Evidence review · AI feedback on writing

AI feedback on student writing

A pooled effect of 0.55 is easy to find. Set beside the grade levels, the software and the study designs behind it, it says little about Grades 4 to 9, and the ChatGPT studies done in schools mostly measure the feedback, not what students wrote next.

Ask whether AI feedback improves student writing and you will meet a respectable number. A 2023 meta-analysis pooled 20 studies and reported a medium effect of g = 0.55 for automated feedback on writing performance.1 Before that figure travels to an elementary classroom, it helps to ask who was in those studies and what software they used. Every one of the 84 effect sizes behind it came from a high school sample (23) or a university sample (61).1 I wanted to know what the evidence says for Grades 4 to 9.

The short answer is that the evidence for children in those grades is thin and mostly about older software. A school-year randomized trial in Grades 7 and 8 found no significant overall effects, including on argumentative writing, from an automated writing evaluation system.78 The one study I could read of ChatGPT feedback with elementary children, in Grade 4, reported large gains, but its design leaves too many questions open for me to lean on it.4 For randomized trials of generative AI feedback in Grades K to 8, the answer from the indexes I searched is none found. Two randomized trials of generative feedback, with high school and university students, favour feedback over none, and neither was in these grades. I know the high school trial only from another paper's description, and in the university trial every group's revised essays still scored below its first drafts. My confidence is moderate that older automated feedback tools are linked to better writing in the high school and university studies pooled so far. It is low for those older tools in Grades 6 to 9, low to moderate for generative feedback with older students, and very low for generative AI feedback in Grades 4 to 9.

The claim, stated precisely

The claim I am testing is a composite: AI feedback improves the writing of students in Grades 4 to 9. That is my paraphrase of how such claims travel, and it is not a sentence from any one paper. Three of its parts need pinning down.

"AI" covers two families of tool. The older family is automated writing evaluation, software that uses natural language processing to give immediate scores and feedback.8 The newer family is the large language model, such as ChatGPT, which writes feedback in sentences. Studies of one family do not tell you about the other.

"Improves" needs an outcome: a better second draft, a better piece weeks later without help, or only a better attitude. The studies below measure different ones. "Grades 4 to 9" needs a sample, and most of the evidence comes from students outside that range.

What the pooled evidence pooled

The meta-analysis is the largest pooled source on automated feedback that I read. Its authors' literature search yielded 4,462 entries, and 20 studies met their inclusion criteria, giving 84 effect sizes and 2,828 learners.1 The individual estimates ranged from -1.73 to 2.99, and 70.24% were positive.1

The average hides a lot. The authors report a 95% prediction interval of -1.02 to 2.12 and note that in some studies the true effect may be negative.1 They also write that the heterogeneity means automated feedback "cannot be understood as a single consistent form of intervention".1

For a teacher's question, the educational-level result matters most. The paper compared only two levels, high school and higher education.1 The pooled effect was 0.50 for the secondary level and 0.58 for the tertiary level, and the paper reports that these two did not differ significantly.1 There is no elementary or middle school category in this paper, so it offers no direct evidence for Grades 4 to 9.

Two further results are often flattened in summaries. Interventions of more than two sessions showed an effect of 0.66, against 0.18 for one or two sessions, but the paper reports that difference as not statistically significant (Q = 1.32, p = 0.250).1 And the effect was 0.57 at post-test and 0.27 at follow-up, and the authors say the follow-up effect did not differ significantly from zero.1 So this does not show that longer use works better or that the benefit lasts.

The authors also say that a fuller test for publication bias is needed and that a later meta-analysis should include unpublished or non-peer-reviewed studies.1 The paper describes two earlier meta-analyses, by Zhai and Ma and by Ngo and colleagues, which I was unable to open. It says Zhai and Ma found significant moderating effects of educational level, among other factors.1 That may be where grade-band evidence lives, and I could not check.

