Exam prep · Updated 9 September 2026
How to Prepare for the New 2026 TOEFL With AI (Safely)
The TOEFL was rebuilt on 21 January 2026, and almost every AI study guide you will find still drills tasks that are no longer on the test.
Practise the tasks that are actually on the test now. Reading and Listening adapt mid-section, scores come back on a 1–6 band scale aligned to the CEFR, and the long independent essay and the four classic speaking tasks have been replaced by short practical items: repeating sentences aloud, answering interview questions with no preparation time, writing an email. AI is genuinely useful for the interview and the writing tasks, close to useless for the repeat-aloud task, and actively dangerous in one specific way — it hands you a template, and ETS warns in its own FAQ that a template gives you "a high probability of providing responses similar to those of other test-takers," which its detection software looks for.
What changed on 21 January 2026
ETS announced the redesign in a press release dated 29 May 2025 and delivered it in January 2026. Two things moved at once: the shape of the test, and the scale your score is reported on.
| TOEFL iBT | Before January 2026 | Now |
|---|---|---|
| Reading | 20 questions, 35 min | 50 items, about 30 min, adaptive |
| Listening | 28 questions, 36 min | 47 items, about 29 min, adaptive |
| Speaking | 4 tasks, 16 min | 11 items, about 8 min |
| Writing | 2 tasks, 29 min | 12 items, about 23 min |
| Score scale | 0–120 total, 0–30 per section | 1–6 in half points, CEFR labelled |
The section times above are the approximate base times ETS publishes, and they add up to about 90 minutes of testing. ETS still tells you to "allow approximately two hours to complete the test," because that covers check-in and instructions as well — do not book your day around the 90-minute figure the prep blogs quote. ETS also notes that "as the test adapts, test time and items may vary."
On scoring, ETS states that the new 1–6 scale runs in half-point increments and that "the overall score will be the average of all four sections, rounded to the nearest half band (for example, if the average of your section scores is 5.25, you will receive an overall score of 5.5)." That is an average, not a sum, so one weak section now drags the headline number in a way adding four 0–30 scores never did. For a two-year transition you also get a comparable 0–120 overall score, and ETS says scores are delivered "within 72 hours or less."
The tasks you are actually preparing for
ETS describes the Reading and Listening sections as using "a multistage adaptive format, adjusting in real time midway through each section based on student performance." Speaking and Writing were rebuilt around shorter, more practical tasks.
| Section | Task types now |
|---|---|
| Reading | Complete the Words · Read in Daily Life · Read an Academic Passage |
| Listening | Listen and Choose a Response · Listen to a Conversation · Listen to an Announcement · Listen to an Academic Talk |
| Speaking | Listen and Repeat · Take an Interview |
| Writing | Build a Sentence · Write an Email · Write for an Academic Discussion |
In ETS's wording, Listen and Repeat means "you will listen to short sentences and repeat them exactly as you hear them," measuring whether your speech is "accurate and clearly intelligible." Take an Interview is "a simulated interview related to academic or campus situations" where you "answer questions about your experiences and opinions." Write an Email asks for "an email in an academic or social situation, such as making a request, giving information, or proposing a solution." And ETS says Writing scoring "focuses on how clearly and effectively you communicate your ideas, not on writing a perfect first draft" — which is a direct argument against polishing your practice answers with a chatbot.
ETS does not publish per-task timings. Prep publishers covering the new format report seven sentences in Listen and Repeat, four interview questions at 45 seconds each with no preparation time, ten Build a Sentence items, about seven minutes for the email and ten for the discussion post. Those counts reconcile exactly with ETS's own item totals — 7 + 4 = 11 speaking items, 10 + 1 + 1 = 12 writing items — so treat them as a reasonable guide, but reported rather than official.
Why most TOEFL AI advice is now aimed at the wrong test
- "Get ChatGPT to grade my 300-word essay." The long independent essay is gone. What you write now is an email and a discussion-board post, both short and both judged on whether the reader gets your point.
- "Practise summarising the lecture in 60 seconds." The integrated speaking tasks that combined a reading, a lecture and a spoken summary are gone.
