DepEd allows AI in action research as a helper, never as the researcher. DepEd Order (DO) No. 003, s. 2026, the Foundational Guidelines on Artificial Intelligence in Basic Education, limits AI in research to “ancillary support functions, such as the analysis of processed data, grammatical refinement, and citation verification”. The same section bars personal and sensitive information from AI systems, requires personnel to declare the AI tools they used in data processing on the order’s Annex A form, and forbids research papers “generated, in whole or in part, by AI without significant human contribution and proper disclosure” (Sec. VI.1.3).

The Research Management Guidelines, DepEd Order No. 16, s. 2017, came nine years earlier and say nothing about AI. Their rules on plagiarism, fraud and authorship still apply to every research proposal and report a teacher submits, whoever or whatever helped write it.

The rules below are those in force on November 21, 2026.

Key findings

  • Support tasks only. “The utilization of AI in data and research management is strictly limited to ancillary support functions, such as the analysis of processed data, grammatical refinement, and citation verification” (DO 003, s. 2026, Sec. VI.1.3).
  • Nothing that identifies a person goes into the tool. “The input, processing, or storage of personal and sensitive information within AI systems is categorically prohibited” (Sec. VI.1.3). A class list, an LRN or a recorded interview stays out.
  • Tools used in data processing are declared. “Personnel are required to formally declare all AI tools employed in data processing using the prescribed declaration form in Annex A” (Sec. VI.1.3). A paper generated by AI without significant human contribution and proper disclosure “shall be subject to applicable administrative sanctions” (Sec. VI.1.3).
  • AI is never an author. The order lists “treating as an author/co-author of a paper” among the inappropriate uses of AI (Sec. VI.1.6(ii)), and under DO 16, s. 2017, “The grantee/s will be the sole author of the research” (Sec. VI.B).
  • A citation of AI use links to the conversation. It identifies the tool and the prompt and “shall include access to the chat history with the AI tool, through a link” (DO 003, Sec. VI.1.7(i)). Since September 2025, APA Style cites a chat by its share link.
  • Chatbots invent references. In a test of 636 citations, 55% of those GPT-3.5 gave and 18% of those GPT-4 gave were fabricated (Walters & Wilder, 2023). In 2025, GPT-4o fabricated 35 of 176 citations, or 19.9% (Linardon et al., 2025).

What DepEd allows AI to do in research

DO 003, s. 2026 covers AI in teaching, assessment, research and office work, and it is strictest about research. A teacher may use AI to help develop instructional materials, subject to DepEd’s quality assurance and approval, and tests, under human validation (Secs. VI.1.1 and VI.1.2); lesson plans follow DepEd Order No. 016, s. 2026, which keeps AI out of core instructional decisions (para. 25). For research the order narrows the role to support, in one section of two paragraphs:

The utilization of AI in data and research management is strictly limited to ancillary support functions, such as the analysis of processed data, grammatical refinement, and citation verification. To ensure full compliance with Republic Act (RA) No. 10173, also known as the Data Privacy Act of 2012, the input, processing, or storage of personal and sensitive information within AI systems is categorically prohibited.

To maintain transparency and ethical standards, personnel are required to formally declare all AI tools employed in data processing using the prescribed declaration form in Annex A. Research outputs must remain fundamentally human-centric and compliant with existing ethical guidelines; consequently, the submission of research papers generated, in whole or in part, by AI without significant human contribution and proper disclosure is strictly forbidden and shall be subject to applicable administrative sanctions. (Sec. VI.1.3)

“Such as” makes the three tasks examples, not a closed list. Searching for sources fits beside them: elsewhere the order lets DepEd personnel “utilize AI tools to support the search for relevant references and sources, brainstorming sessions and development of scenarios/narratives”, and asks that the tools are “not used as a substitute for human production and validation” (Sec. VI.3.1).

What does not fit is AI doing the research. The order names those uses among the inappropriate ones: “falsifying data from raw data, generating official statistical results without human validation, generating speculative data that can be misconstrued as current or actual data, treating as an author/co-author of a paper, writing policies, essays, or speeches solely with AI” (Sec. VI.1.6(ii)).

Two phrases are left undefined. The order does not say what “processed data” are, but the next sentence bars personal information from AI systems, so the safe reading is data with nothing in them that points to a person. Nor does it say how much is “significant human contribution”, or whether a draft a chatbot writes, which the teacher then reworks, counts as ancillary support. “Ancillary” means secondary to the work, and the three functions the order names (analysis, grammar, citations) all work on what the researcher has already written or gathered. The safer course is to write every section yourself and use AI on it afterward; a paper with any part an AI generated must at least show significant human contribution and disclose it, or its submission is forbidden outright (Sec. VI.1.3).

DO 16, s. 2017 still governs what a teacher submits. It defines plagiarism as “intellectual theft, i.e. presenting and/or claiming another’s work as your own without appropriate recognition by way of referencing and citation” and fraud as “a deliberate form of deception intended to result in financial or personal gain” (Sec. III). A proponent signs the Declaration of Anti-Plagiarism in its Annex 3, attesting that the proposal is original and that the proponent “has cited properly all the references used”. Anyone found to have committed plagiarism “will be blacklisted from availing of any other research grant mechanism in the Department” (Sec. VI.D). A paragraph a chatbot wrote and a teacher submitted as their own breaks both orders at once.

Who enforces this? A violation of DO 003 “shall be dealt with administratively” (Sec. IX), and schools are “mandated to rigorously monitor and document the ethical use of AI by teachers and learners” (Sec. XI.2). The research committees that evaluate proposals score them with DO 16’s appraisal guides, which have no line for AI use. As of November 21, 2026, DepEd’s Central Office has issued no order or memorandum on how research committees should handle AI use or the declarations.

