Do templates hurt your TOEFL score? We measured 3,692 answers.

36.2% of responses (1,337 of 3,692) open with a stock phrase, and 55.8% (2,060 of 3,692) contain at least one anywhere. Template use is normal, not a fringe tactic. Any page telling learners that templates are a red flag is telling better than one in three of the cohort that their normal writing is a red flag.

The second finding in this table is that the specific phrases prep sites warn about are not the phrases learners use. "There are several reasons" appears as an opener twice in 3,692 responses. "It is widely acknowledged" twice. "I could not agree more" zero times, and that is a genuine zero from a matcher verified against a known-positive case first. The real template vocabulary is opinion-framing: the top two phrases alone are 24% of openers.

Table 1. Prevalence of fifteen stock phrases, n = 3,692. "As opener" means the phrase matches in the first 90 characters.

Responses that open with a template average 25.198 on the 0-30 scale. Responses that do not average 25.158. The difference is +0.040, 95% CI -0.119 to +0.199, t = +0.49. There is no detectable effect in either direction.

With these group sizes, the minimum detectable effect at 80% power and two-sided alpha = 0.05 is 0.232 points. This is not an underpowered null hiding a large effect: a penalty of a quarter of a point or more would have shown up, and did not.

The same result from the other end

If the grader punished templates, the top band would be depleted of them. It is not. Top scorers use template openers at essentially the corpus rate.

Each response carries three 0-5 sub-scores. This is the test of the specific mechanism the search results assert, which is a language penalty on formulaic phrasing. If templated language were being flagged as formulaic, Language Use would fall while Relevance stayed flat.

Table 2. Sub-score means, n = 1,335 template / 2,348 no-template for each dimension.

The observed pattern is the reverse ordering. Language Use is the flattest of the three, and no dimension is negative. Read this carefully, because it is easy to over-read: all three differences are positive and all three are practically zero. A difference of 0.025 on a 5-point scale quantised in 0.5 steps is one twentieth of the smallest step the grader can award. Nothing here says templates help. What it says is that the asserted mechanism does not appear.

After controlling for length the picture is the same: Relevance beta = +0.029 (t = 2.09), Clarity beta = +0.030 (t = 2.17), Language Use beta = +0.018 (t = 1.16). Three dimensions were tested, so t values near 2 should be read as marginal, and the effect sizes make the question academic.

A number is only defensible if the obvious alternative explanations were ruled out. Five were.

Of the fifteen phrases, nine occur as openers at least 30 times and could be tested individually. Eight are indistinguishable from the baseline, with differences from -0.08 to +0.57 and confidence intervals spanning zero or nearly so. One is not.

Responses opening with “When it comes to…” (n = 75) average 23.573, which is 1.585 ± 0.512 below the no-template baseline of 25.158. The effect survives every control. These responses are longer than baseline, 209.8 versus 161.2 words, and still score lower, so length-adjusted the gap widens to -2.20 (SE 0.286, t = -7.7). It is spread over 56 distinct tasks with no single task contributing more than 7 responses, so it is not one bad prompt. Nine phrases were tested; at a Bonferroni threshold this one still clears comfortably. Removing it from the treatment set moves the main contrast to +0.136 ± 0.160, still null.

Table 3. Per-opener means for the nine phrases with n >= 30, against the no-template baseline of 25.158.

Population

Every non-legacy Academic Discussion task in LingoLeap's task registry (1,480 tasks) and all of its representative sample slugs, 3,729 distinct slugs. Academic Discussion is the only 2026-format task with a real corpus, which is why it is the population.

Retrieval

Each sample was fetched from the evaluation-sample endpoint with 8 concurrent workers and backoff-and-retry on HTTP 503. The backend rate-limits above roughly 8 concurrent reads; an unthrottled first pass at 16 workers lost 784 slugs to 503s, which is a retrieval artefact and not missing data, because every one of them came back on retry. Final: 3,723 records retrieved, 6 hard 404s, 0 unrecovered.

Two fields, and they are different things

The record nests a content field and a markdown field. The content field is the learner's response, and every measurement of template language is taken on that field only. The markdown field is the evaluator's commentary, which for cn-locale records (88% of the corpus) is largely written in Chinese and contains a rewritten model essay. A regex run over the markdown field would measure the grader's prose, not the learner's. It is used only to read scores.

Scoring fields

A numeric score field is present on 3,258 records. For the remaining 465 the score is parsed out of the markdown. On the 2,459 records where both are present they agree on 2,459 of 2,459 (100%), which is the validation for using the parsed value as a fallback. Per-dimension sub-scores parse from the markdown on 3,703 of 3,723 records (99.5%).

Exclusions and analysis set

13 records with no response text and 8 records scoring 0 with 0.0 on all three sub-dimensions (grader failures, not essays) were dropped. Analysis set: n = 3,692 responses across 1,467 distinct tasks. Mean score 25.17 (SD 2.42), median 26, IQR 24-27. Mean length 160 words, median 154.

Treatment definition

Fifteen stock phrases, derived from the data rather than assumed. Every response's first 120 characters were tokenised and the leading 3-, 4- and 5-word sequences counted across the corpus; the recurring formulae were then written as case-insensitive regexes. A response is a template-opener if any stock phrase matches in its first 90 characters.

Known-positive verification

Before any counting, the matcher was run against a response opening “As far as I am concerned, I firmly believe that…”, which matched, and against a non-formulaic opener, which did not. Both behaved as expected, so every zero in Table 1 is a zero from a matcher proven to fire on a positive case.

Sampling

These are representative sample slugs from the task registry, up to about 3 per task, not a random draw from all 16,479 Academic Discussion responses. They are selected for being published sample pages, which may over-represent well-formed responses. Nothing in the selection is correlated with template use, but the population is published samples, not all submissions.

The strongest version of a claim is the one that gets you caught. Six things this study does not establish:

1. 1

   It is not evidence about ETS's e-rater or SpeechRater

   The scores here are produced by LingoLeap's own rubric-trained AI grader. We have measured how an AI scoring system trained on the Academic Discussion rubric treats templates. We have not tested the engine ETS uses, and nobody outside ETS can.

2. 2

   It is not evidence that templates help

   Every positive number in this study is smaller than the grader's own resolution. The correct verdict is no detectable effect, not that templates are safe and certainly not that templates work.

3. 3

   It does not cover speaking

   Academic Discussion is writing. Claims about spoken templates are untested here.

4. 4

   It does not generalise to all TOEFL responses

   The population is LingoLeap's published sample responses to non-legacy Academic Discussion tasks, up to about 3 per task.

5. 5

   It says nothing about content templates

   We measured opening formulae. Memorised example paragraphs, invented statistics and pre-written body arguments are something this study did not look at. A learner who memorises an entire essay is doing something else.

6. 6

   It cannot separate the template from the template user

   This is observational. Nobody was randomly assigned a template. The within-task and within-length-quintile estimates narrow the confound but do not remove it.