look for in TOEFL Writing AI feedback?

How does ETS say TOEFL Writing is scored?

ETS says Build a Sentence is scored by matching against the correct answer, and the email and discussion responses are scored using both an AI scoring engine and human raters (see the Specifications Document and the Technical Manual below). Of the 12 Writing items on the 2026 updated TOEFL iBT, ETS classifies the 10 Build a Sentence items as Machine Scored and the one Write an Email item and the one Write for an Academic Discussion item as AI Scored (ETS "TOEFL iBT Test: 2026 Update Test Blueprint and Specifications Document", hereafter the Specifications Document, no publication date, checked 2026-10-30, page 6; page numbers throughout this article follow the PDF page order). In the same document, machine scoring is a rule-based method that scores selected responses by matching the correct answer or by predefined scoring logic. AI scoring is described as evaluating constructed responses with natural language processing (technology that lets computers analyze human writing and speech) and machine learning, looking at features such as fluency, coherence and grammar (page 3, note 4). Each Build a Sentence item is worth up to 1 point, and the email and discussion responses are worth up to 5 points each (same document, page 15). ETS "Updated TOEFL iBT Test Overview" (hereafter the Test Overview, no publication date, checked 2026-10-30) says the two writing tasks are scored 0 to 5 points on separate scoring guides (a scoring guide is a table that describes what a response at each score looks like), and it prints those scoring guides on pages 23 to 25.

ETS "TOEFL iBT Technical Manual" (ETS Research Report RR-25-12, October 2025, hereafter the Technical Manual, checked 2026-10-30) says these two tasks are scored using both the ETS AI scoring engine and human raters (page 35). The AI engine is trained on scores given by human raters (page 36). AI scoring is monitored in real time, human raters score responses the AI has low confidence in or finds hard to handle, and certified human raters review a random sample again (page 39). The same document says each test taker's responses are scored by two or more raters (page 40). When 1,914 Writing responses were scored by two human raters and compared, the correlation between the human average score and the AI score was 0.86, and the correlation between one human rater and the other was 0.85 (page 39, Table 7). Build a Sentence is scored against the correct answer and uses neither human scoring nor AI scoring (same page).

ETS also states separately which abilities each task measures (ETS "TOEFL iBT Writing Section", no publication date, checked 2026-10-30).

• Write an Email: whether you write clearly, stay on topic, and follow basic writing conventions
• Write for an Academic Discussion: whether you develop your ideas, respond to other people's perspectives, and write in an academic tone

The ETS Information Bulletin (2026-27) says you can request a Writing score review for a fee within 30 days of your test date, and that the review is done by scoring specialists (ETS "TOEFL iBT Information Bulletin", for July 2026 to June 2027, checked 2026-10-30, page 12). What each of the three tasks asks you to do is covered in the Unlock Prep (unlockprep.com) guide "How do I prepare for the three new TOEFL Writing tasks?".

What does Unlock Prep AI feedback show, and what can it not see?

Unlock Prep AI feedback shows, sentence by sentence and in Korean, which sentences in a Writing response to fix and why. AI judges the response on each task's criteria and gives a task score (out of 5) and a section score (the 1 to 6 band, in steps of 0.5). The site describes these as "criteria that reflect ETS scoring criteria" (unlockprep.com/toefl-mock-test, checked 2026-10-30). The scores produced this way are called reference estimates; they are not scores given by ETS, so they do not appear on an ETS score report.

We list the limits first.

• There are no accuracy figures. Unlock Prep does not publish figures comparing its estimates with actual ETS scores (same page). So this article does not give a match rate or error margin either.
• There is no data showing whether ETS scoring and Unlock Prep AI give the same score to the same piece of writing. Unlock Prep does not publish comparison figures like the human-AI correlation ETS reports. There is also no data comparing it with teacher feedback, so we do not say which is better.
• It serves a different purpose from an ETS score. ETS scores appear on your score report, and in a paid review, scoring specialists look at your response again. A reference estimate is a value for finding what to fix in a practice response.
• In our view, it is hard to tell from the writing alone whether a sentence was memorized in advance. The ETS Information Bulletin says scores may be canceled for writing that is substantially similar to other test takers' responses or that contains wording copied without citing the source (same bulletin, page 14). A high reference estimate does not mean a response is clear of this policy.