Two questions before you prompt

Every use of AI in research passes two tests in Sec. VI.1.3: what goes into the tool, and what comes back out of it. Only two of the six combinations are open.

DepEd Order No. 003, s. 2026, Sec. VI.1.3

Two questions before you prompt

Answer both for a task of your own, or open one of the tasks under the grid. Nothing you choose is kept or sent anywhere.

1. What will you put into the tool?
2. What will it give back?
The six answers: what goes in, by what comes back
Rows: what goes in. Columns: what comes back.Help with your own workPart of the research itself
Nothing about a personOpen: Allowed: declare it, check itYour taskClosed: Not allowed: the research is yoursYour task
Coded dataOpen, with a check: Allowed for analysis: check every numberYour taskClosed: Not allowed: invented or unchecked dataYour task
Anything that identifies a personClosed: Not allowed: personal informationYour taskClosed: Not allowed, twice overYour task

Answer both questions to see the rule for your task.

Nothing about a person · help with your own work

Allowed: declare it, check it

Use the tool as a helper, check what it gives you against the source, and list the use in your AI declaration.

  • “The utilization of AI in data and research management is strictly limited to ancillary support functions, such as the analysis of processed data, grammatical refinement, and citation verification.”
    DO 003, Sec. VI.1.3
  • “Utilize AI tools to support the search for relevant references and sources, brainstorming sessions and development of scenarios/narratives”
    DO 003, Sec. VI.3.1(i)
  • “personnel are required to formally declare all AI tools employed in data processing using the prescribed declaration form in Annex A”
    DO 003, Sec. VI.1.3

Coded data · help with your own work

Allowed for analysis: check every number

Give the tool only coded data, then recompute its results yourself (in Excel, for example) before they go into the report. Declare the use.

  • “The utilization of AI in data and research management is strictly limited to ancillary support functions, such as the analysis of processed data, grammatical refinement, and citation verification.”
    DO 003, Sec. VI.1.3
  • “generating official statistical results without human validation”
    DO 003, Sec. VI.1.6(ii), among the inappropriate uses

Anything that identifies a person · help with your own work

Not allowed: personal information

Take out every name, LRN, face, voice and telling detail first, or do the task without AI.

  • “the input, processing, or storage of personal and sensitive information within AI systems is categorically prohibited”
    DO 003, Sec. VI.1.3
  • “internal, confidential, and sensitive documents that contain personally identifiable information shall not be uploaded, shared, or processed through AI platforms, especially those that collect, store, or use input data for model training”
    DO 003, Sec. VII.1.1

Nothing about a person · part of the research itself

Not allowed: the research is yours

Write the paper, choose the data and draw the conclusions yourself; AI may then check the grammar and the citations.

  • “the submission of research papers generated, in whole or in part, by AI without significant human contribution and proper disclosure is strictly forbidden and shall be subject to applicable administrative sanctions”
    DO 003, Sec. VI.1.3
  • “treating as an author/co-author of a paper, writing policies, essays, or speeches solely with AI”
    DO 003, Sec. VI.1.6(ii), among the inappropriate uses

Coded data · part of the research itself

Not allowed: invented or unchecked data

Report missing data as missing, and let no result into the report that you have not computed or checked yourself.

  • “falsifying data from raw data, generating official statistical results without human validation, generating speculative data that can be misconstrued as current or actual data”
    DO 003, Sec. VI.1.6(ii), among the inappropriate uses
  • “Fraud – is a deliberate form of deception intended to result in financial or personal gain”
    DO 16, s. 2017, Sec. III.7

Anything that identifies a person · part of the research itself

Not allowed, twice over

Both limits apply: no personal information goes in, and the research does not come out of the tool.

  • “the input, processing, or storage of personal and sensitive information within AI systems is categorically prohibited”
    DO 003, Sec. VI.1.3
  • “the submission of research papers generated, in whole or in part, by AI without significant human contribution and proper disclosure is strictly forbidden and shall be subject to applicable administrative sanctions”
    DO 003, Sec. VI.1.3

13 tasks teachers ask about

  • Suggest action research topics for a problem I describe in general terms: “Many of my Grade 3 learners read slowly.”

    Goes in: nothing about a personComes back: help with your own work

    Allowed: declare it, check it

    Brainstorming is named in Sec. VI.3.1(i). Which topic you choose, and why, is yours to argue in the Context and Rationale.

  • List studies on repeated reading that I should read.

    Goes in: nothing about a personComes back: help with your own work

    Allowed: declare it, check it

    Searching for sources is a support task (Sec. VI.3.1(i)). Open and read each study before you cite it: chatbots list studies that do not exist.

  • Check the grammar of the Context and Rationale I wrote.

    Goes in: nothing about a personComes back: help with your own work

    Allowed: declare it, check it

    “Grammatical refinement” is one of the three support tasks Sec. VI.1.3 names. Take the corrections, not new sentences that change what you meant.

  • Check that every reference in my list exists, and format the list in APA 7.

    Goes in: nothing about a personComes back: help with your own work

    Allowed: declare it, check it

    “Citation verification” is named in Sec. VI.1.3, and Annex B rates “auto-generating bibliography in APA style” as minimal risk. The tool can be wrong too: confirm each entry at its source.

  • Translate my interview guide into Filipino.

    Goes in: nothing about a personComes back: help with your own work

    Allowed: declare it, check it

    The order’s rules do not list translation as a support task (Annex E mentions it only among the uses of ChatGPT and Gemini); it is help with an instrument you wrote. A translated guide is checked like any instrument: ask a teacher of the language to review it before you use it.

  • Compute the mean scores and a paired t-test from pretest and posttest scores listed under learner codes.