What the school-age review found

A 2022 systematic review looked only at school settings and is closer to the question. It searched for studies in Grades 1 to 12 and identified eight, covering six systems and 1659 students aged 11 to 17.2 Six studies were in middle school, Grades 6 to 9, with 1343 students, and two were in high school, with 316.2 Only one of the eight was experimental, and the other seven used quasi-experimental designs.2

All but one showed a positive effect of automated feedback on at least one writing measure.2 The exception, Ware (2014), found no evidence that the Criterion system improved writing quality, length or mechanics compared with online teacher feedback or peer feedback.2 The review's authors describe studies of these systems in Grades 1 to 12 as scarce.2

What the large language model studies tested

The newer research tests chatbots such as ChatGPT. A 2026 systematic review in Frontiers in Education searched Scopus, IEEE Xplore and ERIC for studies published between November 2022 and February 2026 and included 34.3 Of those, 29 were set in universities, three in secondary education and one in an elementary classroom.3

The authors call this a bias. They write that findings from higher education may not transfer readily to secondary or primary settings.3 Of the 34 studies, eight were experimental and eight quasi-experimental.3

I looked at what each of the four school-level studies measured. By the review's table, the outcome in da Silva and colleagues' study was feedback accuracy, in Steiss and colleagues' it was feedback quality, in Mah and colleagues' it was feedback practices, and in Ezoe and Takagi's it was reflection quality.3 Three of those four are measures of the feedback itself.

I read two of the four in full. Mah and colleagues compared feedback from five language models with feedback from 12 teachers on 22 student essays at middle and high school grade levels.5 It is a qualitative comparison of the feedback itself, so it followed no students' writing forward. The authors found that the models mostly gave corrective feedback at sentence level, while teachers gave more dialogic feedback.5

Da Silva and colleagues had GPT-4 assess 33 opinion articles written by students at Brazilian Army military schools.6 Their method describes a second phase in which students revised their articles with the feedback.6 The results I read answer the learning question by classifying the model's feedback with a four-level framework.6 I found no revised scores in them.

I could not read Steiss and colleagues or Ezoe and Takagi. Both are known to me only from the review, whose table lists Ezoe and Takagi's study as Japanese, in an elementary classroom, with 14 students.3 The title in the review's reference list refers to sixth-grade reflective writing in mathematics.3

One more trial belongs here, though I could not open it. Meyer and colleagues conducted a large randomized controlled trial of 459 high school sophomores.5 The students who received feedback from a language model experienced positive benefits in text revision (d = 0.19), task motivation (d = 0.36) and positive emotions (d = 0.34) compared with students who received no feedback.5 At least two authors of that trial, Fleckenstein and Liebenow, also wrote the 2023 meta-analysis.15 So the trial and the meta-analysis are not independent of each other. I searched the 2026 review for a citation of the trial and found none, so its count of secondary-level studies is not a complete census.

A newer randomized trial with adults points the same way. It was published on October 30, 2026, and I found it when I re-ran the search before publishing. Liu and Shi randomized 312 first-year undergraduates at one university to feedback from a language model, feedback from a teacher or no substantive feedback, and 287 were analyzed.10 Two raters scored the essays without knowing the condition.10 Revised essays scored lower than first drafts in all three groups, with a mean change of -13.24 points for control, -7.09 for AI feedback and -7.04 for teacher feedback.10 The gaps against control (d = 0.60 for AI and d = 0.59 for teacher feedback) therefore reflect a smaller decline, and the authors say they should not be read as evidence that feedback raised essays above their first-draft level.10 The authors judged the two feedback conditions equivalent within a margin of d = 0.35, which they set themselves.10 The equivalence test had modest power (0.55), equivalence did not hold at d = 0.25 or below, and the trial was not preregistered.10 The study ran two supervised sessions, with a single feedback-and-revision cycle on one argumentative prompt, so it cannot show what happens to later, unassisted writing.10

The studies closest to Grades 4 to 9

Two primary studies tested feedback on students in or near these grades and measured their writing.