- "Use your 15 seconds of prep time to plan three points." The preparation window is reported to be gone from the interview task. If you have trained yourself to plan before speaking, you have trained a habit the test no longer gives you room for.
- "Aim for 100 out of 120." Your report is now a set of bands with CEFR labels. Targets published before 2026 need translating before you can chase them.
Note also that only some of this is human-judged. Prep sources report that Build a Sentence and Listen and Repeat are machine-scored against predefined answers — there is no rewarding your creativity on a task with one correct word order.
The template trap
This is the part no competing guide covers, and it is the reason to be careful with a chatbot rather than just enthusiastic. ETS's own testing-policies FAQ answers the question directly.
On memorised answers: "Using memorized answers is a violation of testing policies. ETS raters can recognize when written responses are memorized, and ETS uses detection software to identify writing passages that are similar to outside sources or other test takers' responses."
On templates: "They can be helpful to establish a general structure for your response, but if they do more than that, you're probably violating ETS testing policies ... If you use a template, there is a high probability of providing responses similar to those of other test-takers. So we strongly recommend that you produce your own responses during the test itself."
Read that against how people use chatbots. Ask ChatGPT for a model TOEFL email and you get a well-formed answer — and so does everyone else asking the same question. Convergence is the whole point of a language model: it produces the most likely continuation, which makes chatbot phrasing the least original phrasing available. That is precisely what similarity detection is built to find. The risk is not being caught cheating; it is an honest candidate who revised from a memorised model answer writing something that matches other candidates.
ETS also states plainly that "writing assistance software is not permitted during the test," and that "if you violate the test rules, ETS can cancel your scores, ban you from testing, and refer the matter to law enforcement, as it deems appropriate." Use AI to build the skill, never to supply the words.
Four prompts that fit the new tasks
1. The interview drill with no prep time. This is the highest-value prompt here, because the removal of preparation time is the biggest behavioural change on the test. Speak your answers out loud on a 45-second timer — typing them defeats the exercise.
Act as a TOEFL-style interviewer on one campus or everyday topic of your
choice. Ask me four short questions about my own experiences and opinions,
one at a time. Do not give me the topic in advance and do not give me
preparation time.
After each question, wait for my answer before asking the next.
After all four, give me feedback only on: clarity, whether my pace stayed
natural, vocabulary range, and grammar accuracy. Do not comment on whether
my opinions were good ones. Do not rewrite my answers - tell me what to
change and let me say it again.2. The email critic that refuses to write. Register is what this task really tests: a request to a professor and a note to a classmate are not the same email. Write yours first, unaided, on a timer.
Here is a situation: [describe the scenario and who the email is to].
Here is the email I wrote in 7 minutes, unedited: [paste your own email].
Do not rewrite it and do not give me a better version.
Tell me: (1) is the level of formality right for this reader, (2) did I
actually complete the task I was asked to do or did I miss part of it,
(3) any sentence a reader would have to re-read, and (4) one grammar
pattern I got wrong more than once. Then give me a different scenario and
let me try again.3. The discussion board with opinionated classmates. The academic discussion task asks you to state and support your own opinion in a thread, so practise entering a conversation rather than starting one.
You are a professor posting a discussion question for an undergraduate
class, plus two students who have already replied with opposing views.
Pick any general academic topic. Keep each reply under 80 words.
I will write my own reply. My reply must add something neither student
said, not just agree with one.
After I post, tell me in one line whether I genuinely added a new point or
only restated theirs, and whether I supported it with a specific example.
Do not write a model reply for me.4. The sentence-order drill. Build a Sentence is machine-scored word order, so this is pure grammar reps — the one place where volume beats subtlety.
Give me one everyday English sentence of 8 to 14 words, then show me its
words in scrambled order on a single line. Do not show me the original.
I will type my reconstruction. Tell me only "correct" or "not correct" and,
if not correct, name the grammar rule I broke in under 10 words. Then give
me the next one. Keep going until I say stop.Every one of these makes you produce the language first. That is the same principle behind our guide to building Socratic AI tutors for exam prep, and it is the only way to use a chatbot on a language test without letting its voice replace yours.