    Goes in: coded dataComes back: help with your own work

    Allowed for analysis: check every number

    “The analysis of processed data” is a named support task (Sec. VI.1.3). Recompute the results in Excel before they go into the report.

  • Suggest themes from parents’ open-ended answers, with names and places taken out.

    Goes in: coded dataComes back: help with your own work

    Allowed for analysis: check every number

    Remove every name and every detail that points to a family first. Then code the answers yourself: the themes in the report must be ones you can show in the data.

  • Group my learners by reading level, from the class list with their names and LRNs.

    Goes in: anything that identifies a personComes back: help with your own work

    Not allowed: personal information

    A class list names learners and gives their LRNs. Sort it in your own spreadsheet, or work from codes.

  • Transcribe the recording of my focus group with parents.

    Goes in: anything that identifies a personComes back: help with your own work

    Not allowed: personal information

    A recording carries the parents’ voices, and the order counts voice as biometric data (Sec. VI.1.5(i)). Annex B rates transcription tools minimal risk, but in research Sec. VI.1.3 decides what may go in. Transcribe it yourself and leave the names out.

  • Write my Context and Rationale from my notes.

    Goes in: nothing about a personComes back: part of the research itself

    Not allowed: the research is yours

    Drafting a section of the paper is not among the support tasks Sec. VI.1.3 names, and an essay written “solely with AI” is an inappropriate use (Sec. VI.1.6(ii)). The order does not say how much rewriting would make a chatbot’s draft yours; the safer course is to write it yourself and use AI for a grammar check afterward.

  • Reword paragraphs from a published study so the plagiarism checker does not flag them.

    Goes in: nothing about a personComes back: part of the research itself

    Not allowed: the research is yours

    Reworded and uncited, another author’s work is still plagiarism: “presenting and/or claiming another’s work as your own without appropriate recognition by way of referencing and citation” (DO 16, s. 2017, Sec. III.11). Quote it, or put it in your own words, and cite it.

  • Fill in the posttest scores of two learners who were absent.

    Goes in: coded dataComes back: part of the research itself

    Not allowed: invented or unchecked data

    An invented score is “falsifying data from raw data” and “speculative data” (Sec. VI.1.6(ii)). Report the two learners as missing and say how you handled them.

  • Run my finished report through an AI “humanizer” so a detector reads it as human writing.

    Goes in: nothing about a personComes back: part of the research itself

    Not allowed: the research is yours

    A humanizer rewrites text so that no one can tell where it came from. The order asks for “significant human contribution and proper disclosure” (Sec. VI.1.3), and “repeated or intentional misrepresentation” of AI use can be sanctioned (Sec. VI.1.7).

Source: DepEd Order No. 003, s. 2026, Secs. VI.1.3, VI.1.5(i), VI.1.6(ii), VI.1.7, VI.3.1 and VII.1.1, and Annex B; DepEd Order No. 16, s. 2017, Secs. III.7 and III.11.

The grid, the names of its rows and columns, the tasks and the notes are this figure’s reading of the order; the quotations are DepEd’s words. The order does not define “processed data”; with personal information barred in the same paragraph, the figure reads it as coded data with nothing that points to a person.

The first question is about data, not tools. A chatbot asked to compute a t-test from scores under learner codes is doing “the analysis of processed data”. The same chatbot given the class list with names and LRNs is processing personal information, which the order prohibits. Documents with personally identifiable information “shall not be uploaded, shared, or processed through AI platforms, especially those that collect, store, or use input data for model training” (Sec. VII.1.1). The same provision allows such data only inside AI systems DepEd itself governs, “solely for approved research or operational purposes”, with anonymization and other safeguards. DepEd’s assessment order says the same for classroom work: “teachers shall not input personal learner data or official DepEd documents into AI platforms” (DepEd Order No. 015, s. 2026, para. 37).

Removing names is not always enough. In a class of 32, “the only girl who repeated Grade 2” is one learner, and a parent’s answer can name a street. Take out every detail that points to a person before anything goes in.

Recordings need the most care. Annex B of the order rates “Transcription and captioning tools” as minimal risk, with the example of AI “transcribing parent-teacher meetings or online meetings for inclusivity”. That rating classifies the tool; it does not lift the research rule on what may go in. A recorded focus group carries every participant’s voice, which the order counts as biometric data (Sec. VI.1.5(i)). Transcribe research recordings yourself.

The second question is about authorship. A grammar check returns your own sentences, corrected. A request to “write my Context and Rationale” returns sentences you did not write, which you would then submit as yours. The same is true of a reflection, an interpretation of results, or missing scores filled in with likely values. Editing the result afterward does not make those sentences your own, and the order does not say how much editing would.

What AI may do in each part of the paper

DO 16 sets the parts of an action research proposal (Annex 2.B) and of the completed report (Annex 6.B). The table applies Sec. VI.1.3 to each part; the order itself does not go part by part.

Part AI may help with The researcher’s own work
Context and Rationale Finding studies and DepEd issuances to read; checking grammar and citations The problem, its evidence from your school, and the argument
Action Research Questions Brainstorming possible questions from a problem described in general terms The questions you commit to
Innovation, Intervention, and Strategy Ideas to consider; finding studies of similar interventions The intervention, fitted to your learners
Action Research Methods Translating an instrument you wrote, which a person then checks; suggesting analysis methods to look up The design, the participants, the instruments and the data
Data Analysis Plan Suggesting analysis methods to look up The choice of analysis, and naming any AI tool you will use: the appraisal guide’s 8-point level asks that “tools (ex: software) to be employed are specified” (Annex 4.b)
Discussion of Results and Reflection Computing statistics from coded data, which you then recompute; checking grammar The interpretation, and the reflection on your own practice
References Checking that each entry exists; formatting in APA 7 Reading every source you cite
Abstract, Action Plan, Financial Report Checking grammar The content

Recomputing is quick for the tests action research uses most. The paired and independent t-test, weighted mean and thematic analysis guides show each step in Excel or by hand, so a figure a chatbot gives can be checked against your own.