A randomized trial in the middle grades

The larger one tested an older, non-generative tool. MI Write is an automated writing evaluation system developed and marketed by Measurement Incorporated.8 The trial compared students who used MI Write for one school year, 1,260 of them, with 1,227 who did not.7 It looked at argumentative writing performance, self-efficacy, enjoyment of writing and beliefs in the recursive nature of writing.7 The abstract reports that no significant overall treatment effects were found, though significant treatment-by-district interactions emerged.7

Several design facts come from two companion papers, which I read in full. I am treating them as describing the same trial, because the authors, the tool and the grades match, but I found no sentence that says so. In the teacher study, teachers were randomly assigned, 19 to MI Write and 18 to a comparison group.9 The students whose survey responses were analyzed were in Grade 7 and Grade 8, across three districts, during the 2021-2022 school year.8 The Gates Foundation funded the trial and selected MI Write for it.9 Of the survey paper's eight authors, three were employed by Measurement Incorporated and five declared no conflicts of interest.8

I could open only the abstract record of the main paper, because the publisher blocks scripted access. I cannot say how the writing was scored, whether raters were blind, or what the effect sizes were. The abstract gives 1,260 students in the treatment group.7 The companion survey paper gives 1,299 treatment students.8 I have not reconciled the two figures.

A quasi-experiment with fourth-graders

The other study is the only one I could read in full that tested ChatGPT feedback on writing with elementary children. Kultas ran a quasi-experiment with 60 fourth-graders aged 10 at one school in Van, Turkey, and 30 of them received six weeks of ChatGPT-supported formative feedback on poems.4 The groups were intact classes, because the author says random assignment of individual students was not feasible.4

The result looks striking. The ChatGPT group's mean poem score rose about 30 points to 63.07, while the control group's rose about 8 points to 40.13.4 After adjusting for the pretest, the gap was 22.32 points, with a 95% interval of 18.96 to 25.67.4 The paper reports an eta-squared of 0.757 for the group effect and calls it large.4

I hold back from relying on it, for reasons that come from the paper itself. The control group got only general verbal comments from the teacher to the whole class, and these were not structured according to a rubric.4 So the study compares structured feedback with general comments made to the whole class. Eight students judged to have low basic writing skills were dropped after baseline, five from the ChatGPT group and three from the control group.4

The scoring is described in two ways. The paper says the poems were evaluated by the researcher using a rating scale, and also that outcome assessment was done by a rater who was blind to group allocation.4 Those statements conflict without more detail, and I was unable to resolve them. The study had a single school and a convenience sample, and it reports no financial support.4