What AI cannot do for you here
- Judge Listen and Repeat. Voice assistants are built to understand you in spite of an accent, so they will cheerfully transcribe a sentence you mangled. Record yourself against the source audio and listen back instead — and use official practice for anything you want scored.
- Predict your band. A model has no access to the scale, the adaptive routing or the raters. It will guess with total confidence.
- Write realistic adaptive practice. Reading and Listening now route on your performance mid-section. No chatbot can reproduce that, and its guess at item difficulty is not calibrated to anything.
- Be trusted on facts. Every claim about fees, dates, score requirements or policy needs checking at the source — the habit in our AI hallucination checklist for students.
Where your real practice should come from
- Official ETS practice for calibration. ETS's own TOEFL TestReady platform offers scored section tests and task-level practice with feedback on speaking and writing, with some section-test content free to sample. It is the only practice built against the real scale.
- AI for reps between those tests, using the four prompts above — skill building, never answer supply.
- Your own error log. One line per mistake: task type, what went wrong, the fix. After two weeks it is your study plan.
- Daily speaking outside the test frame. Fluency under time pressure is built by talking, not by drilling. Our guide to using AI as a language-learning speaking tutor covers the conversational half, and the grammar and conjugation workflow covers the accuracy half.
- The application it feeds. A TOEFL score is one item on a graduate application. The rules for the rest of it are stricter than most applicants realise — see what the AI rules actually say for grad school personal statements.
FAQ
Can I use AI to prepare for the TOEFL?
Yes. Nothing restricts what you use while preparing, so having a model run interview drills, critique an email you wrote yourself or quiz you on vocabulary is ordinary studying. Two limits: ETS states that writing assistance software is not permitted during the test, so every habit has to work unaided; and ETS states that using memorized answers is a violation of testing policies, which is exactly what happens if you learn a chatbot answer by heart.
What changed about the TOEFL in January 2026?
From 21 January 2026 score reports use a 1–6 scale in half-point increments aligned to the CEFR, with the overall score the average of the four sections rounded to the nearest half band, plus a comparable 0–120 overall score for a two-year transition. ETS also moved Reading and Listening to a multistage adaptive format that adjusts midway through each section, and rebuilt Speaking as Listen and Repeat plus Take an Interview, and Writing as Build a Sentence, Write an Email and Write for an Academic Discussion.
Is it risky to memorise a ChatGPT template for the email task?
Yes, and ETS says so itself. Its FAQ warns that memorised answers violate testing policies, that "ETS uses detection software to identify writing passages that are similar to outside sources or other test takers' responses," and that with a template "there is a high probability of providing responses similar to those of other test-takers." Thousands of candidates prompt the same chatbot for the same task, so chatbot phrasing is the least original phrasing available. ETS states that consequences can include cancelled scores and a ban from testing.
Can AI check my pronunciation for Listen and Repeat?
Not reliably. The task asks you to repeat short sentences exactly as you hear them, and ETS says it measures whether your speech is accurate and clearly intelligible. General voice assistants are built to understand you despite an accent rather than to judge how intelligible you were, so they will transcribe a sentence you mangled. Use official practice for scoring, and a plain recording of yourself against the audio for everything else.
What is a good score on the new 1–6 scale?
There is no universal answer, because the requirement is set by the programme you apply to. What the scale gives you is a translation: each section carries a CEFR label, and ETS states that a score of 5 on any section aligns with C1 proficiency for that section. Until January 2028 your report also carries a comparable 0–120 overall score, so check whether your programme has published its requirement in the new bands or the old numbers before setting a target.
Bottom line
Check the date on any TOEFL advice before you follow it. If it talks about a 300-word essay, integrated speaking summaries or a score out of 120 without qualification, it was written for a test you will not sit. Practise the interview with no prep time, write the email yourself on a seven-minute clock, drill word order for volume, and take your calibration from official ETS practice. Use a chatbot to ask, time and critique — never to supply the sentence you will remember on test day, because ETS built its detection around exactly that.