AI action research generators and tools

An app sold as an “action research generator” turns a few lines of input into a title, a rationale, research questions and sometimes results. A paper made that way is exactly what Sec. VI.1.3 forbids, a research paper “generated, in whole or in part, by AI without significant human contribution”, and declaring the generator does not cure it: the order allows AI in research only for support functions. Results such an app writes are worse still. Data no one gathered are “speculative data that can be misconstrued as current or actual data” (Sec. VI.1.6(ii)), and presenting them as findings is the “deliberate form of deception” DO 16 calls fraud.

DO 003 regulates uses, not brands. Its table of examples of learners’ AI use names ChatGPT and Google Gemini for brainstorming and Grammarly and Quillbot for “writing improvement” (Table 1). Its Annex E, the DepEd AI Registry as the order prints it, lists ChatGPT, Microsoft Copilot and Google Gemini, each entry dated August 29, 2025 and marked “Pilot”. The notes in those entries are the closest the order comes to saying how to use a particular tool:

Tool (Annex E) Risk classification in Annex E Oversight note in Annex E
ChatGPT (OpenAI) “High Risk (education context with minors; misinformation & privacy risks)” “Teacher reviews and approves outputs; fact-checking required; prohibit entry of student PII; escalation to human officer for sensitive cases.”
Microsoft Copilot “Limited to High Risk (depending on data integration and PII handling)” “Staff validate outputs; sensitive data requires manual redaction before AI use.”
Google Gemini “High Risk (content generation, privacy, and bias concerns)” “Teacher or staff must review generated content; AI not allowed to provide final learning assessments.”

The best AI for action research is whichever does the support task in front of you, with nothing personal in it: a grammar checker for your own draft, a search tool to find studies you will then read, a chatbot to compute statistics from coded scores, which you then recompute in Excel. Look in the tool’s settings for whether what you type is used to train its models, since the order singles out platforms “that collect, store, or use input data for model training” (Sec. VII.1.1).

The research fund will not pay for a paid plan. The Basic Education Research Fund (BERF) “will not be used for” equipment, “Software”, or the “honorarium for resource persons, statisticians, and other service providers” (DO 16, s. 2017, Sec. V.C.i). DepEd Order No. 026, s. 2021 added mobile load and internet subscriptions to the communication expenses the grant covers (para. 3), not software.

Why AI tools invent references

A chatbot asked for sources can return a list that looks right and is partly made up. APA Style puts it plainly: “AI will confidently provide citations, but the citations are to works that are not real, or the works are real but do not actually support the statement.” Studies that checked every reference a chatbot gave found the same, at rates that vary with the model and the subject and never reach zero:

Study What was tested References checked Fabricated
Walters & Wilder (2023) GPT-3.5 and GPT-4, writing 84 short literature reviews on 42 topics 636 55% of GPT-3.5’s; 18% of GPT-4’s
Chelli et al. (2024) GPT-3.5, GPT-4 and Bard, finding sources for systematic reviews 139, 119 and 104 39.6%, 28.6% and 91.4%, counted as a reference with two of its title, first author and year wrong
Linardon et al. (2025) GPT-4o, writing on mental health topics (June 2025) 176 19.9%; and 45.4% of the real ones had errors, most often a wrong DOI
Güneş et al. (2026) Six models answering 120 radiology questions, with four references asked for each answer Every reference each model gave From 3.1% (Claude 3.5 Sonnet) to 60.6% (Gemini 1.5 Pro); ChatGPT models 27.7% to 52.9%

A real reference can also be wrong. Walters and Wilder found substantive errors in 43% of GPT-3.5’s real citations and 24% of GPT-4’s, while “every citation is in APA format”: APA formatting says nothing about whether a source exists. APA’s own blog, testing ChatGPT in 2023, met a fifth reference that “does not seem to be a real article; the digital object identifier given for that reference belongs to a different article.”

Philippine courts have already met the problem. In August 2025 the Sandiganbayan flagged a defense lawyer’s motions whose citations were “non-existent or did not contain the quoted portions cited and had incorrect docket numbers and dates”, and reminded counsel “to fact-check any AI-generated research” (Philstar Life, August 14, 2025). Teachers ask chatbots for sources too: in a 2026 survey of 197 graduating preservice teachers at a public university in Ilocos Norte, who do classroom action research during their internship, 38% said they always used ChatGPT to generate citations (Tejano, 2026).

A reference list for a study of repeated reading in Grade 3

Catch the invented references

Open the checks under each entry, then say what it is. Three of the seven were invented for this figure. Nothing you choose is kept or sent anywhere.

  1. 1Dela Cruz, M. A., & Ramos, J. P. (2019). Repeated reading and the oral reading fluency of struggling Grade 3 readers. Journal of Literacy Research, 51(2), 145–162. https://doi.org/10.1126/science.281.5381.1342

    Open the DOI: Catch the invented references

    doi.org opens “A Neoproterozoic Snowball Earth” by Hoffman, Kaufman, Halverson and Schrag, in Science, volume 281, issue 5381, pages 1342–1346, published August 28, 1998: a geology paper.

    Search the exact title: Catch the invented references

    Crossref finds no article with this title. Its closest matches are other studies of repeated reading, none by Dela Cruz or Ramos.