The evidence at a glance

The evidence, claim by claim
ClaimDesignSample and settingWhat it found
Automated feedback is associated with better writing performance in pooled studies 1Three-level random-effects meta-analysis of (quasi-)experimental studies; journal article20 studies, 84 effect sizes, 2,828 learners; 23 effect sizes from high school and 61 from university samples; no elementary or middle school categoryg = 0.55 overall; 95% prediction interval -1.02 to 2.12. Secondary 0.50 vs tertiary 0.58, not significantly different. Follow-up effect 0.27, not different from zero. The text does not mention ChatGPT or generative AI.
Automated writing evaluation helps writing in school settings 2Systematic review (PRISMA), narrative; 7 of 8 studies quasi-experimental8 studies, 1659 students aged 11 to 17; 6 studies in middle school (Grades 6 to 9), 2 in high schoolAll but one study showed a positive effect on at least one writing measure. The one experimental study (Ware, 2014) found no evidence that the Criterion system improved writing quality, length or mechanics compared with online teacher feedback or peer feedback. Nothing below Grade 6.
A year of an automated writing evaluation tool improves middle school writing 7Randomized controlled trial, one school year; teachers randomly assigned (19 vs 18); non-generative tool (MI Write); main report read as abstract only1,260 students using the tool vs 1,227 not; Grades 7 and 8; three districts; 2021-2022No significant overall treatment effects on argumentative writing, self-efficacy, enjoyment or beliefs about writing; significant treatment-by-district interactions. Scoring method, blinding and effect sizes not read. Funded by the Gates Foundation; three authors employed by the tool's developer.
Language model feedback improves writing, by educational level 3Systematic review, narrative; Scopus, IEEE Xplore and ERIC, November 2022 to February 2026; no pooled effect34 studies: 29 university, 3 secondary, 1 elementaryThe authors flag a university bias and say findings may not transfer readily to secondary or primary settings. Eight studies experimental, eight quasi-experimental.
ChatGPT feedback improves primary school poetry writing 4Quasi-experiment, intact classes, pretest-posttest control group; six weeks; scoring described both as by the researcher and by a blind rater60 fourth-graders aged 10, one school in Van, Turkey; 8 students dropped after baselinePosttest means 63.07 vs 40.13; adjusted difference 22.32 points (95% CI 18.96 to 25.67). Control got only whole-class verbal comments. No delayed test; no external funding.
How language model feedback on student essays compares with teacher feedback 5Qualitative comparison of model and teacher feedback; working paper, not peer reviewed5 language models vs 12 teachers on 22 student essays at middle and high school grade levelsModels mostly gave corrective sentence-level feedback; teachers gave more dialogic feedback. No student writing followed forward.
GPT-4 can assess and give feedback on argumentative writing 6GPT-4 evaluation against instructor evaluation with a rubric; conference paper; a second phase had students revise, but the results read report no revised scores33 opinion articles by students at Brazilian Army military schools; grade not statedFeedback was all at task level and mostly at process level, with none at regulation or self levels. GPT-4 was consistent on structure and references but conservative in scoring.
Language model feedback improves revision, motivation and emotions (described by others) 5Randomized controlled trial by Meyer and colleagues; paper not opened, so scoring method and delay unknown459 high school sophomoresAs described by Mah and colleagues: d = 0.19 for text revision, 0.36 for task motivation, 0.34 for positive emotions vs no feedback.
Language model feedback in an elementary classroom (described by others) 3Classroom intervention in Japan; paper not opened14 students; sixth-grade reflective writing in mathematicsThe review's table gives the outcome as reflection quality and says AI feedback was comparable to teacher feedback.
AI-generated feedback reduces the decline in undergraduates' revised essays compared with no feedback (generative, randomized) 10Randomized controlled trial, three arms (AI feedback, teacher feedback, no substantive feedback); not preregistered; complete-case analysis; two supervised sessions, one feedback-and-revision cycle on one prompt; two raters unaware of condition312 randomized first-year undergraduates, 287 analyzed; one university; feedback from one locally run model, reported as Qwen2.5-27B-InstructRevised essays scored lower than first drafts in every arm (mean change -13.24 control, -7.09 AI, -7.04 teacher), so the gaps against control (d = 0.60 AI, 0.59 teacher) reflect a smaller decline, not improvement over the first draft. Equivalence of AI and teacher feedback within a margin of 0.35 SD the authors set; power 0.55; not shown at 0.25 or below. Intrinsic motivation higher with teacher than AI feedback (d = 0.31) and with AI than none (d = 0.35). No delayed or unassisted writing test. No funding declared.

What the evidence supports and where it stops

What the evidence supports

  • Older automated writing feedback tools are associated with better writing in pooled studies of high school and university students, with wide variation between studies and a follow-up effect not distinguishable from zero.
  • The school-age evidence on automated feedback is eight studies, seven of them quasi-experiments, with students aged 11 and up.
  • In one randomized trial in Grades 7 and 8, an older automated tool showed no significant overall effect after a school year, and results differed by district.
  • The language model studies at school level mostly measured the feedback, not students' later writing; 29 of 34 studies in the largest review were set in universities.
  • One small quasi-experiment with fourth-graders reported large gains in poetry writing after six weeks of ChatGPT-supported feedback against a control group that got little feedback.
  • One randomized trial with undergraduates found that revised essays scored lower than first drafts in every group, with a smaller decline after AI-generated or teacher feedback than after no substantive feedback, within a single supervised revision cycle.