This article covers Writing only.

Where do weak sentences lose points?

In our view, points slip away in sentences that do not show the abilities ETS says that task measures. Below are examples we wrote for illustration. They are not taken from real student responses or from Unlock Prep feedback results, the reasons for lost points are our view, and the text in parentheses is the ETS scoring guide line (the Test Overview, pages 23 to 25) or the ETS Writing page wording that each reason ties to.

Prompt 1, written by us (Write an Email): You are sick and need to hand in a lab report late, so write to your professor asking to push the deadline back two days.

• Weak sentence: "Hello, I am a student in your class and I want to talk about many things this week."
• Revised sentence: "Dear Professor Lee, I am writing to ask if I could submit my lab report two days late because I was sick this week."
• Reason for lost points: The request, which is the purpose of the email, is not in the first sentence, and the reader cannot tell what "many things" means. The greeting and tone also do not fit a message to a professor (Write an Email scoring guide, 4-point line: "Adequate elaboration to support the communicative purpose"; "Mostly appropriate social conventions").

• Weak sentence: "yesterday my laptop broke so i cant send the file until i will fix it."
• Revised sentence: "My laptop broke yesterday, so I cannot send the file until it is repaired."
• Reason for lost points: The first letter of the sentence and "I" are lowercase, and the apostrophe in "can't" is missing (ETS Writing page: basic writing conventions; the Technical Manual, page 37, Table 3, Accuracy/Errors column: "Mechanical errors"). The future tense "will" is used after "until", which expresses time (Write an Email scoring guide, 4-point line: "Few lexical or grammatical errors").

Prompt 2, written by us (Write for an Academic Discussion): Are online classes better for university courses? In the discussion, Maya agrees, and Ken disagrees, saying they make it harder to focus.

• Weak sentence: "I think online classes are good because they are good for students."
• Revised sentence: "I agree with Maya that online classes can help students, because recorded lectures let them replay difficult parts. Ken is right that distraction is a risk, but a fixed study schedule can reduce it."
• Reason for lost points: The reason only repeats the claim and gives no support (Academic Discussion scoring guide, 3-point line: "part of an explanation, example or detail may be missing"). It also does not address the opposing view raised in the discussion (ETS Writing page: responding to other people's perspectives).

The revised sentences state more clearly what the writer originally meant to say. When feedback gives you a revised sentence, also check that the meaning has not changed. Where responses at different score levels split on the scoring guide is covered, with two answers we wrote, in the Unlock Prep guide "Where do you lose points on TOEFL Write an Email?".

What differs between sources, and what is still unknown?

The ETS website (ets.org) has two versions of the Writing scoring guide PDF. The 2026 format scoring guides are on pages 23 to 25 of the Test Overview, and the same wording appears in ETS "Writing Scoring Guide" (writing-rubrics.pdf, Write an Email and Write for an Academic Discussion). The other version (toefl-ibt-writing-rubrics.pdf) is an older format that includes Integrated Writing, which is not on the 2026 task list, so this article does not use it. When you look for a scoring guide, first check that the version has a Write an Email page.

The task table in the Specifications Document lists the email and discussion responses only as "AI Scored", but the Technical Manual says human scoring is also used for the same tasks (pages 35 and 39 to 40). This article cites both documents. The Technical Manual says the AI passes low-confidence responses to human raters, but it does not state the threshold for doing so as a number (page 39). It says the sum of task scores is converted to the Writing 1 to 6 band by weighted equipercentile linking (a statistical method that adjusts for small differences in difficulty between test forms) (page 35), and the conversion table is not printed in the ETS documents we checked. The Specifications Document still carries a note that minor changes may be made before launch (page 3).

What should you check in AI feedback results?

When you get feedback, check four things.

• For each reason for lost points, which of the task's measured abilities in the ETS section above it ties to
• Whether the revised sentence changed what you meant to say
• Whether the same reason keeps coming up in responses from other practice rounds (a recurring reason is what to practice next)
• Whether you record the score separately as a reference estimate, and use the 1 to 6 band on your ETS score report when comparing with school requirements

How to read the 1 to 6 band is covered in the Unlock Prep guide "How do TOEFL 1 to 6 scores compare with 0 to 120 scores?".