    Look up the journal issue

    Volume 51, issue 2 of the Journal of Literacy Research (2019) holds articles on preparing literacy teachers. Pages 145 to 162 fall across two of them: Wetzel and colleagues (pages 138–157) and Fowler-Amato and colleagues (pages 158–176).

    Entry 1 is

    Your call matches.

    Look again: the checks say otherwise.

    Invented

    Every part is believable and no two parts agree: the journal is real, the title is nowhere, the pages cut across two other articles, and the DOI belongs to a geology paper. Delete it, and never cite a study you have not opened.

  2. 2Department of Education. (2019). Guidelines on reading remediation in the early grades (DepEd Order No. 21, s. 2019).

    Open DepEd’s page for the order

    DepEd Order No. 021, s. 2019, issued August 22, 2019, is the Policy Guidelines on the K to 12 Basic Education Program.

    Search DepEd’s orders for the title

    No DepEd order carries this title.

    Entry 2 is

    Your call matches.

    Look again: the checks say otherwise.

    Invented

    A real order number under an invented title. If you meant a DepEd rule on reading, find the issuance itself and cite it by its own number, date and title.

  3. 3Hudson, R. F., Lane, H. B., & Pullen, P. C. (2005). Reading fluency assessment and instruction: What, why, and how? The Reading Teacher, 58(8), 702–714. https://doi.org/10.1598/RT.58.8.1

    Open the DOI: Catch the invented references

    doi.org opens this article in The Reading Teacher, volume 58, issue 8, pages 702–714. The authors, year and title match.

    Search the exact title: Catch the invented references

    Crossref finds it with the same authors, year, volume, issue and pages.

    Entry 3 is

    Your call matches.

    Look again: the checks say otherwise.

    Real and right

    Real, and cited correctly. Read it before you cite it: a real reference can still be cited for something it does not say.

  4. 4Lee, J., & Yoon, S. Y. (2015). The effects of repeated reading on reading fluency for students with reading disabilities: A meta-analysis. Journal of Learning Disabilities, 48(2), 213–224.

    Open the DOI: Catch the invented references

    The entry gives no DOI to open.

    Search the exact title: Catch the invented references

    Crossref finds this article by Lee and Yoon in the Journal of Learning Disabilities, volume 50, issue 2, pages 213–224, dated 2017 (published online in August 2016), with the DOI 10.1177/0022219415605194.

    Entry 4 is

    Your call matches.

    Look again: the checks say otherwise.

    Real, details wrong

    A real study with the wrong year and volume. Correct the entry from the article itself:

    Lee, J., & Yoon, S. Y. (2017). The effects of repeated reading on reading fluency for students with reading disabilities: A meta-analysis. Journal of Learning Disabilities, 50(2), 213–224. https://doi.org/10.1177/0022219415605194

  5. 5Rasinski, T. V. (2012). Why reading fluency should be hot! The Reading Teacher, 65(8), 516–522. https://doi.org/10.1002/TRTR.01077

    Open the DOI: Catch the invented references

    doi.org opens this article in The Reading Teacher, volume 65, issue 8, pages 516–522. The author, year and title match.

    Entry 5 is

    Your call matches.

    Look again: the checks say otherwise.

    Real and right

    Real, and cited correctly.

  6. 6Santos, M. R., & Reyes, J. L. (2021). Repeated reading and oral reading fluency among Grade 3 Filipino learners. Philippine Journal of Reading Education, 14(2), 45–61.

    Search the exact title: Catch the invented references

    Crossref and OpenAlex find no article with this title, and none by Santos and Reyes on repeated reading.

    Look up the journal

    No journal by that name appears in Crossref, OpenAlex or a web search.

    Entry 6 is

    Your call matches.

    Look again: the checks say otherwise.

    Invented

    A Philippine journal and two common Filipino surnames make it easy to believe. A real journal has a website listing its issues; when you cannot find the journal, you cannot cite the article.

  7. 7Therrien, W. J. (2004). Fluency and comprehension gains as a result of repeated reading: A meta-analysis. Remedial and Special Education, 25(4), 252–261. https://doi.org/10.1177/07419325040250040801

    Open the DOI: Catch the invented references

    doi.org opens this article in Remedial and Special Education, volume 25, issue 4, pages 252–261. The author, year and title match.

    Search the exact title: Catch the invented references

    Crossref finds it with the same details.

    Entry 7 is

    Your call matches.

    Look again: the checks say otherwise.

    Real and right

    Real, and cited correctly.

Checks run on November 21, 2026 through doi.org, Crossref, OpenAlex and DepEd’s page for DepEd Order No. 021, s. 2019. The Dela Cruz and Ramos article, the Santos and Reyes article and the DepEd order on reading remediation are invented; their authors’ names are invented too.

The four real works are cited as Crossref records them, one of them with its year and volume changed for the figure. The DOI in the first entry is real: it belongs to a 1998 geology paper in Science.

DO 003 lets AI help with “citation verification” (Sec. VI.1.3), but the check that counts is your own:

  1. Open the DOI. Add it to https://doi.org/ and see where it leads. Liberty University Library’s guide puts the rule simply: if the DOI “Resolves to a different article than the one cited”, “Then the citation is incorrect or hallucinated.”
  2. Search the exact title in quotation marks in Google Scholar, or paste the whole reference into Crossref’s Simple Text Query, which returns the DOIs of the works it matches.
  3. Find the journal and the issue on the journal’s own website when there is no DOI.
  4. Open DepEd’s page for any DepEd order or memorandum, and compare the number, date and title.
  5. Read the source before you cite it. A real study can still be cited for something it does not say.

When none of these finds the work, take it out. Boston University Libraries’ guide ends its steps with “Still nothing? Don’t cite it.” The review of related literature guide covers where to search for free and how to judge a source.