What it does not show

  • That AI feedback improves writing in Grades 4 to 9. No randomized trial of generative AI feedback in Grades K to 8 was found in these indexes.
  • That the pooled effect of 0.55 applies to elementary or middle school students or to chatbots; its samples are high school and university, and its text does not mention generative AI.
  • That longer use works better or that benefits last. The duration difference was not significant and the follow-up effect did not differ from zero.
  • That the Grade 4 result would hold with a fair comparison. The control group got whole-class verbal comments only, and the scoring is described inconsistently.
  • That the Grade 7 and 8 trial's null result applies to generative tools, or what its effect sizes were; only its abstract record was read.
  • Anything about Grades K to 3, or about what happens to students' writing after the tool is removed in any grade band.
  • That the undergraduate trial's result carries to school students, or that it shows what happens to later, unassisted writing; its design covers one revision cycle.

How sure I am

For one narrow claim, that older automated feedback tools are associated with better writing in high school and university students, I would say moderate. There is a pooled estimate from 20 studies, and the school-age review mostly agrees in direction.12 I stop short of high because the prediction interval spans zero and the follow-up effect did not differ from zero.1 Most of the school-age studies were also quasi-experimental.2 The authors of the meta-analysis say a fuller test for publication bias is needed.1 Neither school-age experiment on older tools that I found showed an overall benefit: Ware's study, the one experimental study in the 2022 review, and the Grade 7 and 8 randomized trial.27 Each used a different comparison: online teacher or peer feedback for Ware, and students not using the tool for the trial.27 The positive Meyer trial in high school tested generative feedback, which I grade separately below.

For Grades 6 to 9, I would say low. The randomized trial I found there reported no significant overall effect on writing, with significant interactions by district.7 The quasi-experiments reviewed in 2022 lean the other way, and I cannot reconcile the two without the main paper's numbers.2

For generative feedback with older students, I would say low to moderate: one trial I could not open, and one unregistered trial in which revised essays scored lower than first drafts in every group.510

For generative AI feedback in Grades 4 to 9, I would say very low. The evidence is one small quasi-experiment in Grade 4, one classroom intervention with 14 sixth-graders that I could not read, and studies of the feedback itself.34 For Grades K to 3, none found in these indexes.

The case for optimism runs like this: the pooled effect is medium, the randomized trials of generative feedback with high school and university students favour feedback over none, in one case only as a smaller decline, and the Grade 4 study reports a large effect.14510 I accept that the pooled effect and the two trials favour feedback, and I treat the Grade 4 result as unresolved. My objection is to carrying them across to younger students without a test, because the pooled effect and both trials involve older students, and the Grade 4 result has the design problems above.

What would change my mind is evidence of three kinds. One is a randomized trial of a generative feedback tool in Grades 4 to 8 that scores later, unassisted writing blind. Another is the full text of the Grade 7 and 8 trial, showing what its effect sizes were. The last is a replication of the Grade 4 study with a control group that gets a comparable amount of structured feedback from a teacher.

What this piece cannot know

The field is moving fast, so studies I did not find, or that appeared after my last search on November 12, 2026, may change this picture. I also could not open a university randomized experiment of language model feedback in an educational psychology journal, because the publisher blocks scripted access, or a recent quasi-experiment in middle school that paired generative AI feedback with teacher and peer feedback.

What to try on Monday

If you are asked to adopt an AI feedback tool for writers in the middle grades, the gaps in the evidence justify a small, cheap look of your own. Pick one assignment and collect first drafts. Split the class by alternating names on the roll. One half revises with the tool's feedback, and the other half revises with a short checklist you wrote. Cover the names and score every revision on one trait of your usual rubric, a few minutes each. If that is too much, score a random eight from each half.

Then repeat the check a week later, with the tool and the checklist both put away. Give the whole class a short on-demand paragraph and score it the same way. Compare the two halves on that paragraph. If the tool half scored higher on the revision but no higher on the unassisted paragraph, that pattern is consistent with the tool improving the draft and not the writer. Half a class is a look and not a test, so a small difference means little and a null does not show the tool is fine. Find out who at your school approves tools and student data before you split a class.