How to cite AI in an action research

DepEd sets what a citation of AI use must contain; APA sets the format most action researchers use. DO 003 asks that a citation identify “the specific AI tool and prompt used as applicable” and “include access to the chat history with the AI tool, through a link, which can be downloaded after each chat session making the entire conversation visible and accessible to the reader” (Sec. VI.1.7(i)). Its intent, the order says, “is to promote honesty and critical reflection, not to impose formal citation rules on every instance of AI use”, and the citation “should follow acceptable academic standards”. Its own example is “OpenAI. (2025). ChatGPT (July 7 version) (link to the chat session or history). Used ChatGPT to generate insight or recommendation regarding the symbolism of the tabo in the book El Filibusterismo”.

DO 16 names no citation style, and DepEd’s Action Research Toolkit (2019) gives its examples in APA. Region I’s research style guide lists the APA Publication Manual, 7th edition, among its references and sets its reference examples in APA style (DepEd Region I, Regional Memorandum No. 1415, s. 2025, Secs. 3 and 8). The APA 7 guide for DepEd orders covers the rest of a reference list.

On September 9, 2025, APA Style replaced its 2023 advice on citing ChatGPT. Its new formats fit DO 003 better, because a reference can now point to the chat itself:

  • A specific chat, by its share link: “AI Company Name. (year, month day). Title of chat in italics [Description, such as Generative AI chat]. Tool Name/Model. URL of the chat”. APA’s example: OpenAI. (2025, August 21). High school grammar concepts [Generative AI chat]. ChatGPT. https://chatgpt.com/share/68a77b60-0ee4-800c-9acc-cd3fd573c311
  • A tool in general, when a chat cannot or need not be cited: “AI Company Name. (year). Tool Name/Model in Italics and Title Case [Description; e.g., Large language model]. URL of the tool”. For AI used “to edit, analyze, organize, or refine your own writing”, APA says an in-text citation “plus a general reference to the tool, rather than a reference for each chat, is enough”.
  • In the text: (OpenAI, 2025). The date of a chat is “the specific year, month, and day on which a chat occurred or concluded”.
  • No version numbers by default. The tool’s name or the model’s (ChatGPT, or ChatGPT-5) is enough.
  • Prompts go in the text or an appendix, not in the reference: “prompts can be described in the text, but they are not included in the reference.”

APA adds that “AI itself cannot be an author”. The Committee on Publication Ethics (COPE) asks authors who use AI in writing a paper or “in the collection and analysis of data” to disclose, in the methods section, “how the AI tool was used and which tool was used”. In a DepEd action research that becomes five things:

  1. A reference for each chat, by its share link. The link meets DO 003’s requirement and APA’s format at once. Before sharing, give the chat a plain title in the tool itself, as APA suggests.
  2. An in-text citation wherever the use is described.
  3. A statement in the Action Research Methods saying which tool did what. In a proposal, name the tool in the Data Analysis Plan.
  4. An appendix with each prompt, its date, its link, and what you kept or changed. DO 16’s report outline (Annex 6.B) has no section for AI, so the record goes after the References.
  5. The Annex A declaration, signed.

The log below writes the first four from your entries and fills in the fifth.

APA 7 and DepEd Order No. 003, s. 2026, Sec. VI.1.7 and Annex A

Your AI use log

List each use of an AI tool in your action research. The log writes the references, the in-text citations, a statement for your methods, the record for your appendix and Annex A’s declaration. It opens with a worked example: change it, or start a blank log. Nothing you type is kept or sent anywhere.

The paper
Stage
  1. Use 1
  2. Use 2

References (APA 7)

  • Google. (2026, October 2). Checking a reference list [Generative AI chat]. Gemini 2.5 Flash. https://g.co/gemini/share/your-chat-link
  • OpenAI. (2026, September 14). Grammar check of the Context and Rationale [Generative AI chat]. ChatGPT. https://chatgpt.com/share/your-chat-link

In the text: (Google, 2026); (OpenAI, 2026)

For the Action Research Methods

ChatGPT (OpenAI, 2026) was used to check the grammar of the Context and Rationale. Gemini 2.5 Flash (Google, 2026) was used to check that the sources cited in the reference list exist and format them in APA 7. The researcher reviewed every output before using any of it. The prompts, the link to each conversation, and what was kept or changed are recorded in the appendix.

Appendix: record of AI use

DateToolTaskPromptsLinkKept or changed
September 14, 2026ChatGPT (OpenAI)To check the grammar of the Context and RationaleCheck the grammar of this paragraph. Do not change its meaning or add content.https://chatgpt.com/share/your-chat-linkKept six corrections of verb tense and articles; rejected two rewordings that changed the meaning.
October 2, 2026Gemini 2.5 Flash (Google)To check that the sources cited in the reference list exist and format them in APA 7For each reference below, give its DOI and say whether the title, year and journal match the published article.https://g.co/gemini/share/your-chat-linkTwo entries could not be found in Crossref and were removed; one year was corrected from the article itself.

Annex A declaration, filled in

Consistent with the policy guidelines on the use of AI in basic education, I, [insert full name of personnel/teacher/learner], hereby declare that I have used AI tools to assist in the writing of my action research report “Daily Repeated Reading and the Oral Reading Fluency of Grade 3 Learners”.

I used ChatGPT and Gemini 2.5 Flash to help in the Context and Rationale and the reference list of the action research report for the purposes of finalization of research study. The prompt used were “Check the grammar of this paragraph. Do not change its meaning or add content”; “For each reference below, give its DOI and say whether the title, year and journal match the published article”.

I confirm that I have reviewed, edited, and understood all the content completed. I used AI tools only to support my learning/work/deliverable and not to replace my own thinking, agency, or output. I take full accountability for the final output submitted.

I affirm that the use of AI tools was intended solely to enhance the quality of the declared output. All outputs have been reviewed, adapted, edited, and validated using my professional expertise and agency to ensure accuracy, appropriateness, and alignment with Data Privacy Act of 2012.

Likewise, I affirm that no confidential information or sensitive institutional information data was shared with the AI platform/app identified above in the process.

Before it goes into the paper

  • Nothing to fix in this log.

Sources: the APA Style blog, “Citing generative AI in APA Style: Part 1—Reference formats” (September 9, 2025); DepEd Order No. 003, s. 2026, Secs. VI.1.3 and VI.1.7(i), and Annex A, page 27.

The reference formats are APA’s and the declaration is Annex A’s text word for word, with the log’s entries in its blanks. The statement for the methods and the record are this figure’s wording: adapt them to your paper. The example’s share links are placeholders.

The AI declaration form (Annex A)

Annex A of DO 003 has three sample declarations: a general one, headed “For learners, teachers, and non-teaching personnel”, one for teachers’ instructional materials, and one for learners. Sec. VI.1.3 asks for “the prescribed declaration form in Annex A” without saying which of the three; the general one fits research, since among the purposes its blanks suggest is “finalization of research study”. The DO 003 guide has a filler for all three templates.

Signing it commits you to three things, each in Annex A’s words:

  • “I confirm that I have reviewed, edited, and understood all the content completed.”
  • “All outputs have been reviewed, adapted, edited, and validated using my professional expertise and agency to ensure accuracy, appropriateness, and alignment with Data Privacy Act of 2012.”
  • “Likewise, I affirm that no confidential information or sensitive institutional information data was shared with the AI platform/app identified above in the process.”

A teacher who pasted a class list into a chatbot cannot sign the last line truthfully, and a teacher who did not check the chatbot’s references cannot sign the second.

Divisions have begun to remind teachers of the duty. The City of Imus division told its schools that “Any personnel or learners using AI tools for data processing, research, or instructional materials must formally declare such use using the prescribed AI Use Declaration form (Annex A of the Order)” (SDO Imus City, Division Memorandum No. 130, s. 2026). The 2026 calls for BERF proposals ask for DO 16’s anti-plagiarism declaration; the National Capital Region’s call, issued a month before DO 003, also asks for a “Free Plagiarism Checker Report” with the proposal and with the final output (DepEd NCR, Regional Memorandum No. 059, s. 2026).

Does a grammar check need declaring? The order speaks of “all AI tools employed in data processing” for research (Sec. VI.1.3), and it asks that research papers carry “proper disclosure”. It does not say whether a spell checker counts. Declaring it costs one line of the record, and APA does not require a reference for AI built into common software such as Copilot in Word.

The general rule on declarations is milder than the research rule. “Non-disclosure shall not, by itself, constitute a punishable offense, but repeated or intentional misrepresentation may be subject to applicable academic honesty provisions” (Sec. VI.1.7). That sentence governs everyday disclosure. A research paper generated by AI without significant human contribution and proper disclosure falls under Sec. VI.1.3 instead: its submission is “strictly forbidden” and carries administrative sanctions.

AI detectors and “humanizers”

An AI detector’s score is not proof. DO 003 says so for learners’ work: “AI detectors shall not be used as sole evidence of academic dishonesty” (Sec. VI.1.2(v)). It says nothing about detectors for teachers’ research, but the evidence on detectors does not change with the writer:

  • OpenAI withdrew its own AI text classifier on July 20, 2023, “due to its low rate of accuracy”. It had caught 26% of AI-written text and flagged human text as AI-written 9% of the time.
  • Seven detectors tested on essays by non-native English writers labeled more than half of them AI-generated, an average false-positive rate of 61.3% (Liang et al., 2023). Teachers who write research in English as a second language face the same risk.
  • Turnitin reported in 2023 that documents where it detects “less than 20% of AI writing” have “a higher incidence of false positives”, and it no longer displays a number for scores from 1% to 19%.

The Commission on Higher Education (CHED) set the same limit for colleges and universities: “AI detection tools, where used, shall not be the sole basis for determining academic misconduct” (CHED Memorandum Order No. 21, s. 2026, Sec. 9.2).

A “humanizer” rewrites text so that a detector reads it as human. It changes the score, not the authorship: a paragraph a chatbot generated is still not the teacher’s contribution after another tool has reworded it, and hiding where it came from is the kind of “intentional misrepresentation” Sec. VI.1.7 warns about. The way to lower an AI score is to write the paper yourself and declare the tools that helped.

A plagiarism checker answers a different question. It compares a text with sources it can find, so it catches copied passages; it cannot tell whether a reference is real.

What this means for school heads

DO 003 gives school heads the job of watching AI use in the school. Among their duties: “Establish a monitoring system with systematic record-keeping to track AI tool usage and any AI-related incidents” (Sec. VIII.4.4), and “Facilitate advocacy and orientation for teachers, learners, and parents regarding data privacy, informed consent, and the prevention of academic dishonesty” (Sec. VIII.4.2), along with professional development “focused on the ethical and effective application of AI” (Sec. VIII.4.1). In research, that comes down to these steps:

  1. Ask for the AI record with every proposal you endorse. In the immediate supervisor’s conforme of DO 16’s Annex 1, you certify that the proponent “has/have the capacity to implement a research study without compromising his/her office functions”. Before you sign, ask which AI tools were used and for what, and see the declaration. Keep a copy with the school’s research file: DO 003 asks that declarations kept as academic documentation be “retained for review or reference purposes” (Sec. VI.1.7(ii)).
  2. Teach the two questions in a LAC session. The grid above fits one learning action cell session: teachers sort their own recent uses of AI into its cells and rewrite one prompt so that nothing personal goes in.
  3. Keep school records out of AI tools. Class lists, school forms, LRNs, report cards and recordings of learners or parents never go into a chatbot, whatever the task. The Data Privacy Act guide for schools covers what counts as personal and sensitive information.
  4. Open a few references in each proposal. Use the five checks above on two or three entries, especially any study with Filipino learners that no one on the committee has read.
  5. Never decide on a detector’s score alone. Ask the teacher to walk you through the data, the analysis and the drafts. A teacher who did the work can.
  6. Point teachers in graduate school to their university’s rules. A thesis falls under its school’s policy. CHED’s order of July 21, 2026 says that “Disclosure of AI use in theses and dissertations shall be required” and leaves the format to each school (CMO No. 21, s. 2026, Sec. 11).