If you are asked about a vendor's claim, ask three things. Which grades were studied? Was the control group getting a comparable amount of feedback from a person? Was writing scored later, without the tool, by someone who did not know who used it?

Method

An AI language model (Claude, from Anthropic) ran the searches, read the sources and drafted this text. Where the piece says "I" about searching, reading or checking, the assistant did that work to a brief I set, and the judgments are published under my name. The assistant searched four databases on November 4, 2026: OpenAlex, ERIC, Europe PMC and Crossref. It ran 17 queries then and three more, limited to 2026, on November 12, 2026, before publication. The re-run covered OpenAlex, Europe PMC and Crossref, because ERIC returned an error, and I added one date-filtered OpenAlex check that was not logged. One of the 20 logged queries was written to find null results and another to find retracted work.

A scoping pass on November 1, 2026 read two papers in full and several abstracts before I wrote the protocol on November 4. The protocol was written before any search for this piece was run, but the question was shaped by what the scoping pass found, and the protocol records it. The protocol is kept with the piece and lists the question, the rules and every screened source. Ten sources are cited. Nine were read in full text; the exception, the main report of the Grade 7 and 8 trial, was read as an abstract record only. Every DOI was looked up on Crossref and checked for retraction on OpenAlex. A script checked every figure and quotation in the text against a passage copied from its source, and my own arithmetic was exempt from that match. A model from a different family than the drafter (Llama 3.3 70B) then read the claims beside their passages in batches and judged whether each passage supports its sentence.

Four AI reviewers with assigned roles, an AI editorial gate and an AI persona of me read drafts over several rounds. They are model judgments, not human peer review, and passing them is not evidence that the piece is right.

Sources I could not open did not carry the argument, with labelled exceptions: the Meyer trial, the Steiss and colleagues study, the Ezoe and Takagi classroom study and the Zhai and Ma meta-analysis appear only as described by other authors, and the main report of the Grade 7 and 8 trial appears as an abstract record. The protocol records this as a deviation from its own rule that unread sources cannot carry numbers. Several papers I could not open in full are listed as excluded in the protocol with the reason for each. Among them are Steiss and colleagues (used above only for the review's outcome label), a Grade 4 study in a second journal, a Thai sixth-grade study known only from its abstract and a Moroccan study with implausible effect sizes that I could not verify. A Korean seventh-grade study was excluded because it measured attitudes and not writing. Generative AI papers marked as retracted were excluded. A single AI assistant screened the sources, so there was no second screener. The databases favour English-language research, and studies outside them would be missed. If you know of one I missed, write to me and I will add it with a note.

Funding and affiliations: nobody paid for or commissioned this piece. I have no financial relationship with the study authors, Measurement Incorporated or any AI feedback vendor, with one exception: I pay Anthropic for the model used here, and Anthropic also sells Claude to schools. One study I describe (Mah and colleagues) included a Claude model among the five it compared. I build websites and web strategy for a living, and through thisismyurl.com I also help colleges and training teams learn and work with AI, so I have an interest in how these tools are judged.