Checking that research rests on real sources is part of Strand 1.4 of the Philippine Professional Standards for School Heads, whose Career Stage 1 indicator asks a school head to “Identify relevant research findings from reliable sources in facilitating data-driven and evidence-based innovations to improve school performance” (indicator 1.4.1). The Domain 1 guide covers the strand.

Frequently asked questions

Can DepEd teachers use ChatGPT for action research?

Yes, for support tasks: finding sources to read, checking grammar, checking citations, and analyzing coded data that you then recompute. DO 003, s. 2026 forbids entering personal information into it and forbids research papers generated by AI without significant human contribution and proper disclosure (Sec. VI.1.3). Declare the tools you used on the order’s Annex A form.

Is there an AI action research generator DepEd allows?

No. A tool that writes the rationale, questions or results is generating the research, which Sec. VI.1.3 of DO 003, s. 2026 forbids and which carries administrative sanctions. Results the tool invents are also “speculative data” (Sec. VI.1.6(ii)), and presenting them as findings is the deception DO 16, s. 2017 calls fraud (Sec. III).

How do I cite ChatGPT in APA 7 in an action research?

Cite the chat by its share link, as in OpenAI. (2026, September 14). Title of the chat [Generative AI chat]. ChatGPT. URL of the chat. In the text it is (OpenAI, 2026). Describe the use in the Action Research Methods and put the prompts in an appendix (APA Style, September 2025). The link also meets DO 003’s requirement of “access to the chat history” (Sec. VI.1.7(i)).

Do I need to declare AI that only checked my grammar?

The order does not say. It requires declaring “all AI tools employed in data processing” and “proper disclosure” of research papers (Sec. VI.1.3), and it exempts nothing. Declare it: the City of Imus division already asks personnel to declare AI used for “data processing, research, or instructional materials” (Division Memorandum No. 130, s. 2026), and for lesson plans DepEd lists grammar and spelling checks among the AI uses to be declared (DO 016, s. 2026, para. 25). One line in the record covers it.

Does DO 003 apply to my master’s thesis?

A thesis written for a university follows that university’s policy. CHED Memorandum Order No. 21, s. 2026 requires disclosure of AI use in theses and dissertations and leaves the format to each school (Sec. 11). Schools have a transition period of 12 months by default (Sec. 30), so ask your graduate school for its format. If the same study is also submitted to DepEd as action research or for a BERF grant, DO 003’s research rules apply to that submission.

Sources

DepEd issuances

  • DepEd Order No. 003, s. 2026, Foundational Guidelines on Artificial Intelligence in Basic Education, February 20, 2026: Secs. VI.1 to VI.1.7, VI.3.1, VII.1.1, VIII.4, IX and XI.2; Table 1; Annexes A, B and E.
  • DepEd Order No. 16, s. 2017, Research Management Guidelines, March 20, 2017: Secs. III, V.C.i, VI.B and VI.D; Annexes 1, 2.B, 3, 4.b and 6.B.
  • DepEd Order No. 026, s. 2021, Amendment to DepEd Order No. 16, s. 2017 (Research Management Guidelines), June 21, 2021: para. 3.
  • DepEd Order No. 015, s. 2026, Revised Guidelines on Classroom Assessment, Grading System, and Awards and Recognition for the K to 12 Basic Education Program, June 4, 2026: para. 37.
  • DepEd Order No. 016, s. 2026, Guidelines on Lesson Planning and Learning Design, June 4, 2026: para. 25.
  • DepEd Order No. 021, s. 2019, Policy Guidelines on the K to 12 Basic Education Program, August 22, 2019 (in the reference-check figure).
  • Department of Education and the Basic Education Sector Transformation (BEST) Program. (2019). Action Research Toolkit: Teacher’s Guide, with support from the Australian Government.
  • DepEd National Capital Region, Regional Memorandum No. 059, s. 2026, Call for the Submission of 2026 Research Proposals for the Basic Education Research Fund (BERF), January 19, 2026, Enclosure 1, as attached to SDO Muntinlupa City Memorandum No. 041, s. 2026.
  • DepEd Region I, Regional Memorandum No. 1415, s. 2025, Department of Education Region I Research Style Guide, October 20, 2025: Secs. 3 and 8.
  • SDO Imus City, Division Memorandum No. 130, s. 2026, Foundational Guidelines on Artificial Intelligence (AI) in Basic Education, March 4, 2026.

Other Philippine sources

Citation style and publication ethics

Studies of AI-generated references and detectors

Checking a reference

Written by Mark Anthony Llego for NASH Reviewer. Sources: DepEd Order No. 003, s. 2026; DepEd Order No. 16, s. 2017; DepEd Order No. 026, s. 2021; DepEd Orders No. 015 and 016, s. 2026; CHED Memorandum Order No. 21, s. 2026; APA Style; COPE; studies of AI-generated references and detectors.