Sources

  1. Fleckenstein, J., Liebenow, L. W., & Meyer, J. (2023). Automated feedback and writing: a multi-level meta-analysis of effects on students' performance. Frontiers in Artificial Intelligence, 6, 1162454. doi:10.3389/frai.2023.1162454 Meta-analysis (journal article) Full text read via Europe PMC (PMC10351274). No funding statement found in the text; authors declare no commercial or financial relationships. ↩
  2. Nunes, A., Cordeiro, C., Limpo, T., & Castro, S. L. (2022). Effectiveness of automated writing evaluation systems in school settings: A systematic review of studies from 2000 to 2020. Journal of Computer Assisted Learning, 38(2), 599-620. doi:10.1111/jcal.12635 Systematic review (journal article) Full text read from the University of Porto repository copy; first published online December 2021. ↩
  3. Jiao, Y., & Wang, Q. (2026). Large language models for formative feedback in writing instruction: a systematic review of classroom interventions, feedback quality, and student outcomes. Frontiers in Education, 11. doi:10.3389/feduc.2026.1834085 Systematic review (journal article) Full text read (open access). Narrative synthesis; no pooled effect. ↩
  4. Kultas, E. (2026). Integration of artificial intelligence into primary school classrooms: the role of ChatGPT in improving poetry writing skills and attitudes towards poetry. BMC Psychology, 14. doi:10.1186/s40359-026-04844-4 Quasi-experimental study (journal article) Full text read via Europe PMC (PMC13393496). Single author; no funding declared. ↩
  5. Mah, C., Tan, M., Phalen, L., Sparks, A., & Demszky, D. (2025). From Sentence-Corrections to Deeper Dialogue: Qualitative Insights from LLM and Teacher Feedback on Student Writing. EdWorkingPaper. doi:10.2139/ssrn.5213040 Working paper (not peer reviewed) Full text read. It also describes the Meyer et al. randomized trial in its literature review; that description is the only support in this piece for Meyer et al. (the paper itself was not opened). ↩
  6. da Silva, D. C. A., de Mello, C. E., & Garcia, A. C. B. (2024). Analysis of the effectiveness of large language models in assessing argumentative writing and generating feedback. In Proceedings of the 16th International Conference on Agents and Artificial Intelligence (Vol. 2, pp. 573-582). SCITEPRESS. doi:10.5220/0012466600003636 Conference paper Full text read. ↩
  7. Wilson, J., Palermo, C., Myers, M. C., Cruz Cordero, T. M., Eacker, H., Potter, A., Coles, J., May, H., & Zhang, F. (2026). Impact of Automated Writing Evaluation (AWE) on Middle School Students' Writing Outcomes. American Educational Research Journal, 63(1), 72-108. doi:10.3102/00028312251388655 Randomized trial (journal article, abstract read) Only the ERIC abstract record could be opened; the publisher blocks scripted access. Cited for what the abstract says and no more. ↩
  8. Wilson, J., Zhang, F., Palermo, C., Cruz Cordero, T., Myers, M. C., Eacker, H., Potter, A., & Coles, J. (2024). Predictors of middle school students' perceptions of automated writing evaluation. Computers & Education, 211, 104985. doi:10.1016/j.compedu.2023.104985 Survey study within a randomized trial (journal article) Full text read via Europe PMC (PMC10839244). Describes the students and the software in the trial summarized in source 7. ↩
  9. Wilson, J., Delgado, A., Palermo, C., Cruz Cordero, T. M., Myers, M. C., Eacker, H., Potter, A., Coles, J., & Zhang, S. (2024). Middle school teachers' implementation and perceptions of automated writing evaluation. Computers and Education Open, 7, 100231. doi:10.1016/j.caeo.2024.100231 Mixed-methods study within a randomized trial (journal article) Full text read via Europe PMC (PMC11656462). Reports the funder and teachers' use in the trial summarized in source 7. ↩
  10. Liu, Y., & Shi, M. (2026). As effective as the teacher, but less motivating: a randomized controlled comparison of AI-generated and teacher feedback on students' writing performance and intrinsic motivation. Frontiers in Psychology, 17, 1918748. doi:10.3389/fpsyg.2026.1918748 Randomized controlled trial (journal article) Found in the 2026-11-12 pre-publication freshness search; first published 2026-10-30. Open access; full text read. First-year undergraduate sample; no funding declared. ↩

Citation checks

Run on by research-core/verify_citations.py. A passing check shows a source exists and says what is quoted. It does not show the source is right.

  • 10 of 10 sources opened and read by the checking script.
  • 10 of 10 DOIs resolved on Crossref with a matching title and year.
  • 1 of 1 quoted passages found word for word in the source text.
  • Retraction status checked on OpenAlex: none retracted.